The document says offloading large KV Cache datasets to a more cost-effective external storage tier is emerging as a practical approach for scaling inference clusters. It neatly explains how KV Cache works in LLM inference and discusses how NVMe SSDs can help build a cost-efficient storage layer for AI infrastructure.
The document discusses the mechanisms used by vLLM and LMCache to make SSD-based KV Cache more effective. These are explained with text, diagrams and a couple of Python code examples. For example;
It says "Both vLLM and LMCache support packing KV Cache data from multiple tokens for transfer, improving memory efficiency and I/O efficiency."
The point here is that "External storage systems can use these larger chunks to generate flash-friendly, sequential large-block I/O patterns, improving bandwidth utilization on NVMe SSDs."
With PCIe gen5 NVMe SSDs having capacities rise to 245 TB, and hopefully beyond that next year, there is a lot of scope for building petabyte-level external storage KV caches with good bandwidth and high capacity. Our version of DapuStor's paper, generalized to all NVMe SSDs, can be downloaded here.
]]>HBM is the interposer-connected stacked memory for GPUs that by-passes x86 socket-connected DRAM capacity and bandwidth limits. The Joint Electron Device Engineering Council (JEDEC) published its JESD235 (restricted access) first HBM standard in 2013 and Samsung, SK Hynix and Micron built compliant HBM chips bought, in the main, by Nvidia, its system-building partners, and other GPU builders. Now, 23 years later, Nvidia has revolted against HBM memory controller functionality being in its GPU chip and moved it to the HMB base die unit, freeing up space for more GPU cores.
The change is in Nvidia’s NVHBM custom HBM standard, part of its August NVLInk Fusion announcement. It claims NVHBM delivers up to 30 percent greater memory bandwidth, 15 percent lower HBM power consumption, and frees up to 25 percent more area on its GPU compute die compared with standard HBM4E.
A couple of Nvidia GPU chip layout images show the effect:
Currently HBM has a wide physical interface with many pins connecting it via an interposer to the GPU where the HBM memory controller is located. NVHBM has a narrower and custom PHY interface with the memory controller function moved off the GPU, which means there is more space for AI-processing GPU cores as well as more HBM bandwidth.
However, this is specific to Nvidia and is not an industry standard.
As Nvidia is such a massive HBM buyer - according to reported Morgan Stanley data Nvidia will be responsible for purchasing 37 percent of the global HBM supply in 2027 - the three HBM suppliers are happy to play along and support Huang’s colossus. So much so that they even let Nvidia dictate the semiconductor foundry process they must use, TSMC’s 3nm. Reuters reported Samsung will accede to this request. South Korea’s Seoul Economic Daily said Samsung would however still build JEDEC standard HBM in-house for other customers
Also Amazon’s Amazon’s Annapurna Labs is supporting NVHBM as an Nvidia NVlink partner.
JEDEC has not issued any public comment about NVHBM.
NVHBM is effectively an industry standard because Nvidia is such a dominant HBM buyer and there are only three HBM-manufacturing suppliers all of whom are supporting Nvidia.
When JEDEC announced the HBM4 standard in April last year, the HBM subcommittee chair was Barry Wagner, Director of Technical Marketing at Nvidia. His continued chairmanship might now by untenable.
One JEDEC response might be to accept Nvidia’s invasion of its HBM standard setting scope and restrict its HBM specification activities to the stack above the HBM stack’s base memory controller die. Another might be to continue as before, with an absent Nvidia. A third might be to move in the NVHBM direction and add a base memory controller die with narrower and faster physical channels. This would build on reduced HBM PHY pin count work it is already doing.
]]>When we found about the four storage options: DDN EXAScaler and Infinia, NetApp ONTAP and VAST Data’s AI OS, we were intrigued and approached Nebul with some questions. And then found out it had five options, not four.
We asked: “Could it say how Nebul positions its four storage offerings please? Meaning DDN'sEXAScaler and Infinia, NetApp ONTAP and VAST Data's AI OS.
“DDN says Nebul is using Infinia for KV Cache workloads while Nebul's website positions it for data lake storage. Our understanding is that all four operating systems support NVIDIA's KV cache scheme and so we think it unlikely that Nebul will only be using Infinia for KV cache workloads. Could you confirm that this is the case please?”
Dion Bonné, Nebul’s Chief Product and Revenue Officer, told us: “You're right, we use several storage platforms, each serving different use cases. Let me explain how we position them.
DDN Infinia is our AI-native data lake and inference storage platform. It also supports advanced capabilities such as KV cache offloading through NVIDIA Dynamo and NIXL integration, which makes it particularly interesting for inference workloads.
DDN EXAScaler is used for workloads requiring extreme storage throughput to keep GPUs fully saturated. This is primarily relevant for large-scale AI training.
NetApp ONTAP provides enterprise file and object storage, particularly for customers already using the NetApp ecosystem who want to bring their existing data into AI workloads and applications.
NetApp - https://www.blocksandfiles.com/file/2026/09/29/netapp-unveils-fastest-file-system-on-the-planet-with-great-raft-of-other-announcements/5299808 - Novus is a high-performance parallel file system targeting similar workloads to DDN EXAScaler, but with the added benefit of integration into the broader NetApp data ecosystem.
VAST Data goes beyond storage and provides an integrated data platform. With capabilities such as VAST Database and DataEngine, customers can bring AI-native data services directly into their data infrastructure.
“Like most cloud providers, we don't believe in a single storage platform for every workload. Different architectures have different strengths, and ultimately the customer's use case, performance requirements, and preferences determine which platform makes the most sense.
“On your specific KV cache question: today we use DDN Infinia exclusively for KV cache workloads. That may change in the future as the technology evolves, but we're very happy with our collaboration with DDN and Nvidia. We've been working closely with their teams on optimizing protocols and integrations, with some impressive results documented in the recent whitepaper.
“So, to answer your original question, Infinia hasn't replaced NetApp or any of our other storage platforms. They complement each other, and we continue to invest in these different technologies to give our customers the best solution for their workloads."
]]>A Micron team is building a GPU cluster for the US Department of Energy’s Pacific Northwest National Laboratory (PNNL) to run next-generation, large-scale scientific, AI, HPC and data-intensive workloads. The cluster is called Great Abaco, and Dr. Skyler Windh and his team at Micron developed a high-capacity shared memory system architecture for AI and HPC workloads. It uses CXL 2.0.
The system is described in a 60-minute webinar and the background is this: as Large Language Model (LLM) inference scales across multi-GPU deployments, the Key-Value (KV) Cache becomes a major performance bottleneck. Each GPU typically keeps its own isolated cache, which can spill to host DRAM. When different GPUs process the same prompt prefix, each one repeats the same prefill computation. CXL fabric-attached memory (FAM) addresses this by providing multiple GPUs with a shared, byte-addressable memory pool, allowing other GPUs to reuse KV Cache entries computed by one GPU.
Windh and his team built cross-host KV sharing into Nvida Dynamo and ran it on Micron’s Great Abaco cluster.
Abaco is a shared-memory system architecture for AI and HPC, designed by Micron with Liqid as the memory fabric and orchestration partner. A single Abaco rack presents up to 160TB of DDR5 as an addressable pool that one or multiple hosts in the rack can claim, share, or release without a reboot and without re-cabling. Applications reach the pool through a standard file system interface, so most code runs unmodified.
The cluster uses H200 NVL GPUs in a Liqid PCIe Gen5 chassis, with four Liqid CXL memory chassis forming one shared pool.
On Llama3-70B with 90 percent shared prefixes, a 32 GB cache per server delivered 2.8× the throughput and 2.9× faster time to first token than DRAM of the same size. Hosts reusing each other’s KV blocks got 3× the throughput with half the memory, and CXL memory ran within 99 percent of host DRAM.
Ie. CXL FAM performs on par with a host DRAM baseline.
The Micron team integrated CXL FAM-based KV Cache sharing into an open-source multi-GPU inference framework, with:
A block-level cache manager that allocates in CXL memory
Sequence-hash-based cache discovery across GPUs
An asynchronous transfer engine that moves blocks between CXL and GPU memory without stalling the inference scheduler
A CUDA host-registration workflow that enabled direct memory access (DMA) between GPUs and CXL memory
Liqid SVP for Solution Architecture Martin Cooper says this Micron work shows that shared CXL memory can lift GPU inference throughput 2.8× over DRAM. Liqid puts GPUs and CXL memory on one PCIe fabric, so memory scales separately from compute. The serving stack stays as it is: no vLLM patches and no changes to the GPU servers. A shared pool lets you right-size each server and buy the memory once.
You can access the webinar 's 30-slide presentation deck here, and watch the whole webinar here - it's has many YouTube adverts though. You can download a transcript we created here if you wish, and read it along with the presentation deck.
Download an Abaco technical brief here.
GFAM, referred to in the slides = Global Fabric Memory
]]>However, the profit was down 44 percent compared to the previous quarter’s $3.2 billion profit. The prior quarter’s results included a large one-time Sandisk-related fair value gain (roughly $1.7 billion in the “interest and other” line. Revenue, non-GAAP gross margin, and non-GAAP EPS were all above its high end guidance with growth driven by public clouds and enterprises buying disks for cloud and AI workloads. Gross margins also improved on higher volumes, a mix shift toward higher-capacity nearline HDDs, and better pricing. Its revenue per exabyte shipped was $17.7 million, up 28 percent Y/Y and 8.8 percent Q/Q; price rises are working in its favor.
CFO Kris Sennesael, said: “Our first quarter results demonstrate the continued momentum in our business, with robust exabyte growth and improved pricing driving another quarter of strong financial performance.”
Exabytes shipped rose 21.6 percent Y/Y to 248, up from 231 EB in the previous quarter. Some 88 percent of WD’s revenues came from the cloud segment, 7 percent from client drives (up 119 percent Y/Y) and 5 percent from consumer drives. Ninety percent of its shipped drives were mass-capacity, nearline disks.
CEO Irving Tan said: “We are executing well through our technology transitions to meet the growing demand for storage. As we deepen our customer relationships, broaden the customer base to new areas of the market, and extend our visibility into future demand, we are building an increasingly durable foundation for long-term value creation.”
Financial summary:
Gross margin: 57.8% vs 50.2% in prior quarter
Operating cash flow: $1.6 billion vs $1.12 billion in prior quarter
Free cash flow: $1.4 billion vs last quarter’s $978 million
Cash and cash equivalents: $1.3 billion vs prior quarter’s $1.58 billion
Diluted EPS: $4.64 vs last quarter’s $8.20
WD declared a $0.175 quarterly cash dividend for shareholders, up 17 percent Q/Q.
It noted that, “In connection with the company’s actions to monetize its retained interest in Sandisk and reduce its debt, it completed a number of transactions, including a debt-for-equity exchange, equity-for-equity exchanges, and the private settlement of a portion of the company’s convertible note obligations. In connection with these transactions, the company recognized costs primarily related to discounts given to the counterparties of the transaction.” We think that helps explain the $265 million drop in the cash and cash equivalents number, together with share repurchases. It has a net $0.5 billion cash position.
The company said that it’s ramping production of its next generation ePMR platform, meaning another jump in capacity, to the 40 TB level. Tan said: “We expect this platform to represent close to 30 percent of our nearline shipments by the second quarter and 50 percent by the third quarter of fiscal 2027.”
The next-generation HAMR (heat-assisted magnetic recording) tech is progressing; Tan said in the earnings call: “On our first-generation HAMR product, we are on track to finish qualifications with 2 hyperscale customers in the first quarter of calendar year 2027 and are on schedule to begin volume shipments in the first half of calendar year 2027. We also expect to begin qualifications with 2 additional customers during the first half of calendar year 2027.”
This will be for 44 TB HAMR drives.
That looks like a far shorter HAMR drive qualification period than Seagate endured. Tan noted: “Feedback from our customers have been very positive, particularly around the failure rates that we've seen in our drive performance vis-a-vis what they've seen with other players within the market.”
There’s more: “Starting in the second half of calendar year '27, we will also start qualification of our second-generation HAMR drives, and this will be at 5 terabytes per platter on [an] 11-disk platform.” Expect 55 TB drives.
Tan is confident higher drive IO speeds are needed: “We are expanding the role of HDDs within cloud and AI infrastructure through innovations such as our high-bandwidth drive and dual pivot technology. By combining the economics of HDDs with significantly higher levels of performance, we believe this technology can address a broader range of existing and emerging workloads.”
“We are currently sampling our high-bandwidth drives with multiple customers, and the feedback has been positive with clear use cases and value creation identified.”
He expanded in this nascent opportunity: “Things like high-bandwidth drives provide customers with an opportunity to use HDDs for different workloads that they may have to use alternative storage media for, which will be much more costly. For example, KV cache workloads in AI inferencing and in video streaming as well.”
KV cache workloads are served totally from NVMe SSDs at present. The idea that disk drives with faster IO could cannibalize some much lower latency NVMe SSD KV cache units seems a tad far-fetched.
Nevertheless Tan pointed out: ”We've been actually providing samples to many of our customers in terms of this product so that they can test out different use cases. And I highlighted that's been very positively received.”
Tan said: “We are increasingly seeing storage demand grow beyond our core cloud and hyperscale customers into enterprise environments and into newer end markets such as neocloud and AI frontier labs, resulting in a larger and more diversified customer base. Our customers are making longer-term infrastructure investments and continue to extend their planning horizons into the next decade.”
He added: “Our visibility is extending further and with it our conviction in the durability of demand for data storage on HDDs. … As we ramp production across our next-generation ePMR and HAMR platforms, we expect our exabyte shipments to continue to increase through fiscal 2027 and beyond.”
Tan dismissed Toshiba’s plan to increase its shipped drive unit numbers and move from a 10 percent to 30 percent market share: “There may be some other players in the market who don't have that unique ability to build their own media, build out their own heads, drive that areal density improvement. And they have to compensate that by potentially maybe increasing the number of units just to maintain their share and position in the marketplace.” He reckons vertical manufacturing integration gives WD (and Seagate) the edge over Toshiba, which uses some third-party components; platters and heads for example.
Tan said WD had excellent visibility into demand durability: “if you take our top 7 customers as an example, we have good visibility up to 2030, 2031 for … all of them. …we see quarter after quarter, the demand signal actually increasing over time. And as we move forward into the '28, '29, '30 time frame, we see that demand continuing to be very durable.”
WD can keep raising prices, Sennesael saying: “We continue to operate in a very strong, favorable pricing environment. …The pricing environment is strong because we do execute on our technology and product road map, and we continue to deliver higher capacity drives to our customers that help our customers to reduce their total cost of ownership because you have better rack density, lower power consumption and a lot of benefits for our customers.”
“As we continue to execute on that strategy and continue to deliver higher capacity drives, there is an opportunity to further enhance our price per terabyte over time.”
Sennesael said: “For our fiscal second quarter of 2027, at the midpoint of the ranges, we expect revenue of $5.1 billion (+/- $100 million), non-GAAP gross margin of 60.5 percent, and non-GAAP EPS of $5.80.” That would be 70 percent revenue increase on the year-ago Q2, a heck of a jump.
]]>Solidigm parent company SK hynix kiboshed the notion in a statement saying it “will halt its review of potential pre-IPO and IPO options.” It had “explored various funding options, including a potential IPO, to finance a new fab in the U.S. to meet growing AI memory demand.“
It added: “While an IPO was expected to ease investment burdens and expand capacity for shareholder returns, the review has been discontinued in light of market conditions and share price trends. The need to invest in Solidigm’s growth remains unchanged, and alternatives to enhance shareholder value will continue to be explored.”
Solidigm currently gets the NAND for its SSDs manufactured at a Dalian plant in China, which produces 192-layer 3D NAND, outdated by modern 3D NAND which is at the 270+ layer count and knocking on the 300+layer door. There are two fabs at Dalian, one producing the 192-layer NAND and the other a newer one being equipped after a pause when it was first constructed due to the then poor NAND market. Those days are over of course with AI server demand sending NAND requirements sky high, with NAND builders profits following in their wake.
Equipping the second Dalian plant is being restricted by US technology export restrictions aimed at preventing China getting access to the latest semi-conductor manufacturing technology. Reports say SK hynix has been authorized to install 240-layer NAND equipment at Dalian Fab 2 but that is still two, and soon three, generations behind the latest technology.
The US Trump-led government wants to build up US-based NAND manufacturing and SK hynix has been considering setting up a Solidigm fab in the USA - which will cost $15 to $30 billion. A Solidigm IPO idea was reported in September. This would have emulated parent SK hynix’s own US Nasdaq market entry in July, using American Depositary Shares.
SK hynix has now admitted that: “To secure the funding needed for Solidigm’s growth and ensure efficient capital allocation, SK hynix has been evaluating various options, including direct investment in Solidigm and raising external capital through an IPO.”
Reports said Goldman Sachs and Morgan Stanley had been identified as lead underwriters, but that has come to naught.
A contributing factor to the IPO shelving has been concerns by South Korean SK hynix stockholders that a Solidigm IPO would dilute their shareholdings, making them less valuable.
Also, SK Hynix said: “While SK hynix and Solidigm continue to deliver solid business performance, financial market volatility has increased, contributing to a broader correction in semiconductor stocks.” So: “After comprehensively assessing these business conditions, Solidigm has decided to discontinue its review of potential pre-IPO and IPO options.”
Which leaves SK hynix unable to use the Solidigm IPO route to get funds to build a Solidigm fab in the USA.
Although SK hynix is making massive profits it has to pay for increased high-bandwidth memory (HBM) fabs and packaging facilities. HBM is vastly more profitable than NAND, and in very high demand from customers like Nvidia. SK hynix has mentioned the capital efficiency idea. It would appear that a dollar invested in HBM production expansion will bring in a greater return than a dollar invested in a NAND fab, but NAND is still profitable and it wants to have its HBM cake while eating its NAND cake at the same time.
Perhaps it will look for a joint venture to share the cost of a US Solidigm fab.
]]>…
ConnectWise shut down Cloud Backup after attackers used a stolen vendor credential to access 173 Microsoft 365 customer tenants on October 15. They also pulled SaaS Security and Cloud Management as a precaution. That means MSPs relying on those tools could not back up data, reach archives, or restore files. As of Oct 22, :”As previously reported, we identified unauthorized access to application credentials associated with Cloud Backup, formerly SkyKick. Our teams and independent specialists continue advancing phased restoration, completing the security validation required at each step. We will share service-specific availability and reconnection instructions as they are confirmed. Previously published findings and partner guidance remain current. We recognize the continued disruption to your business and customers and remain focused on restoring service safely. Keep affected connections disabled until formal instructions apply to your service.”
“Existing backup archives remain encrypted, isolated and preserved, with no indication of compromise.”
Expect to wait several more days until normal service is restored.
…
Couchbase, calling itself the Operational Data Platform for AI, today announced an update of its AI Data Plane with new capabilities including native Amazon Bedrock integration with semantic caching, managed Agent Memory and a managed MCP server on AWS, in addition to vector search within Couchbase Operational Insights. Together, these innovations give enterprise teams a single governed platform to replace fragmented data services that slow agent deployments, reducing both infrastructure burden and the cost of running production AI at scale.
Couchbase unifies key-value, SQL, full-text search, vector search, and real-time columnar analytics on a single JSON data set, so agents get operational and historical context in one call instead of a round trip through separate systems.
…
DDN has a new CMO, Adam Berlew, and says he brings experience across cloud, AI, hardware, software and services, with a career spanning marketing, go-to-market and product leadership roles at companies including Equinix (CMO), Atlassian, Google Cloud, Broadcom and Dell. He will lead DDNs global marketing organisation as it looks to strengthen its brand, accelerate customer acquisition and connect more organisations with infrastructure designed to support AI workloads. Prior CMO Jyothi Swaroop left in April to be SMO at ChargePoint.
…
Collaboration supplier Egnyte announced announced Context Studio and AI Control Center; two new platform capabilities designed to help organizations improve how AI uses business context while strengthening control over agent access. Context Studio enables teams to visualize, manage, and optimize the business context behind their AI agents' improving accuracy and reducing AI costs. AI Control Center gives IT administrators a single dashboard to discover every AI agent connected to company content, see what sensitive data each one can reach, and restrict or re-scope that access.
To learn more, visit the Context Studio and AI Control Center webpages or watch the State of Egnyte: Fall 2026 virtual event for a detailed walkthrough of Context Studio, AI Control Center, and Egnyte’s AI journey.
…
JEDEC has established a new task group focused on advancing industry standards for LPDDR persistent memory, with MRAM developer Everspin chairing the effort. The group will focus on simplifying integration and enabling broader adoption of persistent memory across compatible system architectures. Everspin's Steffen Hellmold, VP of corporate strategy, is chairing the task group and brings more than 30 years of experience participating in JEDEC standards activities.
The work is intended to enable controller support for both DRAM and persistent memory, as well as approaches that enable LPDDR DRAM and persistent memory from multiple vendors to operate within compatible system architectures.
…
Austrian banking IT Services provider Raiffeisen Informatik GmbH & Co KG, has implemented a cyber resiliency architecture for its mainframe environment in partnership with Hitachi Vantara. Powered by Hitachi Vantara’s Virtual Storage Platform (VSP) and based on immutable snapshots, the solution reduces recovery time in the worst-case scenario — the total loss of all storage systems across both data center sites — from an estimated seven days to between one and two hours. It also limits potential data loss to two hours.
Raiffeisen’s mainframe environment processes approximately one billion transactions per month for more than 1.25 million end customers. Raiffeisen Informatik operates the IT infrastructure for central institutions of the Raiffeisen Banking Group, including Raiffeisen Bank International (RBI), Raiffeisenlandesbank Niederösterreich-Wien, Raiffeisenlandesbank Burgenland, Raiffeisen Landesbank Vorarlberg, and the UNIQA insurance group. Before the project, Raiffeisen Informatik operated two production storage systems across two data center sites, supported by daily FlashCopies and batch backups written to a virtual tape library. The architecture met the four-hour recovery time objectives for standard failure scenarios.
Now the firm can cope with an attacker destroying all storage systems across both data centers simultaneously.
…
IBM announced Q4 fy2026 results with $17.1 billion revenues, up 5% Y/Y, and $2.2 billion GAAP net income.
Software BU revenue up 8 percent to $7.8 billion.
Consulting BU revenue flat at $5.3 billion.
Infrastructure BU revenue of $3.7 billion, up 4 percent,
It said storage hardware, in the distributed infrastructure segment - where revenues grew 32%, delivered record revenue growth for its Infrastructure BU, gaining market share across the portfolio. CEO Arvind Krishna said: “AI continues to increase demand for storage and we are gaining share through differentiated offerings across Flash, Fusion, tape, and IBM Z storage, including AI-enabled capabilities that help clients scale and manage data for AI.”
CFO Jim Kavanaugh said: “Strength across Storage and Power reflects the growing importance of differentiated infrastructure as clients modernize their environments, address data growth, and support AI at scale. With a sizable backlog in Distributed Infrastructure exiting the quarter, we are well positioned for continued momentum in the fourth quarter. “
In the software BU Kavanaugh said: “Data grew 20%, driven by another strong performance from Confluent, with bookings growth of about 20%.”
…
A labor union in Taiwan carried out the first-ever strike there, with a small number of workers at Micron’s Taoyuan plant, wanting a profit-sharing scheme for employees. This follows a similar union-employer dispute in South Korea where Samsung agreed to a profits sharing deal with employees. Micron said it wanted to maintain a communications dialog with its employees.
…
mountOS is a cross-platform, forkable, time-traveling distributed filesystem over S3-compatible object storage. It provides native, POSIX-compliant filesystem access across macOS, Windows, Linux, and Kubernetes, allowing applications to work directly with object-backed storage without requiring a complete local copy. Unlike traditional fixed-capacity volumes, mountOS uses object storage as its underlying storage layer, with local SSD caching and replicated write handling for low-latency access. It also supports point-in-time filesystem views and writable forks. The system is designed for deployment on infrastructure operated by cloud providers, storage vendors, and enterprises. More information here.
…
Graph database supplier Neo4j announced its Context Engine, designed to give AI agents a shared understanding of business data and relationships across systems, enabling more accurate and trustworthy enterprise AI. It creates a shared knowledge layer that helps AI agents understand relationships between data and retain knowledge from previous interactions. Capabilities include:
Business Ontology – creates a structured map of an organisation's data, processes and relationships, helping AI agents understand how different parts of a business connect
Context Build and Runtime – connects existing business data sources to the shared knowledge layer, allowing AI agents to access relevant context when carrying out tasks
Context Studio – provides a central platform to build, manage and govern the knowledge and workflows used by AI agents
Context Memory – gives AI agents persistent memory of previous interactions and decisions, allowing them to reuse knowledge and improve their reasoning over time
Neo4j Context Engine is now available in private preview. Learn more here.
…
Vector database and search company Pinecone added on-demand mode to Pinecone Bring Your Own Cloud (BYOC). It follows BYOC becoming generally available on AWS, Google Cloud and Azure three weeks ago. BYOC runs Pinecone inside a company’s own cloud account, so the data stays with the company while Pinecone operates the database. Teams in legal, healthcare and financial services use it for exactly that reason.
…
Data management supplier Precisely announced the Precisely Automate MCP Server and native OData support in Precisely Automate. This release enables AI assistants and applications to interact with SAP automation workflows through a standards-based interface, while supporting SAP Clean Core requirements. OData is SAP's preferred API standard for connecting applications and exchanging data, allowing AI applications and agents to interact with SAP through modern integrations.
The Precisely Automate MCP Server, built on the open Model Context Protocol (MCP), exposes existing Automate Studio scripts and Evolve workflows as AI-callable tools — giving any MCP-compatible AI platform a standards-based connection to SAP automation. One configuration works across Claude, ChatGPT, Microsoft Copilot, and other AI platforms that adopt the protocol, without rebuilding integrations as the AI landscape evolves. Learn more here.
…
Reuters reports Samsung and SK Hynix are upgrading their China fabs to produce NAND flash chips with 200+ layers, aiming to fend off fast-growing Chinese rival YMTC.
Samsung (Xi’an plant): This facility accounts for roughly 40% of Samsung’s global NAND output and currently produces chips around the ~200-layer range. Under an internal project called “Monami,” Samsung is converting it to produce 286-layer V9 NAND. Targets cited by sources: ~5,000 wafers/month by December 2026, ramping toward 40,000 wafers/month by December 2027. Samsung has said the upgrades are to meet demand (including from U.S. customers), comply with U.S. rules, and do not expand overall capacity or add new facilities.
SK Hynix (Dalian plant): Acquired via the 2021 purchase of Intel’s NAND business (now Solidigm). It currently makes ~192-layer chips. Plans are to move to 240 layers or more. Sources indicate it has U.S. approval for up to 240 layers there and may seek further approval for R&D on 300+ layer technology (which could support future U.S. production). This site is Solidigm’s only NAND manufacturing base and is relevant to its planned U.S. IPO.
…
Sandisk announced the Guardian Series portfolio of high-endurance storage designed for use with 24x7 compatible home security cameras, dash cams and built-in monitoring systems. They enable continuous recording with the durability to withstand harsh environmental conditions, and the continuous overwriting cameras require. The new portfolio includes:
SANDISK Dash Cam Card – Engineered to work with a wide variety of dash cams, the SANDISK Dash Cam Card is purpose-built for continuous, high-endurance recording in demanding driving conditions. The card supports high-res recording from Full HD up to 4K UHD and is engineered to withstand heat and humidity1 on even the longest drives. It is available in capacities ranging from 32 GB to 512 GB, with the 512 GB2 cards tested to endure up to 40,000 hours of recorded drives.
SANDISK Security Cam Card – Built to work with a wide variety of security cameras, the SANDISK Security Cam Card is engineered to operate continuously alongside – or in lieu of – cloud storage to help keep your footage even when the Wi-Fi drops. Available from 32 GB to 512 GB, it is temperature, water, and dust proof for year-round outdoor camera systems and supports high-res recording from Full HD up to 4K UHD1 for crystal clear video.
The SANDISK Guardian Series memory cards are available now from select retailers starting at $28.99 U.S. MSRP for 32GB.
…
Symply announced GA of SymplyATOM2, the next generation of its free LTO tape drive management and diagnostic software. Built for macOS and Windows, SymplyATOM2 provides an accessible set of tools for managing, maintaining and troubleshooting Symply LTO tape products. Its graphical interface is designed to help people who are new to LTO, whilst also providing the detailed diagnostics and controls required by experienced tape users and support professionals. SymplyATOM2 can be downloaded from the Symply Support Portal.
…
Data warehouser Teradata announced a portfolio of deployment and platform capabilities designed to give governments and enterprises greater control over the data, AI, and technology that power their most critical operations. This includes expanded capabilities for Teradata Factory with AI Studio and the availability of Teradata Sovereign Cloud on Google Distributed Cloud air-gapped environment. Teradata Factory with AI Studio capabilities include:
Tera, the agentic coworker for enterprise data work, including Tera Harness and Tera Context Engine
AgentStack, the complete AI agent development capabilities to build, manage, observe and operate agents through natural language, pro-code, low-code, and Model Context Protocol (MCP) interfaces
Model services, which provide machine learning operations and AI model management with pre-integrated model and bring-your-own-model support
Data pipelines for ingesting multimodal unstructured data, vector embeddings, vector storage and hybrid search for RAG
Analytics, native notebooks backed by in-database analytics and Teradata's Elastic Python Compute, built on Ray and Arrow, bring AI to where data lives
Teradata Sovereign Cloud on Google Distributed Cloud air-gapped brings Teradata’s high-performance enterprise-scale analytics into fully isolated environments. The expanded Teradata sovereign capabilities for Teradata Factory with AI Studio will be available in Q1 2027. Teradata Sovereign Cloud on Google Distributed Cloud air-gapped is available in Private Preview.
…
Mirantis is a B2B open-source cloud computing software and services company that provides platforms for managing Kubernetes, containers, OpenStack, and AI infrastructure across hybrid, multi-cloud, on-premises, and edge environments. It announced a joint integration with VAST Data to simplify how neoclouds and AI infrastructure deploy and operate secure, multi-tenant AI infrastructure at scale. By bringing together the Mirantis k0rdent AI infrastructure orchestration and the VAST AI Operating System, the companies are creating a repeatable operational model spanning GPU compute and the data services that power AI workloads.
Previously, deploying compute and storage for an AI cloud tenant required operators to coordinate separate infrastructure systems, bare-metal servers, Kubernetes clusters, storage, and tenant access. Identifying network and storage details, applying specialized domain expertise and manually maintaining isolation between tenants was labor-intensive and introduced multiple opportunities for configuration errors.
With the Mirantis and VAST integration, k0rdent AI orchestrates the infrastructure so storage is automatically created, configured, scaled and retired alongside tenant compute resources.
…
Veeam announce GA of Veeam Intelligent ResOps for Microsoft 365, the first generally available offering in the Veeam Intelligent ResOps (iResOps) portfolio and an add-on for Veeam Data Cloud for Microsoft 365. Customers connect Veeam iResOps for Microsoft 365 to Veeam Data Cloud in a few clicks, with no data migration and no re-architecture required. They get new visibility across their Microsoft 365 estate, identifying known and unknown AI agents, AI-driven file changes, discovering Redundant, Obsolete, and Trivial (ROT) data, and locating any sensitive and regulated data to validate protection. With clear visibility across their Microsoft 365 data and AI estate, organizations can safely accelerate their use of AI, reduce storage costs, improve compliance, and recover data with precision. Veeam is also announcing Private Previews of Veeam iResOps for Veeam Data Platform and Okta, respectively.
Veeam Intelligent ResOps for Microsoft 365 is generally available today as an add-on to Veeam Data Cloud for Microsoft 365, in Foundation and Advanced editions, and is also available through Veeam Cloud & Service Providers (VCSPs). Veeam iResOps for Veeam Data Platform and Veeam iResOps for Okta are available now in Private Preview to qualified customers. Private Previews for GitHub, Azure DevOps, and Google IAM begin in Q4 2026.
…
Virtuozzo Infrastructure System (V/IS) 8.0 is the latest major release of Virtuozzo’s infrastructure platform (formerly called Virtuozzo Infrastructure). It is SW-defined infrastructure for service providers, SaaS companies, channel partners, and enterprises to deliver virtualization, storage, networking, compute, and AI/GPU workloads. V8.0 enables smaller deployments for customers who need only one or two servers, using local storage without a shared storage network. Plus:
Expand into AI with updated Nvidia support. Dedicate GPUs to workloads or share them across virtual machines, with passthrough and vGPU on the same host.
Simplify VMware migration. The built-in migration tool is now generally available, supporting vSphere 8 without VMware’s VDDK.
The release also brings a RHEL 10 operating system support, simpler setup with new onboarding wizard, improved object storage performance, and greater control over API access. Virtuozzo positions V/IS 8.0 as a more flexible and often lower-cost alternative to VMware.
…
ReRAM developer Weebit Nano has published a technical paper on how ECC is implemented and used in ReRAM-based systems. It goes into the specific failure mechanisms it sees in ReRAM, how they differ from flash, and what this means for selecting the right ECC approach in real products. Access the paper here.
…
xMEMS Labs announced the start of mass production of its solid-state XMC-2400 and XMC-4800 µCooling chips.
]]>Nebul, a neocloud or AI CSP, based in Leiden, Holland, was founded in 2022 by CEO and data center specialist Arnold Juffer. It offers a full-stack and sovereign cloud with provisioning of CPUs, DPUs, GPUs, and APUs, with 1-click deployment, managed Kubernetes, managed databases, and Nvidia Hopper (H100/H200) and Blackwell (B200 and later) GPUs. These are hosted in energy-efficient European data centers and can scale from single GPUs to large clusters (hundreds to over 1,000 nodes in some descriptions). Nebul raised approximately €20 million ($22.4 million) in mid-2024 to expand its sovereign AI cloud and data centers across Europe, support hiring, and meet growing demand from enterprises and governments. Infinia is DDN’s object-based OS, and a companion to its traditional Lustre parallel filesystem software.
DDN co-founder and CEO Alex Bouzari said: “Every time a GPU rebuilds context it has already processed, it is spending cycles repeating work, By making that context available across Nebul’s GPU fleet, Infinia helps deliver faster inference and more productive use of GPU capacity while keeping data within the sovereign environment.”
We’re told DDN Infinia extends Nebul’s data controls to the KV cache used during inference. Tenant, subtenant and dataset scope remain attached to the stored context, while cache capacity can be managed without interrupting inference.
Juffer said: “With DDN Infinia, we can reuse model context across our GPU infrastructure instead of spending valuable GPU cycles recomputing it. That makes inference faster and more efficient while keeping the complete AI workload within the sovereign environment.”
KV cache testing covered workloads ranging from short prompts to large-token contexts under concurrent demand. Nebul measured results on its own platforms, with the same eight GPUs, model and workloads on both sides. Some of the results:
Nearly 10x faster time to first token with DDN Infinia than Nebul’s own cache on CPU and memory.
Up to 5x faster token responses than alternative offerings.
2.5x more input tokens per second from the same GPUs.
Nearly 60 percent lower cost per million input tokens, by reloading context instead of recomputing it.
However, Infinia is not Nebul’s only storage OS, it being promiscuous in a storage sense, and listing NetApp’s ONTAP, DDN ExaScaler and VAST Data as specific storage options, now joined by Infinia.
An April 2025 VAST success story said “Nebul is anchored by a multi-petabyte VAST Data Platform Cluster, which provides the basis of performance and scalability needed to support their clients’ rapidly growing AI workloads.”
Nebul's CRO Bam Gobets, now at Regatta Data, said: “We’ve carefully selected VAST Data for our cloud platform and data management backend. VAST Data allows us to provide the best performance needed for properly saturating GPUs, and seamlessly scaling customer solutions for model training, production inference, and all stages of data processing across the AI pipeline.”
Juffer added: “The modern architecture of VAST was essential to us. We understand the market will end up where VAST is going.”
Apparently not. NetApp said in a September 2025 Nebul customer story that Nebul offered GenAI services like private GPT, agentic AI, private inference APIs, and RAG – supported we understand by NetApp AFF A90 storage arrays running ONTAP. Its “DataOps Toolkit, multi-protocol functionality, and especially the S3/NFS duality provides extreme value to Nebul’s business processes as well as the flexibility and data pipeline integration Nebul can offer its end customers.”
A video by NetApp praises Nebul as well.
A final point; Nebul is positioning Infinia as data lake storage.
The name Nebul come from nebula, meaning a cloud or mist as well as a bright cloud of stars in astronomy.
]]>The company is currently shipping G9, its ninth generation, with 276 layers. The first product using this tech, the 2650 client SSD, was launched in July 2024. Micron held a Technology and product webinar on its Investor Relations website on October 21. Its President and Chief Technology and Products Officer Scott DeBoer discussed Micron's technology foundation, multigenerational roadmap, customer co-design, and long-term research positions in DRAM, HBM, and NAND.
DeBoer said in the NAND section: “We've been in a very strong position with our Gen 9 NAND, and this leadership is going to extend with Generation 11. And this is a significant leap. We have optimised this memory. We've actually gone back and ensured that this memory is optimised for the AI products of the future and intentionally, which is why it goes from Gen 9 to Gen 11. We've made substantial improvements that this node really brings relative to any prior NAND.”
G11, compared to G9, has twice the bit density, an up to 1.5x increase in I/O speed and needs 25 percent less energy to transfer a bit.
DeBoer said: “This G11 technology is the world's most advanced NAND. It will ramp in the second half of next year. Also, it's based on wafer on wafer bonding technology … We’ll be at more than 400 layers and we'll give more specifics on that next year closer to the launch date.”
He also discussed future NAND product technologies that could use G11 3D NAND. One was a higher-speed SSD with smaller block sizes enabling much higher IOPS. The second was a higher bandwidth SSD, with stacked layers connected to a base die by through silicon vias (TSVs) like high-bandwidth memory (HBM), and similarly placed next to an accelerator processor (GPU).
DeBoer said: “There's really two focus areas here. One is on high IOPS, AI class NAND, which is [an] SLC-based design implemented in small blocks to design ultra fast memory storage products. And this is a key piece that meets several emerging use cases noted here and really is a major focus for Micron in terms of getting these products out the market over the next couple years. On the other side of this architecture and design path is high bandwidth AI-class NAND focused on other emerging use cases. And these ultimately are more costly and have to be implemented in partnership with customers in a way to ensure that they're meeting the mark and the cost point to drive the value proposition.“
The high-bandwidth NAND would not be positioned along the accelerator’s shoreline like HBM, “as we don't think that has merit, but along closer positioning points to the compute itself to make use of this high performance NAND and the viability of that path.”
He did not mention the very similar high-bandwidth flash (HBF) technology being promoted by Sandisk and SK Hynix. Both have stacked layers of 3D-NAND dies above a base logic or controller die and a connection to a nearby accelerator, possibly/probably by using an interposer.
The high-IOPS SSD would be in the 25 - 100 million IOPS area while next-generation SSDS could hold up to 512 TB. Like the high-bandwidth AI-class NAND, the high-IOPS NAND would also be located close to the accelerator.
DeBoer also mentioned future Micron AI-class NAND being used in AI PCs and workstations, automotive, robotics, mobile and wearable products.
BNP Paribas Equity Research senior analyst Karl Ackerman told subscribers that Micron's HBM roadmap runs through HBM6, with vHBM and vEIM starting production In 2028: "Micron noted that HBM4 36GB 12-high entered high-volume manufacturing in March 2026, and HBM4E is currently sampling with key customers. The company targets high-volume manufacturing of HBM4E in 2027, HBM5 in 2028, HBM5E in 2029, and HBM6 in 2030.
"Additionally, Micron announced vertical HBM (vHBM), which places the DRAM stack directly on top of the processor rather than beside it, delivering greater bandwidth than conventional HBM. The trade-off is this inhibits the cooling of the processor, which is why we expect vHBM to be deployed in conjunction with traditional HBM. Meanwhile, vEIM (vertical edge inference memory) applies the same stacked structure to edge devices, placing low-power LPDDR on top of a logic die, and Micron’s first-generation vEIM targets 2TB/s of bandwidth at less than 1 pJ/bit (this approach is similar to what d-Matrix has pioneered). Both new lines are targeted for volume production in 2028 and have three generations on the roadmap through 2030."
And: "Micron built the HBM4 base die on its own CMOS process, while competitors transitioned to foundry-made base dies. However, HBM4E’s speed and power targets require more advanced CMOS, so Micron is now moving to foundry-made base dies. HBM4E is also the first generation for which Micron offers custom HBM alongside the JEDEC-standard product, which we think includes Nvidia’s NVHBM. For custom parts, Micron and customers jointly design the base die, integrating the customer’s logic IP with Micron’s memory IP."
Gen 1 - 32 layers
Gen 2 -64 layers
Gen 3 - 96 layers
Gen 4 - 128 layers
Gen 5 - 176 layers
Gen 6 - 232 layers at end 2022 - branded G8 externally in Nov 2024
Gen 7 - 276 layers - branded G9 externally
Gen 8 - planned 300+ layers, branded G10 externally but abandoned in Oct 2026
Gen 9 - planned 400+layers, branded G11 externally Oct 2026.
We think this is marketing-based and a response to Samsung having its V9, 10 and 11 generations, SK hynix with V9, 10 and 11 ditto, and Sandisk/Kioxia with their BiCS 9, 10 and 11.
According to Imec, Wafer-on-wafer (WoW or W2W) bonding is a semiconductor 3D integration technique in which two (or more) fully processed silicon wafers are aligned and bonded together as complete wafers, after which the bonded stack is diced into individual multi-die chips.
]]>This Spaceborn Computer-4 (SBC-4) should travel to the moon as part of Astrolab’s FLIP rover for an upcoming mission for the Moon Base program.
HPE has so far sent three SBC generations to the International SpaceStation (ISS) and is now, literally, going further by having one operate on the moon’s surface. SBC-4 will be part of a mission that’s intended to launch in late 2026 aboard a SpaceX Falcon Heavy from Launch Complex 39A at Kennedy Space Center, Florida, as part of NASA’s Commercial Lunar Payload Services (CLPS) initiative. After touching down at the Nobile Crater region near Mons Mouton at the lunar south pole, Voyager Technologies’ Griffin-1 lander will transport Astrolab’s FLIP rover, carrying HPE Spaceborne Computer-4 as a payload.
Scott Wiest, VP, fellow and global sales chief technologist, HPE, said: “For more than 70 years, HPE innovation has supported some of the world’s most important space missions. The next chapter of lunar exploration will depend on the ability to turn data into insights at the point of discovery. With HPE Spaceborne Computer-4, we are bringing AI and high performance computing to the lunar edge, helping the acceleration of scientific discovery and building the infrastructure for the next era of exploration.”
Griffin-1 is a commercial lunar lander developed by Astrobotic Technology, now operating as Voyager Lunar Systems under Voyager Technologies. Voyager acquired Astrobotic in 2026; the lander is built at facilities in Pittsburgh, Pennsylvania. It will carry more than 1,100 pounds (455 kg) of cargo to the lunar surface. Astrolab's FLIP rover will be the largest commercial payload ever sent to the lunar surface.
The lander will carry about 10 scientific and demo payloads, including Astrolab’s FLIP rover (FLEX Lunar Innovation Platform).This is a ~500 kg technology demonstrator and precursor to Astrolab’s larger FLEX rover. It carries instruments and hosts systems like HPE’s Spaceborne Computer-4. FLIP is designed to test mobility, batteries, tires, avionics, sensors, and software for future lunar vehicles.
Jaret Matthews, founder and CEO of Astrolab, said: “Our FLIP rover is designed to accommodate a variety of payloads, and this mission will demonstrate how commercial lunar mobility platforms can support technology development and scientific research.”
The FLIP rides on top of the Griffin lander, stacked on a cone/platform with ramps. Assuming a successful soft landing near the lunar south pole, the rover will reach the moon’s surface like this:
It separates from the lander and begins a controlled descent down steep ramps.
A fabric tether system wrapped around the rover’s rear wheels provides the control, allowing it to “rappel” or carefully lower itself down the incline onto the lunar surface. This approach was developed by Astrolab and its partner Venturi to handle the constrained geometry of the Griffin lander and various possible landing orientations/tilts.
Once on the surface, FLIP drives away to begin its operations.
There’s a great animation on the Astrobotic website.
The SBC-4 is purpose-built as a compact (~249 × 132 × 26 mm), lightweight (~2.6 lb) aluminum, self-contained, and fanless compute module, designed to survive the harsh environment of the lunar South pole. It is equipped with both a low-power CPU and GPU. The compute module utilizes Kioxia SSD flash storage, specifically BG6 series client/M.2 NVMe SSDs; specifically two 1 TB models (mirrored via software RAID 1 for ~1 TB usable capacity).
These are not radiation-hardened products, being 2023-era, commercial off-the-shelf client SSDs, using the M.2 form factor, he host computer’s DRAM, and a PCIe gen4 bus. The FLIP rover has a limited power budget and these drives only draw 5W.
The SBC-4 will run Large language models (LLM) to deliver insights, analyze visual data, and support autonomous monitoring at the lunar edge. Tests will process LiDAR data for mapping, navigation, terrain analysis and infrastructure development, while an AI services container will demonstrate reliable deployment of scalable AI service applications on the Moon. The HPC side of things involves high-performance LINPACK (HPL) and high-performance conjugate gradient (HPCG) calculations being run.
SBC-4 software optimizes the required interaction with Earth-based resources and enables secure delivery of insights and results from the Moon back to Earth.
NASA and Deloitte Consulting are involved in SBC-4 workloads as well:
NASA will run an AI neural net model trained on data from Earth, the Moon, and on the International Space Station, demonstrating coordination of three separate AI models, in a Federated Learning for Unbounded and Intelligent Decentralization (FLUID) experiment.IT will also conduct a tuned LLM test with semantic processing abilities to assist in searching and interpreting onboard documents.
Deloitte's Silent Shield, a new system to detect cyber intrusions in space, will provide near real-time, predictive cyber-analytics and anomaly detection.
Cyber intrusions on the moon seem a tad unlikely and the Deloitte contribution appears somewhat premature.
HPE will support Astrolab’s mission control in Hawthorne, California using HPE Private Cloud to support real-time insights, delivered in a turnkey solution. HPE CloudOps Software will provide a unified, flexible environment for managing and automating workloads.
HPE’s Spaceborne Computer program has so far sent three systems to the International Space Station (ISS).
Spaceborne Computer-1 (SBC-1, also just called Spaceborne Computer)
Launched: August 2017 on SpaceX CRS-12.
Hardware: Commercial off-the-shelf (COTS) HPE Apollo 40-based HPC server system. No traditional radiation hardening; protection relied on software monitoring that could throttle or adjust the system under stress.
Purpose: Proof-of-concept experiment to show that unmodified (or lightly adapted) commercial servers could survive launch, operate reliably in space, and run scientific workloads. It achieved roughly 1 teraflop and became the first teraflop-class system in space.
Duration: Operated for 615 days (well beyond the planned one-year mission) before returning to Earth in 2019. It demonstrated that software-based resilience could handle cosmic radiation and other space hazards.
Spaceborne Computer-2 (SBC-2)
Launched: February 2021 on Northrop Grumman NG-15.
Hardware: Combination of an HPE Edgeline EL4000 (edge-focused system with CPU + GPU) and an HPE ProLiant DL360 Gen10 (traditional 2-socket HPC node). Later storage upgrades added substantial Kioxia capacity.
Purpose: Expanded edge computing and AI capabilities on the ISS. It enabled real-time processing of experiment data (e.g., image processing, DNA sequencing, healthcare, 3D printing, disaster recovery), dramatically reducing the volume of data that needed to be downlinked to Earth (sometimes by orders of magnitude). It ran dozens of research experiments and supported cloud bursting to Microsoft Azure in some cases.
It proved more serviceable and capable than the first system and stayed operational for years.
Spaceborne Computer-3
Launched: January 30, 2024, on Northrop Grumman NG-20 (SS Patricia Robinson).
Hardware: Updated Edgeline + ProLiant configuration with a major storage expansion - over 130 TB of Kioxia SSDs (including 8× 1 TB XG6 NVMe M.2, 4× 960 GB RM6 SAS, and 4× 30.72 TB PM6 enterprise SAS drives). This was the largest amount of storage sent to the ISS on a single mission.
Purpose: Further advance AI/ML workloads, larger datasets, and edge computing for scientific research on the ISS. HPE describes this as the program’s third journey/iteration to the station, building directly on the successful SBC-2 platform with software and storage improvements for more complex applications.
All three ISS systems used commercial off-the-shelf HPE servers (with software techniques for radiation resilience rather than specialized radiation-hardened chips) and demonstrated that modern computing can operate effectively in low Earth orbit inside a space station. Unlike the space environment outside the ISS, the interior has atmospheric pressure, temperature control, and insulation from solar radiation. Spaceborne Computer-4 is the next step: a much smaller, lower-power, fanless module designed for the lunar surface on Astrolab’s FLIP rover. There the environment is a vacuum and cold.
Sunlit surfaces near the south pole can reach roughly 200–300 K (−73 °C to about +27 °C / −100 °F to +80 °F), depending on slope, season, and exact location. Nighttime / shadowed areas drop much lower, often to ~40–100 K or below.
HPE claims “70 years of leadership in space exploration.” We understand this is based on:
1957: HP instruments (including the HP 560A Digital Recorder) were used by the Stanford Research Institute to track Russia’s Sputnik 1, the first artificial satellite. This was before NASA existed.
1958: HP equipment supported Explorer 1 (the first U.S. satellite) and the Pioneer Station deep-space antenna. NASA was founded that year.
1960s (Mercury, Gemini, Apollo): HP supplied medical recording systems for astronaut vital signs (e.g., on Gordon Cooper’s Faith 7 Mercury flight in 1963), digital voltmeters and oscilloscopes for Gemini, and critical cesium-beam frequency standards / atomic clocks that synchronized the Apollo tracking and communications network.
Apollo 11 (1969): HP diodes, PIN switches, and related components were used in the Eagle lunar module and in the astronauts’ communication systems during the first Moon walk.
1975: An HP-65 programmable calculator served as a backup for rendezvous and guidance calculations on the Apollo-Soyuz Test Project (the first international docking).
Later decades: HP calculators flew on Space Shuttle missions, HP medical equipment was used in microgravity research, and more recently HP/HPE systems (workstations, printers, and the Spaceborne Computer series starting in 2017) have operated on the International Space Station.
You might want to reflect if all of this constitutes “leadership in space exploration,” compared to, say, the Apollo moon landing program and SpaceX’ re-usable rockets. But SBC-4 on FLIP is stirring stuff, no question.
Astrobotic is a 2007 Carnegie Mellon University spinout. Voyager Technologies completed its acquisition of Astrobotic in July this year, for up to roughly $300 million ($162 million in cash and stock upfront, up to $129 million in performance-based earnouts, plus assumed debt). The unit now operates as Voyager Lunar Systems, with former Astrobotic CEO John Thornton continuing in a leadership role.
The core of Astrobotic’s funding has been NASA (especially CLPS task orders) and other U.S. government awards. That will be supplemented now by the resources of parent company Voyager Technologies.
Astrobotic has not had a successful Moon landing as of October 21, 2026. It’s Peregrine Mission One in January 2024 had a propellant leak shortly after separation from the Vulcan rocket, failed to reach the Moon, and re-entered Earth’s atmosphere after about 10–11 days. Here’s wishing it, and HPE, the very best of luck for FLIP.
]]>Nvidia OpenShell is an open, secure runtime for AI agents, governing agent execution, access, and where inference goes, giving agents the access they need to execute the tasks they’re sent. It combines sandboxed execution, controlled resource access, credential protection, and formal policy verification, with enforcement outside the agent workload. It's integrated with hardware-level security mechanisms based on the BlueField-4 chip. NemoClaw is an open-source collection of blueprints and a reference stack for deploying specialized, always-on autonomous AI agents more safely and with governance controls. Agent state persistence is the core feature of VAST’s new Foundation Stack for Governed Agent Sandboxes (FSGAS), announced at Nvidia GTC in Berlin.
Alon Horev, CTO and Co-Founder at VAST Data, said: “Every enterprise that puts an agent into production will eventually need to answer what it did and why it did it. An agent might exist for minutes, spawn thousands of other agents and then disappear, but its work has to persist and its actions have to stay understandable long afterward. Together with Nvidia, we’re showing how the VAST AI Operating System provides the durable foundation underneath secure agent environments, so organisations can scale agentic systems without sacrificing persistence, governance or control.”
A VAST blog states: ”Agent sandboxes are quickly becoming the de facto approach for implementing guardrails around AI agents. Sandboxes should prevent agents from going rogue and limit the blast radius of any questionable actions, while still providing agents access to all the tools they need.
Nvidia’s Kari Briski, VP of Generative AI Software, added more context: “As AI agents take on more complex work, enterprises need infrastructure that can support secure execution while preserving the data, context, and state agents depend on across their lifecycle. By combining Nvidia NemoClaw and OpenShell with the VAST AI Operating System, organisations can build agent sandbox environments that are secure and designed to scale from experimentation to production.”
FSGAS uses a new Agent Sandbox Foundation Stack, and VAST says agents can spawn other agents, execute tasks and disappear, with fleets creating and destroying thousands of isolated workspaces an hour. Long-running agents may also work across multiple sessions, sleeping, resuming and delegating tasks to short-lived subagents along the way. Whether an agent runs for minutes or across multiple sessions, its progress, working data and memory need to persist independently of any individual sandbox or compute node – remaining available, governed and auditable wherever execution resumes.
FSGAS provides a scalable, persistent data infrastructure that enterprises need to securely run thousands of AI agents and sub-agents in production. It works like this;
As new agent tasks are initiated, VAST uses zero-copy clones to provision persistent storage for isolated OpenShell sandboxes without creating full copies of the underlying data. VAST separates agent execution from persistent state so agents can pause, resume or transfer work across agent environments without starting from scratch.
VAST DataBase maintains a persistent record of agent activity, including tool calls, prompts, tokens, costs and decisions while the VAST Management System provides real-time visibility into data flows and infrastructure activity across agent fleets.
By decoupling persistent agent state from individual compute nodes, the FSGAS architecture allows agents to pause, resume, or move across eligible infrastructure without losing their progress, working data or context.
Previously introduced VAST foundation stacks include:
Enterprise research assistance with Nvidia Nemotron - AI-Q based Foundation Stack provides a foundation for building custom AI researchers that can operate across private, enterprise data sources, synthesising hours of research in minutes, using the VAST AI OS for persistent and secure context, scalable reasoning pipelines, and trusted agent execution.,
Video intelligence with Nvidia Cosmos - a Video Search and Summarisation (VSS)-based Foundation Stack enables organisations to ingest massive volumes of live or archived video and extract insights through semantic indexing, summarisation, and interactive Q&A,
VAST plans to make the automated deployment available through a public GitHub repository, providing a reproducible path for organisations to deploy the architecture in their own environments.
The aforementioned blog notes: “This foundation stack builds on the recently announced VAST DataEnclave, which delivers a trusted and cryptographically secured runtime for running leading commercial AI models alongside sensitive data on customer-controlled infrastructure.” Read more here; it's worthwhile.
HPE will support the Nvidia OpenShell Reference Architecture for Safe Agents, delivering full-stack AI factory governance, isolation, and security with Nvidia Vera Rubin, Nvidia Vera CPUs, and Nvidia BlueField-4.
]]>Nutanix will support Nvidia’s forthcoming Vera CPU with its NKP (Nutanix Kubernetes Platform), develop a jointly-engineered server system with Supermicro to run Nutanix’ agentic AI software, and work with Netris to combine its Nutanix Cloud Platform (NCP) with Netris’ Network Automation, Abstraction, and Multitenancy (NAAM) platform. This provides multi-tenant AI infrastructure with network automation, hard tenant isolation, and self-service network operations across bare metal and virtualized environments with NCP. So we might see a Supermicro Vera Rubin server running Nutanix software available through Netris neocloud and enterprise customers under an AIaaS deal.
EVP for Nutanix Product Management Thomas Cornely said: “Organizations need more than just advanced infrastructure to move agentic AI into production – we believe they need an AI cloud that simply connects intelligence, agents, enterprise applications, and data across CPUs, GPUs, virtual machines, containers, on-premises environments, and cloud. Through our collaboration with Nvidia, Supermicro, and Netris, we plan to help customers accelerate deployments of agents and AI services on a unified platform with built-in security, performance, and governance.”
Nutanix plans to work towards Day Zero support for Nvidia Vera CPUs, beginning with NKP Bare Metal, to serve highly performance-sensitive AI factory environments, with planned expansion to virtualized Nutanix environments. As enterprises adopt agentic AI at scale, they are anticipated to require infrastructure that can support increasingly demanding workloads across the full AI environment, not just GPU-intensive tasks.
Nutanix notes agentic applications combine inference (think GU) with tool calls, code execution, retrieval, orchestration, and outcome processing (think CPU). It says that, by pairing Nvidia’s purpose-built CPU architecture – engineered for high-speed agent execution and reduced latency – and OpenShell to create safe, private runtime environments for agents with NKP, the collaboration aims to give organizations a predictable, high-performance foundation for their most advanced AI agents.
Regarding the Supermicro deal, Nutanix says that, by combining Supermicro servers and Nvidia GPUs - think Vera Rubin - with Nutanix Agentic AI (NAI) and Nvidia Spectrum-X networking, enterprise customers reduce custom configuration and integration delays, get on-premises GPU capacity and take on runaway AI costs. Customers should get a predictable path to deploy, manage, and scale AI inference workloads without the bother of a complex, multivendor environment.
Nutanix and Netris are collaborating for neocloud-like customers to automate network-layer multi-tenancy and turn complex GPU infrastructure into secure, cloud-like, multi-tenant AI services.
Nutanix will provide a compute foundation across virtual machines, containers, CPUs, and GPUs, along with storage, security, Kubernetes, and Nutanix Agent Gateway, model-as-a-service and AI platform services. Netris will provide network automation, abstraction, and multi-tenancy across the physical network fabric, with hard isolation (enforced on networking hardware) from compute through the physical network.
As a result, AIaaS providers - AI CSPs - will be better able to onboard tenants and deliver AI services with a self-service operating experience across bare metal and virtualized AI infrastructure.
VMware, Nutanix’ main competitor, does not yet have any public Vera support plans, being x86 and not Arm-focussed; Vera being an Arm CPU. Nor does VMware have joint-engineering server deals with Supermicro, although VMware does certify Supermicro gear as vSAN-ready nodes and collaborate around VMware AI factory management.
VMware does not have a Netris partnership, but Netris documentation lists support for VMware vSphere.
]]>The company is a neural machine translation (NMT) service headquartered in Cologne and founded in 2009 as Linguee, an online dictionary, by CEO Jarosław “Jarek” Kutylowski. DeepL is a unicorn, valued at around $2 billion after a $300 million funding round in May 2024 led by Index Ventures. It has more than 200,000 business customers.
Jeff Denworth, a VAST Data Co-Founder, said: “DeepL is a great example of how quickly the infrastructure requirements behind AI are changing. As AI moves into production, infrastructure reliability becomes just as important as raw performance. Written language, real-time voice and the next generation of AI experiences all place different demands on data, but researchers should not have to redesign the underlying infrastructure every time the model or product changes. VAST AI OS was built to deliver that reliability at AI scale, giving teams a foundation they can trust, and then getting out of their way.
DeepL has several offerings:
DeepL Translator — Text and document translation (DOCX, PPTX, PDF, XLSX, etc.) that preserves formatting. Supports 100+ languages. Known for natural, human-like output, especially on European language pairs. Offers glossaries, tone control (formal/informal), and alternatives.
DeepL Write — AI writing assistant for improving grammar, style, tone, and clarity (available as a Pro add-on).
DeepL Voice — Real-time speech translation and captions for meetings (integrations with Zoom, Microsoft Teams, Google Meet) and conversations. Launched in 2024 and it now has advances enabling it to preserve characteristics of a speaker’s voice across languages.
DeepL API — For embedding translation, editing, and related features into apps, websites, and internal tools.
Emerging capabilities include workflow automation (e.g., Translation Flow) and mentions of a DeepL Agent for more autonomous language-related tasks.
It added AWS as a sub-processor earlier this year; all its translation code is containerized, making the port simpler. This addition was to to support global scale and performance. Now it has a hybrid infrastructure that combines DeepL administered servers in data centers located in Europe with cloud infrastructure provided by AWS.
When you use DeepL, the text you enter or upload is processed instantly to provide the services and is not stored. If you choose to save data such as glossaries or translations, it remains stored until you delete it. There’s more information about this here.
DeepL says it competes with general tools like Google Translate by prioritizing high-quality, natural output and business-ready features. As an example, consider the English word “oversight.” It has two overall meanings and seven possible words for them in German;
1. Supervision, control, or management
Aufsicht (general supervision, oversight of a project or process)
Überwachung (monitoring, surveillance, or regulatory oversight)
Kontrolle (control, checks, or audit oversight)
Aufsichtsrat (for an oversight board / supervisory board)
2. A mistake, omission, or slip-up
Versäumnis (an oversight or failure to do something)
Versehen (an accidental oversight / oversight by mistake: aus Versehen = by accident)
Fehler (error or mistake)
DeepL claims it deals with such contronyms - a word with two opposite meanings - better than other machine translation tools because it has a better understanding of context.
We understand DeepL currently runs large-scale AI training and inference on NVIDIA DGX SuperPOD systems hosted at EcoDataCenter in Falun, Sweden (100 percent renewable energy, liquid-cooled, AI/HPC-optimized):
DeepL Mercury (2023): DGX SuperPOD with 68 × DGX H100 systems (each with 8 H100 GPUs). Ranked on the Top500 list with ~21.85 PFLOPS Linpack performance. Supplied in part by DELTA Computer Products.
Arion (2025, first in Europe): DGX SuperPOD with DGX GB200 NVL72 systems (Grace Blackwell Superchips — 36 Grace CPUs + 72 Blackwell GPUs per rack, NVLink, InfiniBand). Enables much faster training and inference (e.g., translating the entire web in ~18 days vs. ~194 previously).
Arion is actually DeepL’s third major SuperPOD deployment and a substantial upgrade over Mercury. It is Europe’s first commercial DGX GB200 SuperPOD and underpins the company’s ability to train larger models and deliver faster, higher-quality language AI services.
These DGX systems include local high-speed NVMe storage for operating system, caching, and temporary data, but we don’t know what shared storage backend has been used for the large training datasets and checkpoints - up until now.
This is how we understand its general backend storage environment has evolved since 2017:
The Early On-Premises Architecture (2017): DeepL started with a highly centralized, bare-metal footprint with basic active-passive server configurations connected directly to local GPU worker nodes. There were local SSD arrays engineered to handle the high-throughput, low-latency requirements of their initial Convolutional Neural Network (CNN) translation models. The CEO personally racked early machines.
The Iceland Supercomputer Scale-Up: As training sets expanded, DeepL deployed a 5.1-Petaflop supercomputer housed at the Verne Global data centre in Keflavik, Iceland. The backend storage handling this cluster was built to sustain the rapid parallel read workloads required for deep learning training, utilizing a highly tailored local file system setup optimized for high IOPS.
Expansion to Sweden (Falun): As the company scaled up its Large Language Models (LLMs), it shifted a significant portion of its heavy high-performance computing (HPC) data footprint to EcoDataCenter's Main Site in Falun, Sweden. For model training, this infrastructure utilizes liquid-cooled NVIDIA DGX SuperPOD configurations, which natively lean on ultra-fast NVMe-over-Fabrics storage reference architectures to feed data to their Blackwell-generation systems.
The Shift to a Hybrid Infrastructure: While model training remains on proprietary hardware under a strict Zero-Trust framework, DeepL updated its core infrastructure to incorporate Amazon Web Services (AWS) as a global cloud sub-processor. This means their overall backend storage is currently a hybrid setup—blending local, high-performance NVMe physical storage inside their European data centres with scalable cloud storage layers provided by AWS to support instant global API distribution.
DeepL is now revealing some details of its backend storage; opening its storage kimono, so to speak, and deploying VAST to provide a higher-performance, more resilient data environment while making more efficient use of the NVMe infrastructure it already operates.
Based on current deployment plans, DeepL expects to achieve roughly three times the storage capacity from its existing hardware, giving research teams greater room to scale model development without a corresponding increase in physical infrastructure. Which is rather amazing, and implies that VAST’s SW will adopt the existing SSDs and make better use of them through data reduction, fewer copies, and possibly other techniques.
Stefan Andresen, DeepL’s Chief Scientist, said: "DeepL’s researchers are constantly pushing Language AI into more demanding areas, and the infrastructure underneath them has to keep pace. VAST gives us the performance and reliability we need while allowing us to make significantly more efficient use of the infrastructure we already operate.
"Just as importantly, it gives us a path to make data more easily accessible across our own data centres and, over time, cloud environments, reducing the manual data movement that can get in the way of research. That becomes particularly valuable as we continue investing in data-intensive areas such as DeepL Voice."
DeepL will use VAST’s SyncEngine technology to migrate data from its existing environment onto the VAST platform, helping streamline the transition. The AI OS will be used for both file and object storage.
]]>There are underlying questions here, such as: what do we mean by AI agent accuracy? And: how do we test it?
We understand that Holland’s Optiver, a global automated trading firm, saw agent accuracy go from 42 percent to 74 percent on real researcher questions across its petabyte-scale data ecosystem after layering in a governed context facility. Is 74 percent accuracy good enough for production use? Most probably not.
Online collaboration platform Miro has more than 20,000 datasets spread across Snowflake, dbt, and Looker, spanning domains from finance to marketing to product analytics. When Miro connected Claude Code directly to Snowflake to enable natural language querying for business users, the early results were promising for simple questions but note for more complicated ones.
A question like “give me the active users” would return something plausible but wrong. Ronald Angel, Lead Product Manager at Miro, and his team measured this precisely. Using an evaluation bank of 900 questions, testing text-to-SQL accuracy, created manually by data experts across business domains, they benchmarked accuracy at less than 40 percent. A sample question was: “What percentage of our premium accounts have two-factor authentication enabled?”
Accuracy was measured end-to-end: whether the agent produced the correct answer (right tables, joins, filters, and results) for each question.
The benchmark ran regularly and was updated with model or pipeline changes.
When Claude Code was connected directly to Snowflake (with ~20,000+ datasets and no strong context layer), accuracy was under 40 percent . Wrong-table routing occurred in more than 30 percent of cases.
Ronald Angel, Product Manager, Data Platform, Miro, said: “Starting with Snowflake metadata alone, our analytics agent answered about half of our benchmark questions correctly. After layering in DataHub Cloud as our context platform, including data product documentation, cross-source context and business meaning derived from our query history, we nearly doubled accuracy from close to 50 percent to around 90 percent .”
Rather than continuing to connect Claude Code directly to Snowflake with no context layer, Miro placed DataHub between the orchestrator and the SQL engine, using the DataHub MCP server to expose data entity descriptions, tags, glossary terms, column schemas, lineage, and search to the agent.
Miro used Claude Code to orchestrate question decomposition and DataHub MCP to surface governed metadata, with Snowflake Cortex executing the final SQL.
It tracked accuracy as they added DataHub capabilities as a context/routing layer (via DataHub MCP exposing descriptions, tags, glossary terms, lineage, schemas, etc.) in stages:
Reviewing the benchmark questions still failing at the 85 percent level, Miro’s team found that many failures traced not to routing logic but to sparse model documentation or deprecated columns the agent could not distinguish from live ones.
The fix was to reduce the number, of data assets used by the agent and clarify definitions. Specifically, Miro curated only the data products critical to its AI use cases, documenting them with business definitions, dimensions, and measures in dbt, loaded into DataHub via MCP. The remaining 20,000+ assets were intentionally left out of scope.
It said that, when the right asset is well-documented and exposed as the only semantic view the agent can reach, routing errors disappear. The agent does not need to choose between 50 tables. It sees one. This is called a deterministic mode and has >90 percent accuracy.
Read the DataHub Miro case study here.
Before you can assess AI agent accuracy, you have to define it. That’s a simplistic statement but the actuality is complex. Look at Miro with 900 text-to-SQL test questions, and a 4-stage project progressively improving accuracy, with the final stage dispensing with >20,000 non-essential data sets and setting data definitions, etc. in stone as it were. Many team members were involved in this over time. It was a far from a simple task and only applies to one scenario; text-to-SQL generation. And it still didn’t get to 100 percent.
Consider the idea of an AI agent acting on your behalf and booking airline tickets. Would a 90 percent accuracy rate be acceptable? It books a passenger identity, an airline, departure airport, a destination, a departure date, a departure time, a flight class, a seat, identifies a travel rewards program, and pays a price. Which of these ten things could it get wrong and still be acceptable?
None of them, of course. These are all mission-critical, so to speak. We need 100 percent accuracy.
Verifying AI agent accuracy is going to become profoundly important.
]]>Meta’s Prometheus is a gigawatt-scale GPU supercluster located in New Albany, Ohio, used to build Meta’s Muse family of foundational AI models. It is being scaled up from its current hundreds of thousands of GPUs to to around 1 million GPUs functioning in lockstep to train Meta Supercluster Labs’ next frontier models.
Currently there is an array of hundreds of thousands of GPUs across multiple interconnected buildings. There are two Nvidia GPU platforms within Prometheus:
Catalina, which uses the Ariel variant of Nvidia’s GB200 NVL72 (Blackwell); a two-rack configuration in which 36 GPUs per rack are interconnected via NVLink to form a unified 72-GPU domain.
Clemente, which uses Nvidia’s GB300 NVL72 (Blackwell Ultra) and improves on Catalina at both the GPU and the platform level. GB300 raises HBM capacity per GPU by 1.5x, from 192 GB to 288 GB, and dense FP4 throughput by a similar factor, while roughly doubling attention (softmax) acceleration. A full 72-GPU domain is deployed in a single rack.
A third Nvidia GPU platform is due, San Miguel, based on Vera Rubin NVL72 GPUs. This will be will be a generational leap in per-GPU performance and will integrate directly into Prometheus’s existing network fabric. As the cluster matures, it will deliver even more compute per watt by augmenting its capacity (primarily interconnected GB300s) with Vera Rubin.
Meta is building a second supercluster in Richmond Parish, Louisiana, called Hyperion.
There are several Prometheus datacenter buildings at New Albany. Inside a building an Ethernet-based fabric is used. Prometheus cluster buildings are interconnected by a multi-layer hub-and-spoke Backend Aggregation (BAG) network that spans from 2 km to over 100 km. This uses Arista 7800-series chassis running deep-buffered Broadcom Jericho3 silicon, delivering multiple Pbps of bandwidth per region pair.
This is a custom, multi-tier storage stack built on Meta’s Tectonic foundational layer, with BLOB-storage (Binary Large OBject) interfaces and specialized hardware platforms for bulk and high-performance needs.
The storage subsystem features a 4-layer hardware stack:
L1/L2: Memory and local flash on GPU hosts.
L3: Regional flash-backed BLOB storage.
L4: Source of truth: durable HDD-backed BLOB/Tectonic storage (global or regional).
There is distributed caching (leveraging spare GPU-host memory, reusing components from Meta’s Owl system), metadata caching, prefetching in the data loader, hedged reads, and dynamic concurrency control. The design prioritizes predictable low tail latencies so training steps are not stalled by storage.
Owl is Meta’s production system for moving massive volumes of hot internal content efficiently by combining peer-to-peer data transfer with intelligent centralized control. It is a high-fanout (from a few clients to millions of processes) content distribution system for large, frequently accessed (“hot”) objects inside Meta’s private cloud, delivering executables/binaries, code artifacts, AI models, and search indexes to millions of hosts.
The delivered content can range from a few GMB to several TB in size. There are separate data and control planes:
Decentralized data plane: Ephemeral peer-to-peer distribution trees. Data streams from sources through peers as soon as chunks arrive (avoiding full store-and-forward delays). Trees are short-lived and per-chunk.
Centralized control plane: Tracker services maintain detailed, up-to-date metadata on peers, their cache contents, locations, and ongoing downloads. Trackers decide sources, retry behavior, caching/eviction, and tree construction.
The Prometheus storage stack has migrated away from its Llama-training design of having an exposing an NFS-like FileSystem interface on top of Tectonic (see bootnote). This architecture continues to be used widely within Meta, but Prometheus is moving to BLOB-based storage.
This “is motivated by the need for unified storage access to massive data lakes in the BLOB-storage layer as well as the need for high performance” to maximize GPU utilization.
The storage software stack has two main layers:
Tectonic - the foundational, regional, multi-tenant block storage fabric. It provides high durability/availability via erasure coding, media tiering (HDD + flash), and smart placement of hot/warm/cold data.
BLOB-storage layers on top of Tectonic: These expose object-storage (and historically file-system) APIs.
There is a unified flat metadata schema backed by ZippyDB (for fast O(1) path-to-address lookups), elimination of a dataplane proxy in favor of a “fat client” SDK that streams data directly from storage servers, and regional deployments colocated with GPUs.
Meta now deploys a regional BLOB-storage stack colocated with GPUs in every AI region.
When the SDK on the client receives this API call, it now issues a getReadPlan(“/bucket/path”) request to the API server. The API server does O(1) lookup per chunk to the new metadata store to map the path to (blockId, offset, size) tuples. It then returns the ReadPlanResult to the SDK. The SDK has Tectonic BlockClient embedded within it, and so is now able to stream data from these blocks directly from Tectonic.
Two caches are used:
Distributed data cache: Prometheus leverages the spare memory on the GPU hosts as a distributed data cache for frequently and concurrently accessed data. To achieve this it reused components from Meta’s Owl subsystem: integrating the peers in the Owl subsystem directly into the BLOB-storage client SDK so that all data access goes through this data cache.
Readplan metadata cache: Readplan refers to the mapping from path to storage address. It caches the read-plan for frequently accessed BLOBs in a distributed-memory store similar to memcache.
We can now flesh out the storage scheme outline. Meta says the core idea is to have a dataloading scheme leveraging the memory and flash on the GPU host as L1 and L2 caches, and regional BLOB-storage fabric backed by flash as the L3 cache. The global BLOB-storage HDD-based data lake is the backing store, what Meta calls “the ultimate source of truth." A dataloader accesses the caches and this cache-backing store system through the BLOB-storage SDK.
Note that:
Dataloader prefetch: Dataloaders prefetch the next batch of datasets into memory while processing the current batch. This prefetch will surface as a read operation at the BLOB-storage SDK level.
Deep prefetch: We expose an explicit prefetch() API as part of the BLOB-storage SDK. The dataloader will trigger explicit prefetch of the data needed during the next few minutes by invoking the prefetch() API in the background. This API triggers hydration of data from remote storage onto the local region L3 cache and also prewarms the metadata cache.
Automatic data life cycle: Data in the L3 regional disaggregated flash tier is typically held for a configured period of time to allow reuse across epochs in a training cycle. We support custom eviction policies, including TTL and LRU policies. The eviction policies are also capacity/quota aware.
Once Meta stared using this dataloading scheme, average job ingestion time dropped from 150 minutes to 10 minutes. The maximum very large job ingestion time of 89 hours fell to just over 3 hours; 182 minutes.
Storage hardware
The storage hardware includes:
Olympic and Yosemite V3.5 Crater Lake: Flash/SSD-based platforms optimized for lower-latency, higher-throughput access.
Grand Canyon: HDD-based system for bulk training-data storage.
Support for general-purpose compute platforms (e.g., Yosemite variants) for cluster management.
These allow independent optimization of capacity, latency, and cost across the data pipeline
Yosemite v3.5 has a blade/sled chassis design with up to a dozen 10 RU servers in a 40 RU rack. It supports carious compute, memory/CXL, PICe, accelerator and SSD expansions.
Yosemite V3.5 uses OCP-aligned datacenter NVMe SSDs (E1.S for capacity/performance storage and M.2 for boot), not commodity consumer drives. They are purpose-built or qualified datacenter NVMe devices.
Crater Lake is a server designed for the Yosemite v3.5 architecture, supporting 4th Gen Intel Xeon (Sapphire Rapids / Eagle Stream) processors. It has a a single-socket 1OU (1 Open Unit) design and is intended as a base compute node. The Open Unit is specific to the Open Compute Project’s Open Rack Standard, being 48 mm (about 1.89 inches or 1⅞ inches) high, not the traditional EIA 1U = 1.75-inch rack unit.
Grand Canyon is a 4OU-tall, single-drawer dense chassis that holds up to 72 × 3.5-inch HDDs. There are three build options:
Dual storage server — Two compute nodes + 72 HDDs (all drives powered for higher performance/access speed). Fully standalone.
Single storage server — One compute node acting as a head node, which can connect to additional JBOD chassis.
Pure JBOD — No compute; used as a disk shelf. In head-node + JBOD setups this can scale to 216 drives (or more in multi-chassis configurations).
It’s another design, like Yosemite and Crater Lake, that Meta contributed to the OCP.
Prometheus uses a distributed checkpointing architecture that asynchronously writes model state (including weights, gradients, and optimizer data) to multi-tier storage. It employs delta checkpointing and in-memory staging to minimize impact on training throughput. This fits the overall tiered model (fast local/in-memory staging → flash → durable HDD-backed storage) and supports high goodput (e.g., >85% on pre-training runs)
This specialized custom storage scheme has been developed by Meta to minimize GPU waits in very large scale training clusters. If organizations are running similar scale AI training workloads, meaning they are hyperscaler or near hyperscaler class, then they could write the software themselves or, if smaller and/or the standalone object storage trio; Cloudian, MinIO and Scality, can provide the functionality they require, use that software instead.
The OCP-class NVMe SSDs should be available from OCP suppliers and commodity disk drives are commodity disk drives.
It will be interesting to see if Meta uses its Prometheus clusters for AI inferencing and agentic AI workloads, which won’t have, we suppose, such large job ingestion times. Neither will individual inferencing jobs require huge numbers of GPUs working in lockstep.
In 2024 Meta said Hammerspace was its data orchestration software supplier, supporting 49,152 Nvidia H100 GPUs split into two equal clusters and used in two Llama 3 AI training systems. Meta’s Grand Teton OCP hardware chassis houses the GPUs, which rely on Meta’s Tectonic distributed, flash-optimized and exabyte scale storage system. This was accessed though a Meta-developed Linux Filesystem in Userspace (FUSE) API and used for AI model data needs and model checkpointing.
Both the Tectonic and Hammerspace-backed storage deployments use Meta’s YV3 Sierra Point server fitted with the highest-capacity E1.S format SSDs available.
A previous AI Research SuperCluster, with 16,000 Nvidia A100 GPUs, was used to build Meta’s gen 1 AI models and “continues to play an important role in the development of Llama and Llama 2, as well as advanced AI models for applications ranging from computer vision, NLP, and speech recognition, to image generation, and even coding.” That cluster used Pure Storage FlashArray and FlashBlade all-flash arrays.
The Meta BLOB-storage SDK is the client-side library that lets applications, like AI training jobs, efficiently read and manage BLOBs from Meta’s global storage fabric, with direct streaming, Owl-based distributed caching, and AI-oriented optimizations.
]]>That bridge is set to become bigger.
Huawei's focus on AI and sustainability aids the EU in lifting its AI capability vis-a-vis the US and China, and supports its net-zero carbon reduction goals, while also offering an alternative to US suppliers.
There were more than a thousand attendees, and many watched a welcome video pitch from Italian government Minister of Enterprises Adolfo Urso stressing Italy’s desire to play a leading role in the digital transition. Various enterprise-class customers also pitched their use of Huawei gear. We are looking here at a well-established Chinese supplier of integrated compute, networking, storage, and public cloud IT services with solid green credentials, meaning less electricity usage.
What struck this observer was how comfortably Huawei occupied its place in the EU’s IT supply ecosystem.
There is a growing antipathy between the EU and what it considers the boorish and ill-informed US view of its activities. Data compiled by the Pew Research Center and Eurobarometer shows 75 to 80 percent of citizens across major Western European nations express a distinct lack of confidence in the current US administration's handling of world affairs.
On the US side here is venture capitalist Marc Andreessen of a16z: “Europe seems, I just take them at their word, determined to be left behind. They seem enthusiastic about being left behind. They seem to really badly want that," He continued, "I think they would be very happy if AI didn't happen inside the EU. I think they'd be very happy if social media didn't exist. I think they'd be quite happy to shut up the internet.”
Of course, the EU can’t do those things and falls back on regulatory control attempts instead, while disliking and disagreeing with Andreessen's views. He might say emperors never like being told they have no clothes.
The EU’s ties to the USA are weakening while its links with China are growing stronger, but these bring threats in their wake.
Jamie Dimon, Chairman & CEO, JPMorganChase & Co, thinks China’s manufacturing strength could overwhelm the EU’s industrial base. At the Economic Club of Chicago, he said this month: “The Chinese are tough. They can do what they want. I don't think they have to be a natural enemy of ours. But the other thing about them is if you don't have a common trade barriers of sorts that are fair, it's going to be hard to stop them.
“And if you go to Germany today, they don't have a proper defense industrial base, but their real industrial base was extraordinary cars, machine tools, all this equipment. I'm sure if you're in business, you have a lot of them in your plants. That's being attacked today by the Chinese and their stuff is good. And they're going to lose. And you can see it right in front of you, but it's going to take a while to play out. And we have obviously bad politics around this, but that to me is the most important thing.”
Let’s focus on one part of this: the contention that China's "stuff is good" and Germany is "going to lose."
Huawei in IT is Chinese industrial prowess exemplified. Its product portfolio encompasses smart consumer devices (watches, mobile phones, tablets), telco networking, enterprise IT (compute, storage, networking, data lake software, and public cloud services) in all its complexity and comprehensiveness, digital power gear for data centers, intelligent automotive systems, and in-house semiconductor development and manufacturing.
It has developed software and GPU hardware, and networking to be able to compete with Nvidia’s ecosystem, including KV Cache equivalence and AI factory systems with rack-level GPU pods. It has storage products that have propelled it to the number 2 supplier position world-wide, even though it is being denied access to the US market.
Huawei lacks Nvidia GPUs, hard disk drives and LTO tape. It has compensatory alternatives for the GPUs and tape but not disk drives, which means its OceanProtect backup systems are flash-based and hence expensive compared to HDD-based protection arrays. But they are faster, of course.
We believe Huawei also has a profound dislike of the American government because of the USA’s technology export restrictions, and the prolonged 3-year (2018 to 2021) house arrest in Canada of its CFO, Meng Wanzhou, the daughter of its founder Ren Zhengfei. The US charges against her were eventually dropped.
What we have here in Huawei is an extremely well-motivated and world-class Chinese IT champion with enormous expertise and energy working to build its base in Europe. The lack of disk drives, Nvidia GPUs and software is not going to stop it. Dell, HPE and the array of mainly US-based networking and storage suppliers face a determined and expert competitor. Huawei has a good bridgehead in the EU that is set to grow larger at these US suppliers' expense.
]]>In other words, they need more of their unstructured data made AI-ready.
If your product pitch is making data ready for AI then you have to find customers whose data is not ready for AI and show them that making it ready is worth while; that there is a discernible AI project ROI. Finding customers can be made easier by getting the tech media to publish stories about survey reports showing x percent of potential customers don’t have AI-ready data. So here goes…
The survey shows two-thirds of respondents (68 percent) describe their organization as mature or maturing on AI, defined as setting standards, automating business processes, agentic workflows and applying governance models. Yet only 7 percent have classified and cleaned the majority of their file data. As well, 57 percent say they have half or less of their data ready and 13 percent cannot say how much is ready at all.
Krishna Subramanian, COO and cofounder of Komprise, said: “Our survey showed that although most enterprises are maturing concerning their adoption of AI, only 21 percent-26 percent of enterprises have quantified ROI on the critical areas of operations and revenue growth. Enterprise IT and AI leaders recognize that data intelligence is lacking but the pathway to ensuring high data quality is not straightforward. It begins with visibility and analytics across storage silos, unstructured data classification and industry-specific metadata extraction.”
Komprise says it unlocks unstructured data for AI while optimizing data placement to cut storage costs. The top 5 takeaways from its survey are:
AI returns map to data readiness.Low AI ROI correlates with low unstructured data readiness. To generate measurable and meaningful ROI on AI, improving data quality must be job number-one among IT, data and AI leaders.
Data readiness is an ongoing tactic Data readiness does not resolve as organizations grow adoption of AI across departments and functions. By investing in AI data management, enterprises can get ahead. Top pain points are metadata enrichment/extraction and clearing up redundant, obsolete and trivial (ROT) data to discover the right data for each project. Sophisticated metadata management builds trust and lowers costs
Metadata analysis and extraction across the complete data lifecycle creates a foundation of trust, where AI users and engineers can easily find the most relevant, specific data for a project, without running up AI costs and risks.
Lakehouses are behind on unstructured data. A scant 4 percent of enterprises have developed schema for most file data. Until unstructured data carries the metadata that lakehouses require, these platforms deliver a single source of truth for structured data only, eroding their usefulness for AI.
Automation is still immature for unstructured data preparation Filtering data for security risk is moving forward but less than one third filter for quality and 31 percent are not preprocessing unstructured data.
The survey noted customer service is the highest area for quantified ROI at 38 percent. Quantified ROI for employee productivity and decision speed clocked 28 percent each, followed by revenue growth at 26 percent and operational savings at 21 percent. There are no details of what quantified ROI actually means for the surveyed organizations, just that there has been quantifiable ROI.
Many storage and data protection suppliers have issued AI data readiness reports. For example:
Seagate — 2026 Data Infrastructure Readiness Report (survey of 2,700+ enterprise tech decision-makers). Key findings: 99 percent expect AI to increase storage needs, but only 38 percent feel fully prepared; data quality/readiness and storage infrastructure ranked as top AI deployment challenges.
NetApp — Multiple reports, including the AI Space Race Report, Data Complexity Report (2024), and an IDC-sponsored Enterprise AI Maturity Study (2025). These examine data readiness, infrastructure optimization for AI, and gaps between perceived and actual preparedness (e.g., 84 percent still said storage was not fully optimized for AI in one study).
Hitachi Vantara — State of Data Infrastructure Global Reports (2024 and 2025 editions, surveying 1,200+ IT leaders). These focus on data infrastructure maturity, data quality as a driver of AI success, governance/security gaps, and the link between infrastructure readiness and AI ROI.
Cohesity has a Gaia solution brief discussing AI data readiness
Commvault — State of Data Readiness reports (e.g., for ANZ and Asia). These cover data management complexity, AI adoption risks, cyber resilience, and readiness challenges in regulated environments.
Iron Mountain — Reports such as The Information Supercycle: Aligning Data, Infrastructure and Capital for AI Scale and earlier AI-readiness research with FT Longitude. Findings highlight that data concerns delay or derail AI projects for a large majority of organizations, with emphasis on unstructured/legacy data and governance.
Veeam — Surveys examining the gap between confidence in data management/protection and proven AI readiness, recovery capability, and governance.
Dell Technologies — Studies such as the Modern Enterprise Readiness Study that address data readiness challenges for AI/GenAI alongside broader enterprise readiness.
Read the 13-page Komprise survey report here to find its take on the topic.
Maturing is defined as "beginning to establish standards and use AI more strategically, automating key business processes and applying governance models across the organization."
Mature is defined in the survey as: "deployed across multiple interconnected departments with agentic workflows and a unified governance and security strategy in place."
]]>Israeli startup Model9, co-founded in 2016 by CEO Gil Peleg, an ex-IBM SW engineer, and two others, replaced complex, slow and costly mainframe tape backup storage with Virtual Tape lIbrary (VTL) software sending data to public cloud object storage and S3-supporting on-premises stores. It was bought by BMC in 2023. Peleg and the team joined BMC, but he and ex-Model9 CRO Dan Shprung left in September 2025 to startup Geniez AI and develop software to connect AI Large Language Models and agents with mainframes and extract real-time data for AI use. Just over a year later it’s been sold to Rocket Software, in what looks like rocket-level progress.
Milan Shetti, president and CEO of Rocket Software, said: “This is about continuing to transform what is possible on the mainframe. By bringing Geniez AI’s specialized expertise into Rocket, we are accelerating our vision to help enterprises securely apply agentic AI to their systems with speed, confidence, and control.” Shetti is an ex-SVP And GM for HPE’s storage and big data business unit, joining Rocket in 2021 as a resident for its z mainframe unit.
Rocket Software is an IBM mainframe and i/Power-focussed IT modernization supplier and is owned by Bain Capital. It has, it says, a strong and long-standing partnership with IBM. Geniez says its AI Framework augments mainframe teams with expert AI agents operating directly on the mainframe in every domain, monitoring live systems, diagnosing incidents, remediating security vulnerabilities and executing agentic workflows using natural language. It requires no prerequisites and installs in hours.
When started up, Geniez AI said it was opening up mainframe’s data storage for real-time use by non-IBM LLMs and agents. At that time Peleg told B&F: “We’re going to bring GenAI to the mainframe just like in Model9 we brought object storage to the mainframe.”
Peleg, now SVP, SoftwareEngineering at Rocket, said: ”Geniez AI was founded on the belief that GenAI-based business applications can fundamentally change how enterprises understand and operate their mainframe environments. Joining Rocket gives us the scale, customer reach, and deep mainframe expertise to bring that vision to more organizations and help shape the next era of intelligent, mission-critical operations.”
Rocket gains Geniez AI’s technology and six employees, and they will help accelerate Rocket’s AI initiatives, particularly Rocket EVA, the company’s agentic AI platform for core systems, giving it new capabilities. Rocket Software and the Geniez AI team are working, Rocket says, to shift traditional, expertise-intensive mainframe management to intelligent, AI-assisted operations that make core systems easier to understand, manage, and modernize.
In other words, using natural language to manage the mainframe instead of complicated command line interfaces. Rocket wants to help mainframe customers apply AI agents that can reveal system insights, support action, and automate mission-critical workflows, with appropriate governance and control of course. Geniez AI gives it expanded agentic AI capabilities, particularly we suspect in the data sources access area.
The acquisition cost was not revealed. Geniez raised $6 million in a 2025 seed round, and Israeli outlet Calacalist estimates the deal value could be up to $100 million. That seems high to us.
Will Gil Peleg do it a third time; bring new technology to mainframe users? For him, thus far, the mainframe is the gift that has not stopped giving.
A Geniez AI blog says: "Core system data that typically resides on mainframes is different. It is authoritative, highly structured, and often spans years of operational history. This data captures real customer behavior, financial movement, entitlements, and transactions. These signals matter most for prediction and decision-making."
And: "Historically, core system data has been consumed in batches hours or even days after events occurred. For traditional reporting, this was acceptable. For modern AI, it is not. … Without real-time access, GenAI models operate on a lagging view of reality, forcing teams to compensate with heuristics, thresholds, or conservative assumptions."
Geniez AI provides real-time access to that data.
]]>…
BioMemory says the world's first DNA data storage service is live as it has launched DNA Shield. “For the first time, organizations can store their most critical data in DNA, as a service, and applications are open now.Every organization holds significant data it cannot afford to lose: the keys that unlock everything else, the formulas that took years to perfect, the records it must stand behind for decades. Most of it sits on connected systems within reach of an attack, on media that wears out. DNA Shield writes that data into DNA and keeps it completely offline, where no network can reach it. It is written once and built to last for generations without migration.” Find out more here.
…
BitTitan announced updates to MigrationWiz that deliver 3 to 5 times greater document migration throughput, shortening timelines for large, complex migrations across Microsoft 365 and Google Workspace. The updates improve scheduling, indexing and parallel processing across supported OneDrive, Google Drive, Google Shared Drive, Microsoft Teams and SharePoint Online document workloads. They also help reduce delays, failed items and the manual troubleshooting required to keep projects moving.
MigrationWiz helps organizations move files, folders, permissions, and metadata while maintaining file hierarchy and access requirements throughout the transition, whether a project involves gigabytes of user data or terabytes of content across a complex environment. Find out more here.
…
Cerabyte has preserved a digitally restored copy of Fritz Lang’s Metropolis on its ceramic on glass storage media, designed to keep the film intact for centuries. The company says that, nearly a century ago, Fritz Lang’s Metropolis imagined a future shaped by robots, synthetic identities and mass manipulation influencing generations of science fiction, from Star Wars and Blade Runner to The Matrix. As the film approaches its centenary, Cerabyte is using it to pose a question: How will the enormous volumes of digital information being created today remain accessible, and unaltered for the next hundred years? The answer? Cerabyte technology, hopefully.
At the OCP Innovation Village, Cerabyte will demonstrate how content stored on its ceramic media can be independently read using optical imaging, including with a smartphone.
…
Industrial software company HighByte announced a partnership with Databricks to help manufacturers and industrial companies tap into their operational data for analytics, machine learning, and agentic use cases. HighByte supports the Databricks Data + AI Platform’s direct streaming interface with its native Zerobus integration, and builds upon a comprehensive set of already supported integrations with the open analytics platform.
…
Leland Jutcher has developed JuiceMount and tells us: “Every small post house I know already owns a NAS, and still pays per seat for a cloud mount to edit off it. I wanted my own storage to show up in Finder like a local drive, so I built JuiceMount. It is an open-source alternative to the cloud mounts you cover, like LucidLink.
* Mounts NAS storage as a Finder volume at /Volumes/JuiceMount.
* Streams blocks through an SSD cache, with offline pins for pinned directories.
* Free and open source (Apache-2.0), $0 per seat, built on JuiceFS.
It is beta, macOS 14+ with Apple Silicon tested and TrueNAS SCALE as the tested server path.”
Get the repo and install notes here.
…
Micron plans to invest up to $50 million over five years to support the US Department of Energy’s Genesis Mission Graduate Fellowship Pilot Program and related national laboratory research. The fellowship creates four-year PhD pathways that pair super (artifical) intelligence research with a scientific discipline, plus hands-on research at a national lab and in industry. It’s also the first initiative tied to Micron Research Labs, the company’s planned Boise-based research hub.
…
Peer Software, which ships PeerGFS real-time file replication and synchronisation software, says Danish engineering company Liftra is a customer and has eliminated file replication delays, enabling engineers across worldwide operations to access the latest CAD files faster, collaborate more effectively and keep complex projects moving. Liftra designs and manufactures specialist lifting and transport solutions for the global wind energy industry. With engineering teams working across Europe, the Americas and Asia-Pacific, the company needed a more efficient way to manage and synchronise large volumes of engineering and project data between sites.
…
Rubrik expanded Project Hourglass, its partner alliance for securing agentic AI development, adding AHEAD, Trace3 and WWT and giving participating partners access to Rubrik Code Guardian in private preview. Code Guardian uses Anthropic’s Claude Mythos 5 to red-team an air-gapped copy of customer code, identify validated multi-step attack paths, and prioritize the risks that are actually exploitable, without touching a live repository or production environment.
…
Sandisk announced a partnership with Meta to be the exclusive provider of next-generation flash storage solutions for Meta VR Glasses when they launch in Spring 2027. Designed in collaboration with Meta, the new offerings will be part of the official ‘Made for Meta’ ecosystem. These include Sandisk microSD Express memory cards and USB-C flash drives.
…
Silk, which supplies speedy public cloud block storage, has closed a $45 million growth capital facility led by Avenue Capital Group with participation from Sequoia Capital, S Capital VC, Vintage Investment Partners, Ibex Investors, West Coast Equity Partners and others. Silk will use the proceeds to expand engineering investment in its data platform, scale go-to-market coverage of neocloud, sovereign and inference cloud providers, and deepen its enterprise footprint across Amazon Web Services, Microsoft Azure, and Google Cloud.
Silk founder and CEO Dani Golan said: “The shift from training-centric infrastructure toward autonomous agent-based production is happening now. Organizations are spending enormous sums on compute only to wait on the data layer beneath it. This capital lets us build ahead of agentic demand and bring AI to the data.”
The financing follows a period of sustained commercial expansion for Silk, including validated support for the latest AI-optimized cloud instance families and the appointment of Aaron Shoffa as Chief Business Development Officer to lead the company’s alliances with neocloud, sovereign and inference cloud providers. It last raised $30 million in 20-25 and total funding is now, we reckon, $388 million.
…
Reuters reports SK hynix wants approval from Washington to export US chipmaking equipment to the Solidigm Dalian NAND fab in China to increase production there as a condition for building a factory in the USA. SK Hynix subsidiary Solidigm wants to install new equipment to produce NAND chips in unused space at SK Hynix's Dalian, China, site. In September, we noted Solidigm currently manufactures NAND chips at an old Intel plant in Dalian, China, now run by SK Hynix, and uses different technology from SK Hynix; floating gate instead of charge trap. It also uses fewer 3D NAND layers - 192, than SK Hynix which is at the 321 and 375 levels, and other manufacturers. Any upgrade to the latest NAND technology is obstructed by US tech export to China restrictions.
SK Hynix told Reuters "The installation of equipment for Mod B is being carried out in close consultation with and under authorization from the US government. No capacity expansion is being sought for the Dalian Mod B, which was completed in 2023." Solidigm did not comment.
…
WEKA CTO Shimon Ben-David has left the company. He became its CTO in 2020. A LinkedIn post revealed the news and hints he is betting involved in some kind of new venture.
]]>According to Korea’s The Elec media outlet, Korea’s second memory and NAND supplier is looking at future PCIe speed increases and recognising that a single SSD controller CPU won’t keep up as capacity and bus speed increases mean more read and write IO performance is needed. So it is developing small specialised processing cores; H-TPUs or Hynix-Tensor Processing Units, to look after low-level and repetitive read and write command processing, metadata management, and buffer control.This offloads the main CPU which still looks after higher stack work such as storage-policy management, error recovery and exception handling. These H-TPUs are different from Google’s AI-focussed TPUs.
Koo In-je, SK hynix’ VP of system-on-chip (SoC) design, presented at the International SoC Design Conference (ISOCC) 2026 in Incheon, Korea, and said: “SK Hynix decomposed the read and write processes of traditional SSD controllers into smaller tasks and structured the system so that numerous small H-TPUs process them in parallel.”
Were the main controller SSD to look after all the processing needs for future high-capacity and high bus speed - think PCIe 6, 7 and 8 each doubling the speed of their precursor - then it would need to increase its physical size and power draw beyond practical limits.
The H-TPUs are physically small cores, taking up 4 percent of a standard general-purpose CPU core’s area (SK hynix uses ARM Cortex-R8 CPUs in its latest SSDs.) They also need less power; about 2.5 percent of the standard core. Also, when not active, they are set into an idle state to save power. We’re told they have a 4-stage pipeline, a sequential execution design, and will be given specialised functions.
We understand SK hynix will deploy H-TPUs in its next-gen enterprise SSDs and Storage Next SSD controllers, including the AIN P high-performance, AIN B high-bandwidth and AIN D high-capacity drives. It will likely also feature in the SK hynix-Sandisk High-Bandwidth Flash (HBF) Technology.
]]>IBM Storage Ceph is IBM’s enterprise-supported distribution of the open-source Ceph software-defined storage platform for file, block and object storage. It’s based on the open-source Ceph project (RADOS + RBD + CephFS + RGW/S3). IBM took over the enterprise Ceph offering from Red Hat in early 2023, and it It runs on standard x86 server hardware (with optional IBM Storage Ready Nodes that bundle validated hardware + software) and is designed for hybrid-cloud, AI/ML, data lakehouse, and large-scale enterprise workloads.
We note that IBM Storage Ceph has a BlueFS storage feature with up to three device levels:
WAL – fastest (write-ahead log / journal),
DB – intermediate (metadata),
SLOW – the main/slow data device; the large block device that stores the actual object data, usually HDDs, and a spillover target.
DB is an optional dedicated, faster device (typically SSD or NVMe) that stores BlueStore’s internal metadata. Its BlueStore uses an embedded RocksDB database for metadata (object-to-block mappings, etc.). The DB device holds as much of this RocksDB data as possible. If the DB device fills up, metadata automatically spills over onto the primary data device.
This latest Ceph release features:
BlueFS Spillover Cleaner - An intelligent BlueFS Tiering Manager has been implemented to optimize file placement across DB and SLOW devices,
Enhanced NVMe/TCP Gateway Lifecycle Management - enhancements to the NVMe-over-Fabrics (NVMe-oF) Dashboard, focused on improving usability, operational efficiency, and admin control,
RGW S3 Object Browser enhancements - advanced object discovery through metadata and tag-based
search, streamlined bulk management capabilities for large data sets, simplified data management, and improved security governance,
Virtual-Host S3 domains configuration from within the Ceph dashboard is a simplified and centralized way to configure virtual-hosted-style S3 access for Ceph Object Gateway environments through the RGW service
configuration interface,
CephFS Subvolume Quarantine capability - implemented for CephFS volumes and subvolumes, to protect
environments following a ransomware attack or other malicious file system tampering.
The BlueFS Spillover Cleaner ensures high-value, frequently accessed metadata remains on faster DB media while less active data is placed on lower-cost storage. This enhances app performance by automatically promoting hot RocksDB files to the DB device, so improving metadata access efficiency, contributing to more consistent Ceph cluster performance and lower latency for business-critical workloads.
Find out more in blog post here.
]]>VergeIO has had fresh but unrevealed funding this year and ex-VMware execs have joined its board; former VMware CTO Kit Colbert and former VMware ecosystem leader Zia Yusuf. It also recruited a new CEO, Parag Patel, the former VP for World-Wide Sales, Hyperconverged Software, which is what VergeOS is; very hyperconverged. We understand VergeIO is making a strengthened push to be a VMware alternative for the several hundred thousand VMware customers facing Broadcom-owned VMware price rises. It aims to replace VMware by having a better and more cost-effective hyperconverged virtual machine product and by providing stronger AI support than VMware.
CEO Parag Patel said: “VergeIO is focused on simplifying infrastructure design and operations by helping customers exit legacy stacks, reuse hardware to combat data center inflation, manage any number of distributed data centers, and prepare for on-premises AI workloads. Today's release lets IT execute a faster, less disruptive, and more complete VMware exit.”
A quicker migration from VMware comes from v26.2, with its changed block tracking technology, adding asynchronous I/O to network block device (NBD) transfers, making both the initial full backup and the final transfer 120 percent faster and shortening cutover. That’s a 2.2x quicker.
V26.2 supports Microsoft Failover Clustering within VergeFS, its internal storage capability, which aggregates commodity drives inside servers into a highly scalable, high-performance vSAN. Windows workloads like Microsoft SQL Server typically rely on Microsoft Failover Clustering, which requires a shared disk and SCSI-3 Persistent Reservations. That requirement has often led customers, VergeIO says, to a dedicated storage array, stalled many VMware migrations, or left customers running two hypervisors. The Failover Clustering support fixes this issue.
We’re told distributed architectures like retail point-of-sale often run single-server data centers that replicate back to a core VergeOS implementation. VergeOS 26.2 lowers the requirements for these sites to one server with two network connections and no core network, so customers can bring those locations along in their VMware exit.
With the fresh funding and its trio of ex-VMware exec, plus EMEA expansion, it seems clear that VergeIO sees a substantially reinvigorated market in which to capture VMware refugees by offering cost savings, easier migrations and, we expect, coming AI inferencing and agent support.
VergeOS 26.2 is available now and you can get more details from a data sheet.
]]>This Gartner report defines “defines the privacy user experience (UX) as the components of an organization’s privacy program that directly touch an individual. These components provide transparency and control over individuals’ personal data, enabling them to manage and exercise preferences and rights. Privacy UX provides organizations with a compliance-backed foundation for responsible data use by consolidating and synchronizing individual choices across all touchpoints, thus enabling robust consent management and efficient subject rights processing.”
Gartner says: “The privacy UX enables individuals to control how much personal data to expose, to whom and for what purpose. It encompasses all the touchpoints where individuals meet and engage with organizations.”
According to Veeam, its privacy UX capabilities offer organizations one way to understand, secure, and govern personal data across production and backup environments, so that consent, preferences, and rights hold up across the full data lifecycle.
CEO Anand Eswaran said: “You can't protect what you don't understand, and you can't promise privacy you can't prove. Privacy is a commitment that has to survive across every copy of their data, in production and in backup. Veeam is the first to unify data security, governance, privacy, compliance and resilience together, and that's why we lead this market worldwide."
Veeam gets into this MQ through its 2025 acquisition of Securiti, so gaining the ability to scan a production data set, such as a customer's entire AWS data estate, and map it using graph technology. It gained data item relationship intelligence and mapping, with users, essentially, being trackable data items with properties and relationships. Veeam now competes with whole new class of other suppliers in this privacy UX area and a look at the MQ diagram shows who they are:
There are 11 vendors with 5 leaders, Ketch being ahead of the others. Veeam is unique amongst the 11 in combining privacy controls across production and backup data.
It says that, through its the DataAI Command Graph, it applies data security posture management (DSPM) and classification across both live and backup data. This closes a long-standing blind spot, because personal data persists in copies long after it's been governed in production, and it supports prioritized recovery, so organizations can restore the most sensitive data first during an incident.
You can download a copy of Gartner’s Privacy UX MQ from Veeam’s website.
]]>Uvi says every AI agent an enterprise deploys starts work knowing nothing about the company it works for. It has to reconstruct that context from scratch, and the company pays a frontier model every time it does. It has knowledge graph technology to give Claude, Gemini, GPT and every other AI tool a company uses the missing context behind its decisions, conversations and code, creating a context reasoning layer the company owns once and can make available to every agent it runs.
CEO and co-founder Gilad Salinger said: "Enterprise context is a problem that gets bigger as AI adoption grows. Early use cases exposed accuracy gaps. Today, the conversation has shifted to cost and tokens. But as companies move into more complex agentic workloads, the real bottleneck becomes speed. And speed depends on context: understanding what the organization knows, what other agents are changing in parallel, and the history behind those decisions, then narrowing it down to exactly what each agent needs to keep moving."
Keeping a semantic layer, a knowledge graph and the reasoning layer on top relevant, permissioned and in sync with live data is the heavy lifting no engineering team wants to own. Uvi says it does it for them.
Salinger and CTO Dror Wolmer, both in Tel-Aviv, founded Uvi in 2023 to build the substrate enterprise AI agents need to act inside organizations, not just summarize documents. It was initially called Naboo. Salinger is an ex-VP technology at Nutrifresh Services, which acquired Fetchy, a company he founded, while Wolmer is an ex-security researcher and CTO of Cognigo, which was acquired by NetApp.
Naboo said it was building a “Reasoning Layer” (also called an enterprise context or semantic layer) for AI agents so they can act inside organizations rather than just summarize documents. Enterprise AI agents often fail or become expensive because organizational knowledge is scattered across systems (code, tickets, PRs, Slack, docs, monitoring) and the why behind decisions is rarely explicitly recorded. Naboo/Uvi addresses this by constructing a Decision Graph.
Decisions are modeled as first-class nodes (with owner, trigger, blockers, dependencies, and supporting evidence).
It uses live joins/ETL across tools (GitHub/GitLab, Jira/Linear, Slack, Confluence/Notion, Datadog/Splunk, internal services, etc.).
Agents query it via GraphQL or MCP (Model Context Protocol) and receive a structured chain of relevant decisions instead of a list of documents or similar chunks.
It runs on-prem or in the customer’s virtual private cloud (VPC), with native RBAC/permissions enforced at retrieval and zero data egress.
This is positioned as more than RAG (Retrieval-Augmented Generation) , enterprise search (e.g., Glean-style), or a standard knowledge graph—focused on intent, company-specific language/semantics, and actionable reasoning chains.
Uvi says RAG was a breakthrough for giving LLMs access to external knowledge. It works by embedding documents into vectors, then retrieving the most semantically similar chunks at query time. For a customer-support bot searching a knowledge base, this is effective.
But enterprise R&D environments aren't knowledge bases. They're living systems with code repositories, ticket trackers, pull requests, Slack conversations, monitoring dashboards, CI/CD pipelines, and documentation - all interconnected, constantly changing, and governed by strict access controls.
When an AI agent needs to help a developer with a task, it doesn't need “the 10 most similar document chunks.” It needs to understand the specific context of that task: which code is involved, who owns it, what decisions were made, what changed recently, and what the developer is allowed to see. RAG can't provide this. A Reasoning Layer built on a Decision Graph can.
Uvi runs in production at Lemonade and monday.com, both public companies, and at Melio, now part of Xero. It connects to Slack, Codebase, Coding Sessions, Meeting Summaries, Jira, Confluence, and more, with additional integrations in development. Employees see only what their existing access rules allow, and every answer traces back to its source.
The company says that, at monday.com, asking what shipped last week does not lead to a single record. A feature may exist across a ticket, a pull request and a feature flag that only make sense together. Without understanding those relationships, an AI agent can return a list of tickets and still fail to answer the question.
A Uvi blog discusses a benchmark test around questions such as: "Do we have a guide for creating a GitHub app?” It ran the same set of questions two ways: once through the context layer, once through a do-it-yourself agent over the same sources, and measured cost per correct answer, total model spend divided by the questions actually answered right, not just answered. The do-it-yourself side came out about 5x more expensive per correct answer, because it burned tokens wandering for answers a good map would have handed it directly.
It says that, built the way most enterprises are building today, an AI answer costs $0.75 to $1.41 and can exceed $2.00. Built on Uvi, the same answer costs just $0.10 to $0.16. On open ended questions where agents became stuck in search loops, Uvi was up to 46 times cheaper.
Uvi is model-agnostic, saying that AI models can change but a company’s context does not. Companies can change models and agents without forcing each new system to relearn how the business works.
The company has now raised a $7.5 million seed round led by CT Innovations with participation from Cardumen Capital, Manu VC and 91 Ventures. Uvi will use the funding to build its U.S. team, expand its go to market efforts and continue product development.
Find out more about Uvi’s technology here.
Wolmer is the author of three US patents, eg; US-12468695-B1 · Issued Nov 11, 2025: "System and method for querying knowledgebases utilizing a language model based on consolidated identity context." This uses knowledge graph technology and natural language input to traverse a graph and find a context.
Several other companies are exploring the idea of an enterprise having organizational memory, context for AI agents and models, such as HYCU, Jedify, and MinIO.
]]>When a GPU’s high-bandwidth memory runs out of capacity cached key:value items used in AI token processing, tier down to the server’s x86 DRAM, which also has capacity limits. The Computer Express Link (CXL) pools external memory and makes it available across the PCIe bus to connected servers and other devices, so enlarging the DRAM KV caching pool. Fabless semiconductor supplier Marvell’s original Structera CXL controllers have upped their original CXL 2.0/PCIe gen5 game by supporting CXL v3.2 and PCIe gen6 as well as more memory channels.
Will Chu, EVP and GM of Custom Cloud Solutions at Marvell, said: “Marvell is the only provider with a complete end-to-end CXL portfolio spanning memory expansion (Structera X), near-memory acceleration (Structera A), and PFMA (Photonic Fabric Memory Appliance), pod-scale CXL memory-sharing appliance, giving customers a complete portfolio from direct-attach to disaggregated memory pooling. We are delivering industry-leading CXL controllers, co-developed with leading hyperscalers to address the most demanding AI infrastructure requirements.”
Marvell’s first generation Structera controllers were the X 2404 and 2504 and its second generation X 3406 and 3506 with this naming scheme:
First digit: CXL generation (2 = CXL 2.0; 3 = CXL 3.x / 3.2).
Second digit: DDR generation (4 = DDR4; 5 = DDR5).
Last two digits: Memory channel count (04 = 4 channels; 06 = 6 channels).
The X 3406 and 3506 provide:
Expanded memory capacity for AI with six DDR4/DDR5 memory channels, 50 percent more than the previous generation, supporting up to 9TB of DDR4 or 6TB of DDR5 memory per controller.
Accelerated memory-bound AI workloads: Up to 5x higher LZ4 compression throughput enables 2–3x more effective memory capacity, helping improve large language model serving and DLRM (Deep Learning Recommendation Model) performance.
More hosts as CXL 3.2 support enables memory sharing across up to four hosts with up to 256 GB/s of host bandwidth, combined with enterprise-grade RAS and post-quantum security.
Jim Handy, general director of Objective Analysis, and author of a CXL market report, said: “CXL is starting to gain widespread adoption in AI systems since they reap strong benefits from its very high bandwidth and support of large memory sizes. In addition, today’s memory shortage and subsequent high prices are driving IT managers to use CXL to re-deploy used DDR4 DIMMs into DDR5 systems for immediate cost savings. In the near future, these same managers will use CXL’s memory pools to reduce stranded memory and drive greater savings, and they will find controllers that offer compression, as this one does, to be even more interesting.”
Marvell said the prior-generation X 2404/2504 architecture is “already shipping to hyperscalers,” and two hyperscalers offered support for its new generation.
An Alibaba spokesperson said: “Our collaboration with Marvell brings together its leadership in CXL technology with Alibaba’s experience in large-scale AI and cloud infrastructure to develop memory architectures designed for the next generation of AI data centers.”
Meta director of Hardware Engineering Yashar Bayani said: ”Meta is pleased to work with Marvell to advance CXL solutions that address the growing memory demands of compute infrastructure. As systems scale, increasing memory capacity while improving infrastructure efficiency and cost becomes critical. CXL provides a flexible approach to expanding memory resources and optimizing infrastructure economics, and we look forward to working with the ecosystem, including Marvell to help enable the next generation of scalable infrastructure.”
Marvell Structera X 3406 and Structera X 3506 CXL memory expansion controllers are expected to sample in Q1 2027.
Marvell has a strong focus on photonics and we expect it to introduce external memory pools connected to processors across high-speed photonic links. CXL will likely adapt to photonics mainly by using optical interconnects (silicon photonics and fiber) as the physical transport layer while keeping the CXL protocol, coherency, and memory semantics intact. PCIe at gen 6 and 7 could also adopt photonics physical transport.
]]>TapeArk says this makes decades of tape-held information available for AI analysis and processing: “Your archive is a longitudinal dataset waiting for liberation.” It’s highlighting how this can help pharmaceutical suppliers, saying Almirall had 50 years of research data, 400,000 R&D documents on offline tape. Moving this to Azure made it searchable online. Canada's Ontada migrated 150 million cancer (oncology) documents into Azure.
It says tape issues are well-known.You need old tape drives to read legacy tapes and old software possibly as well. The latest LTO-10 tape format has no backward compatibility. Tapes deteriorate and need copying to fresh media; re-silvering. Migrating old tapes to newer formats is a formidable problem with large and long-lived tape archives. So, TapeArk says, do away with the whole mess and migrate to virtual tape libraries (VTLs) and drives in the public cloud.
It proposes that they audit (no cost), you ship your entire tape archive to one of its data centers where they identify the tape reels, read the contents into local cache, generate metadata and tags, and then perform a mass public cloud ingest. The old tapes are then returned or tossed away. Its Nexus product is a cloud based platform for the management of legacy tape-based data that allows you to view, analyze and restore data from your legacy tape environments without the need to maintain any physical tape storage, legacy tape hardware, software or data center space at all.
You can then reclaim the tape library data center space and use its power and cooling for something else.
Tape Ark connects a customer’s existing backup tape application to the virtual tapes. Business customers pay for the service using a SaaS model with no need for capital expenditure (CapEx) for equipment or software, or related operational expenditure (OpEx) for managing and maintaining the legacy environments. The Nexus SaaS service is sold based on the number of physical tapes you have, the volume of data they contain and the number of environments you need to be supported. There are low to no tape entry fees and Nexus is available on multi-year tailored plans.
TapeArk suggests that a year of LTO-1 tape vault storage ($0.40+/tape/month) buys abut 4 years of Azure storage ($0.00099/Gb) for an LTO-1 tape, 75 percent less expensive.
The Tape Ark services are available globally, with mass ingest facilities located in Australia (Perth), India (New Delhi), the UK (Exeter), USA (Dallas) and Canada (Toronto). Tape Ark also offers mobilization to customer sites for in-house data ingest (subject to connectivity requirements).
Learn more by browsing TapeArk’s blog webpages.
TapeArk was founded in 2017 by Australian Guy C.Holmes. This followed the 2015 acquisition of the oil and gas data management division of his Perth-based SpectrumData company by US based Katalyst Data Management. SpectrumData was an Australian specialist in data storage, management, tape transcription, scanning, migration, and recovery services, focused on the geoscience/oil-and-gas industry and government clients in the Asia-Pacific region.
Tape Ark takes a company’s tape collection and copies the data into the public cloud. Tapes are written into the cloud as virtual tapes and are paired with a virtual tape drive that also resides in the cloud. If you need to do a restore, you can use the Tape Ark access portal to load the virtual tape into the virtual tape drive and the restore can be performed without the need for you to maintain any physical tape drives, legacy software licenses or pay for any physical offsite storage of media.
The company supports the AWS and Azure clouds. It also works with Google Cloud, Wasabi, Backblaze, and others. Its Nexus SaaS platform and tape migration/virtualization services are designed to run on the customer’s preferred cloud, with explicit support for AWS and Azure environments and Virtual Tape Library (VTL) options in those platforms. TapeArk says its moved >1.3 exabytes of tape data to the public cloud so far in 380 tape-to-cloud projects.
Several other suppliers provide public cloud VTLs: AWS Tape Gateway, FalconStor, OpenVTL, and Spectra Logic with Geyser Data TAPAS.
]]>DigitalOcean is a 14 year-old Infrastructure-as-a-Service CSP offering cheap virtual machines in its cloud that’s expanded into offering AI IaaS services to its customers, ranging from GPU hardware to agent development services. VAST Data provides disaggregated all-flash storage with a substantial AI-focussed software superstructure called the AI Operating System which has been adopted by many Neoclouds. The companies say that AI agents need fast, governed access to context, but developers today assemble it from separate object stores, vector databases and data pipelines. Together, they are collapsing that stack into one integrated system for inference and agentic applications.
Renen Hallak, Founder and CEO of VAST Data, said: “Agents are only as good as the context they can reach. They require an integrated environment where compute, data and security work together as one system. VAST and DigitalOcean are building that platform for agents. By combining DigitalOcean's AI-Native Cloud with the VAST AI OS, we're creating an optimized path without forcing developers to stitch together disconnected services. Together, we're giving models and agents the context they need in real time, with the simplicity and economics enterprises need to operate AI at scale.”
DigitalOcean is a publicly-traded American cloud service provider founded in 2012 and offering IaaS (infrastructure-as-a-service) facilities. it says it focuses on making cloud hosting affordable and straightforward for developers, startups, and small-to-medium businesses (SMBs) and is positioned as a simpler alternative to larger providers like Amazon Web Services (AWS). The flagship service is Droplets; easy-to-spin-up virtual machines (VPS-style Linux servers) available in various sizes and regions, starting at low monthly prices (e.g., around $4–6/month for basic ones). DigitalOcean is headquartered in Broomfield, Colorado, US, with 15 globally distributed data centers.
It’s pivoted from a traditional developer-focused cloud (centered on simple Droplets/VPS) into an “AI-Native Cloud” aimed at inference and agentic workloads. This was exemplified with its 2023 acquisition of Paperspace for $111 million in cash to obtain GPU-powered infrastructure and AI/ML tooling, helping DigitalOcean offer easier ways for developers and SMBs to train, build, and deploy AI models.
DigitalOcean now offers high-end GPU compute: GPU Droplets and bare-metal GPU servers with Nvidia (H100, H200, B300, L40s, RTX series) and AMD (Instinct MI300X, MI325X, MI350X, etc.) hardware. Options include single-GPU and multi-GPU (including 8x) configurations, plus spot/interruptible capacity for lower costs.
Its datacenters, such as its Richmond facility are being upgraded with liquid cooling, high-speed non-blocking RDMA fabrics, and capacity for dense GPU workloads. It says owning its own silicon instead of renting capacity improves unit economics at scale.
In effect, DigitalOcean has has moved from basic cloud hosting into the Neocloud market, providing a full AI stack centered on making agent deployment and inference practical and affordable for its core customers.
With VAST Data, DigitalOcean delivers Agent Data Services (ADS), a fully managed data layer where agents and AI applications read the context they need and write back what they learn, leveraging the reliability and scale of the VAST AI OS. Every piece of a customer's data, from agent logs and traces to video, documents, and images, becomes agent-ready the moment it arrives and searchable by meaning in a single call.
ADS unifies storage, database and vector search in one system. Data is indexed and queryable as it lands, with no separate vector database or ETL pipeline to manage. VAST says everything lands in one tenant-isolated catalog, with built-in governance controlling what each agent can read and write so every agent operates only within the data it is permitted to access.
According to VAST, DigitalOcean slashed boot times from 120 seconds to 15 seconds and cut operational overhead by unifying storage and databases. It eliminated GPU idle time and reduced operational infrastructure overhead.
A Youtube video tells you more.
Paddy Srinivasan, DigitalOcean CEO, said: “The inference and agentic era needs a cloud where every layer is built to work together. Expanding our partnership with VAST brings the VAST AI OS into the DigitalOcean AI-Native Cloud, so our customers get the performance and scale of a world-class data platform with the simplicity and scale they expect from DigitalOcean. Together, we're making it easier and more cost-effective for developers and businesses to run AI in production.”
]]>The CM9-DLC products have cold-plate, direct liquid cooling support and they also support air-cooling.The drives are cooled by having a liquid-cooled metal enclosure touching the top or both sides of the device.
Maulik Sompura, Senior Staff Director, Product Management, SSD business unit at Kioxia America, said: “The rise of AI infrastructure is driving a fundamental shift in data center design toward higher power densities and advanced cooling technologies. The Kioxia CM9-DLC Series addresses these evolving requirements by combining industry-leading enterprise SSD performance with direct liquid cooling support in the E3.S form factor. This enables customers to deploy denser, more energy-efficient AI systems while providing the performance, endurance, and capacity needed for data-intensive workloads, including emerging AI architectures that rely on large-scale memory and KV cache resources.”
Kioxia gave its CM10 series of SSDs liquid-cooling support earlier this year.
Deploying liquid-cooled CM90-DLC drives will involve the system builder matching a host server’s cold-plate/manifold system, coolant loop, and undergoing qualification. Kioxia does not provide any information, such as mounting, thermal contact surface, or specific cold-plate compatibility, beyond saying the drives “support cold-plate-enabled server designs.”
It says that system builders should ensure their server chassis complies with modern standards like the SFF-TA-1008 rev. 3.0a specification, which discusses cooling-interface areas and mechanical keying. This spec includes enclosure modifications designed to establish proper flush contact between the drive housing and the external cold-plate manifold where the liquid coolant circulates.
The liquid-cooled CM9 drive performance numbers are the same as for the air-cooled CM9.
Kioxia CM9 Series drives are currently sampling to select customers and will be showcased at the 2026 OCP Global Summit, taking place October 12 – 15 in San Jose, Calif.
]]>Its backup appliances feature tiered backup storage with a deduplication free landing zone for fast ingest and restore, and globally deduplicated retention storage which can further reduce already deduplicated backups from Commvault, Veeam and other vendors. Its appliances scale out, with deduplication working across the systems, and there is an AI-powered retention time lock for ransomware recovery.
It says it has a double-digit Y/Y growth rate, well over 5,000 customers in 108 countries, and its competitive 4-quarter average win rate is more than 70 percent.
CEO Bill Andrews tells us: “ExaGrid continues to grow and has become the largest independent backup storage vendor in the world … We are getting into and winning more and more larger deals including new customer deals over $1 million. In fact, we had the most bookings from over seven figure deals in our history.”
And there’s more: “We continue to generate cash every quarter and as you know we have zero debt of any kind. We don't even have an accounts receivable line. … Our win rate remains strong. Our customer retention rate remains strong. We do over 50 percent of business now, outside of the United States. We have far more sales reps outside the US than inside the US.”
ExaGrid, with its 70 percent win rate, replaces hard disk drive primary storage behind the backup app, such as Dell PowerStore, HPE Alletra and NetApp eSeries, then Dell Data Domain (PowerProtect) and HPE StoreOnce. The most common competitive suppliers are the big 3 storage vendors: Dell, HPE, and NetApp. It periodically sees Everpure, but less than before, amd Hitachi Vantara.
Andrews says: ”We are not seeing Everpure SSD anymore as the price of SSD has gone through the moon and is now too expensive for corporate data center backup.”
“We are,” he says, “beating the cy2026 business plan for bookings, revenue, profitability, free cash flow, EBITDA, etc." All in all: “Our competitive win rate, customer retention rate, NPS score, and percentage of customers on maintenance and support remain the highest in the backup storage industry. Our growth rate is accelerating [and] We are having a very solid year and the company is pumped up.”
Decades ago NetApp tried to buy Data Domain in 2009 only to have EMC snatch it away at the last minute. Dell then bought EMC seven years later and now is the largest backup appliance vendor. With this corporate memory it has been a tad surprising that NetApp has not acquired ExaGrid. It would be a straight tuck-in acquisition with ExaGrid complementing ONTAP and StorageGrid systems and not competing with them in any way. The constantly growing ExaGrid business would provide positive cash flow and an opportunity for NetApp reps to cross-sell NetApp offerings to ExaGrid customers. It would also give NetApp a way to tap away at Dell's Data Domain base and slowly reclaim some of that lost business.
ExaGrid was founded in 2002 by CTO Dave Therrien who is on its board along with CEO Bill Andrews and representatives from VCs Investor Growth Capital, Lead Edge Capital, and Sigma Partners. It’s headquartered in Marlborough, Massachusetts. Bill Andrews has been President and CEO since 2005.
]]>Vera bills itself as a "verified intelligence platform" “that makes AI tell the truth.” The aim is to find old data, like product prices, that conflicts with newer data, and replace the old with the new so that a sales proposal prepared by a sales chatbot is truthful.
Vera is designed to act behind or as an AI chatbot and has two components: a data ingestor and a natural language-fed query answer facility.
The ingestor cleans a company’s unstructured data, including documents, PDFs, slide decks, chats and emails where most knowledge actually lives, before any AI uses it. It breaks unstructured content into individual items, called claims, recognizing the specific facts inside them, such as a price, a spec, a date, sales course, or a policy. It checks each claim against every other claim. Conflicts and outdated facts are resolved, or flagged for an expert (maven) to settle, before AI uses them. A verified claim is said to be "sealed."
The query answer part accepts input from a natural language-fed sales chatbot and returns truthful answers from the cleaned data. For example, it could figure out the answer to a question like “What’s the list price for a 6-unit Kinetic site with Halo and Sentinel Enterprise? The customer’s last quote came off the FY26 price book. Does that price still stand?”
Mark Snow, Unicorn co-founder and CTO, said: “For 30 years I was the engineer in the room when the customer asked the hard question, and for 15 years I helped teach the engineers in those rooms for Cisco’s partners. Again and again, we watched good people lose deals because they were working from bad data, or because a new feature already did what the customer needed and nobody knew it existed yet. Our team built Vera so the right answer shows up with the source attached, and so the people who know the most get to spend their time lifting up the next person.”
UnicornIQ explains in its press materials, "tech OEMs and VARs suffer from a brutal reality: a field rep has a customer meeting tomorrow morning, so they scramble through a maze of shared drives and outdated portals to duct-tape together a pitch. They end up quoting old pricing, referencing discontinued product features, or missing context on what the specific customer bought last quarter."
It says generative AI promised to fix this, but when a model doesn’t know something it fills the gap with an answer that sounds right. Retrieval-augmented generation (RAG), the standard fix, searches a company’s documents but never checks whether they agree with each other or are still true. Users are then presented back with whatever answer the AI decided was correct that day.
Specific subfeatures of Vera include:
An autonomous enablement engine stops the field from pitching blind off outdated PDFs. Vera tracks live product changes and automatically updates learning paths, coaching materials, and documentation in real time so your field is always selling the current reality. “Vera, how do I pitch this new feature to ACME Corp? Build me a compliant leave-behind leveraging current pricing and our latest company brand guidelines.”
A full channel partner program is billed as a complete operational layer built for the channel ecosystem. Vera handles partner records, contracting, deal registration and attribution, tiers, rebates, and incentives, actively orchestrating the entire partner lifecycle.
A CRM & execution Layer works alongside enterprise systems like Salesforce rather than replacing them. Vera surfaces stale deals, flags missed touch points, and runs scheduled actions like drafting replies, turning static knowledge into active revenue generation.
Pitch coaching & accuracy auditing lets reps practice pitches inside of Vera or ask Vera to sit in on live customer meetings. Vera actively evaluates them not just for delivery habits, but for product and pricing accuracy against the verified knowledge base.
Vera is available now. Access the free demo trial or contact sales/partnership teams at unicorniq.ai.
UnicornIQ does not say Vera is an AI chatbot or large language model (LLM), but departed co-founder Joe Onisick implies quite strongly that it is an LLM in its own right. We have asked the company to clarify this.
In a February 2026 YouTube video Onisick likens Vera to a data hygiene black box, speaking of how Vera looks at data sources: “I can parse them and distill the information, and then I create a vector database which is just a source of truth. It’s not really metadata about your data. It’s the facts in text with a vector index in a gravity-weighted database.” That sounds like an LLM.
The data is exposed via an API and MCP (Model Context Protocol) server. Vera is designed to fit in with whatever AI stack a customer has and “sit in the back," for example, behind a sales AP chatbot. So, Onisick says: “Just add a prompt to that existing agent and tell it to check for truth first.” Then the chatbot agent connects to Vera to use its information.
If Vera gets asked a question it can’t answer, than it says so and requests a human maven, a trusted specialist, to provide an answer in a human-in-the-loop exercise. A maven-supplied answer is then added in to Vera’s knowledge base and used from then on.
Icertis and UnicornIQ are not related. Each ship a separate product named Vera, and nothing public links the companies or the products. Icertis Vera is the contract-intelligence layer inside Icertis’s contract lifecycle management platform.
]]>It’s free, read-only, and discovers the AI agents operating across an organization's applications to find out what each one can read or write, whether it acts unattended, how many people can invoke it, is anyone listed as its owner, and and then identifies which of those AI agent-interacting applications lack independent backup. HYCU says agents increasingly run with broad permissions and write access to systems of record. A misconfigured instruction, a compromised credential, or an agent acting outside its intended scope can change or delete data across multiple applications at machine speed. Where those applications have no independent backup, that data cannot be recovered.
HYCU founder and CEO Simon Taylor said: "Every organization we talk to is adding AI agents faster than it can count them, and most of those agents hold permissions nobody regularly inspects. If you can't see it, you can't protect it. aiR Graph gives any organization that first look for free. It shows which agents can change or delete data, and which of the applications they reach have no independent copy to recover from. That second answer is where resilience starts.”
Jason Buffington, Principal Analyst, Data Protection Matters, points out: “Mapping agent permissions against backup coverage is the gap assessment that most IT and security teams didn't realize they needed and will appreciate immediately."
The aiR software was announced in May to query multiply-sourced backup data to find cyber-resilience risks no single-source risk visibility app can discover. The aiR Graph offering identifies agentic risk to applications. In detail, aiR Graph does this:
Connects read-only to Microsoft Entra ID, including Microsoft Entra Agent ID where it is enabled, and to Okta, including Okta AI Agents where the organization subscribes. Entra ID and Okta are the first supported discovery sources, and HYCU will be adding more.
The first scan maps the total applications in the organization and flags all copilots and agents in the organization.
For every agent, aiR Graph shows its OAuth scopes, the data sources it can read or write, whether it has organization-wide access, whether it can act unattended, how many users can invoke it, and whether anyone is recorded as its owner or sponsor.
Risk assessment with severity reflecting capability, not just ownership. The more an agent can reach and change, the higher it ranks: unattended access to every user's data, broad access combined with write or delete permissions, or an account still enabled after its agent identity was deleted.
It shows which applications agents reach most and which of those HYCU R-Cloud can protect, so teams can back up the systems of record most exposed to agent activity first.
Assessment Report - one click produces a point-in-time report with an executive summary, recommended triage actions, and a full inventory for management, audit, or board review.
Once aiR Graph shows that an agent can write to, for example, a Git repository, a Salesforce org, or a Confluence space, the next step is an independent, immutable backup of that application in HYCU R-Cloud, which protects more than 100 workloads. That copy is what lets an organization roll back to the moment before an agent went wrong.
HYCU Chief Product Officer Anant Chintamaneni said: "We made a deliberate choice to keep language models out of the risk assessment. Every finding in aiR Graph comes from an explicit rule applied to data read from your identity provider. Run it twice on the same tenant, and you get the same answer, and every finding traces back to the field that produced it. When a security team takes that report to an auditor or a board, it has to hold up.”
HYCU aiR Graph has extensible design for additional discovery sources, and is available to any organization, whether or not it is a HYCU customer, directly from HYCU and through its partner network. It’s is part of HYCU aiR, which lets organizations search, query, and run purpose-built agents across their backup data.
To join the aiR early adopter program, visit https://www.hycu.com/air.
]]>This AI Inference Factory brings together validated open-weight models, a disaggregated inference-serving layer, and Scality AI Data Infrastructure (ADI), which uses policy-governed autonomous operations to manage massive datasets across the AI lifecycle.It gives enterprises, government agencies and neo-cloud providers a supported alternative to cloud-based AI services, without having to assemble, integrate and maintain the entire software stack themselves. It puts ADI object storage under a vLLM/Dynamo serving layer as a shared KV cache, with S3-over-RDMA and a native KV connector.
Scality co-founder and CEO Jérôme Lecat said: “With AI moving into mission-critical production environments, organisations need greater control over where inference runs, how their models are managed and what happens to their data. For 15 years, Scality has built data infrastructure that thousands of customers around the globe rely on to operate 24/7. AI Inference Factory brings that experience to on-premises AI, giving organisations the reliability and sovereignty they need to run critical AI workloads on their own terms.”
The company says on-prem inference has data sovereignty and potential cost advantages as, with the cloud-based inference model and per-token pricing, the more useful an AI workflow becomes, the more it costs. Also model versions and quantisation can change at the provider's discretion, affecting the workflows built on them.
Scality’s AI Inference Factory supports specialised chatbots, AI-assisted software development, agentic applications and other inference-intensive workloads. It combines 4 software layers in a stack:
Validated open-weight models maintained as part of the supported stack;
Disaggregated inference serving that separates prefill from decode so each can scale independently;
A control plane that authenticates and meters requests, routes them to the appropriate GPUs holding the relevant context and schedules workloads against SLA targets; and
Scality ADI, which provides high-performance object storage for models, enterprise data and inference state.
Scality’s ADI (Autonomous Data Infrastructure) was launched in May as an enterprise-focussed data infrastructure management product called ADI (Autonomous Data Infrastructure) to place data in four performance, cost, and protection storage tiers with policy-driven AI agent workers.
ADI integrates with standard AI stacks and deploys at multi-petabyte scale on NVMe, with RDMA access to hard disk drives for virtually unlimited KV cache capacity. A single namespace automatically tiers data across TLC, HDD and, optionally, tape, allowing organisations to align storage performance and economics with different stages of the AI data lifecycle.
It says ADI serves as the shared storage layer for the AI Inference Factory, extending key-value (KV) cache beyond GPU high-bandwidth memory (HBM) capacity limits. This enables an organization’s AI inference infrastructure to persist and retrieve model context instead of requiring GPUs to recompute it, improving GPU utilisation and reducing inference costs.
Scality clams ADI provides a shared, multi-petabyte cache that GPUs read at latency the same order of magnitude as GPU memory, meaning the operation’s latency is within roughly the same factor-of-ten range as a GPU’s memory access, but obviously slower, “allowing them to retrieve existing context and remain focused on inference.” We understand this is the same as every other KC caching scheme.
Co-founder and CTO Giorgio Regni said: “The KV cache on Scality ADI is fast enough to sit in the serving path. Restoring a context from ADI is the same order of magnitude as GPU memory, and 14 times faster than recomputing it, with the GPUs staying busy the whole time. Storage is no longer the reason to keep the KV cache inside the GPU server.”
He added: “That opens the door to disaggregated serving. One pool of GPUs handles prefill, processing the prompt and writing the KV cache to ADI. A second pool handles decode, reading the cache back and generating tokens. With Scality ADI as the shared cache, any decode GPU can pick up any context, and there is essentially no size limit. Prefill no longer interrupts decode, and both pools run at full load.”
As Regni says, the AIIF architecture separates AI Inferencing prefill from decode, with ADI acting as the shared cache between GPU pools. Prefill computes the KV cache for a new prompt, and decode generates the answer, one token at a time. When both run on the same GPUs, every new prefill interrupts the decodes already running. Once split, each pool does one job full time and scales on its own, with both working from the same shared KV cache. This allows available decode GPUs to access existing context without tying it to a specific GPU server. Scality cites two examples of this separation improving inference performance:
DistServe (OSDI 2024) measured up to 7.4 times more requests served within the same latency, moving the cache directly between GPUs. In Scality's design, prefill writes the KV cache to Scality ADI, and decode reads it back, so any decode GPU can pick up any returning context, and the tier puts essentially no limit on the size of this cache.
Moonshot AI's Mooncake reported 75 percent more requests on Kimi's production traffic by adding a shared KV cache pool.
Scality testing demonstrated that ADI can reduce GPU consumption while maintaining the same level of performance:
1.9-second load time for Gemma-3 27B, almost 10x faster than local NVMe, with the model streamed in parallel across the cluster over RDMA;
166 ms warm time-to-first-token on a 14K-token context restore from ADI, only 83 ms behind HBM;
14x faster KV cache retrieval than recomputation on a 14K-token context and up to 72x faster on a 439K-token context;
A KV cache more than 80x larger than a single GPU’s memory, keeping up to 1,000 concurrent sessions resumable without recomputation;
No measurable impact on token generation, as context is restored before the first token;
97 percent of network line rate for data transfer between GPUs and storage.
Scality validates and maintains the complete stack. It’s compatible with open-source harnesses and agentic frameworks such as OpenCode, Hermes, Goose, LangGraph and Pydantic AI, and runs validated open-weight models such as Mistral, Gemma, gpt-oss, Qwen, Kimi, GLM, and DeepSeek. The software deploys on standard servers from Dell, HPE, Lenovo and Supermicro.
The stack components are open code, meaning customers can inspect how inference state is stored and moved, and submit contributions for Scality review.
Scality AI Inference Factory packages the whole serving stack behind one endpoint. It includes vLLM, Dynamo, the Scality AI Connector, Scality ADI, and more. This AI Inference Factory is available now as a software license or as a fully managed service. Check out a technical deep dive into the AI Inference Factory architecture and test results here.
Virtually all filesystem storage suppliers are supporting Nvidia's STX KV Caching scheme. This specifies hardware configurations requiring BlueField-4 DPUs, Spectrum-X networking, and dedicated CMX flash storage tiers. Think DDN, Everpure, HPE, Hitachi Vantara, IBM, NeApp, Nutanix, VAST Data and WEKA. Object storage (S3) Nvidia CMX partners include Cloudian and MinIO (with its petabyte-scale MemKV caching system.)
Cloudian’s HyperScale AI Data Platform (AIDP) is a turnkey, on-premises appliance that’s an alternative to public cloud S3 data sources. It’s an S3-compatible, RDMA-based, data store for AI models and agents running on Nvidia Blackwell GPU hardware and software and BlueField DPUs, and complying with Nvidia’s AI Data Platform reference architecture. It lets customers run production AI on their own IT infrastructure, keeping sensitive data under their direct control. Cloudian says it saves up to 60 percent on cost by avoiding the recurring token, egress, and inference fees that make public cloud AI expensive at scale.
This is the same kind of message as Scality’s but without the prefill-decode separation.
MinIO says that, with MemKV, an entire GPU cluster can access a common pool of context at microsecond latencies that keep pace with inference, rather than waiting on millisecond-latency. Again, this is the same kind of message as Scality’s but without the prefill-decode separation.
Mooncake is the serving platform for Kimi, an LLM service provided by Moonshot AI. Both the Transfer Engine and Mooncake Store are open-sourced. Mooncake features a KV Cache-centric disaggregated architecture that separates prefill and decode clusters. It also leverages underutilized CPU, DRAM, and SSD resources in GPU clusters to build a disaggregated KVCache pool.
]]>There is an enhanced version of the GP Series SSD using gen 3 XL-FLASH with >20 million IOPS 512-byte random read performance. The GP1 announced in August offered around 10 million IOPS. XL-FLASH, provides higher IOPS, finer-grained 512-byte data access, and lower power consumption per I/O compared to conventional SSDs. This gen 1 XL-FLASH uses, we understand, MLC (2 bits/cell) with a faster SLC (1 bit/cell) mode, a 16-plane architecture, and read latency under 5 microseconds.
Kioxia America’s Neville Ichhaporia, SVP and GM of the SD business unit, said: “Kioxia is making game-changing high-performance flash storage solutions for AI infrastructure. As AI drives new performance demands, scalable, efficient storage becomes the foundation for innovation. We’re helping data centers build that foundation by pioneering storage solutions designed specifically for AI-driven infrastructure."
Gen 3 XL-FLASH triples read performance and raises write performance by about 150 percent. It uses still shorter word lines and bit lines to reduce RC delay, and more planes per die than Gen 2’s 16. Energy efficiency is expected to improve by about 40 percent for reads and 80 percent for writes. Because performance scales faster than efficiency, energy per I/O is still projected to rise.
The LD4 is a single port, QLC (4 bits/cell), E1.L format drive with a PCIe 5 x 4 lane interconnect, with ordinary capacity options of 15.36 and 30.72 TB. We say ordinary because these are not high-capacities for QLC drives which are now up at the 246 TB level with Kioxia’s own LC9. The LD4’s architecture is valid for extension to 122.88 TB according to Kioxia and it’s meant for hyperscalers and large data centers.
We think there is a lot of space inside the E1.L enclosure and that Kioxia’s 15.36 TB capacity makes it ¼ full, 30.72 TB ½ full, and 122.88 TB completely full, with a likely 61.44 TB offering coming at the ¾ full point.
The LD4 is built from Kioxia’s BiCS 8 218-layer flash. Kioxia had an LD2 drive back in October 2023, with a PCIe 3 interconnect, the same 15.36 and 30.72 TB capacities and E1.L format. We think it used BiCS gen 5 112-layer flash in a QLC format. There was no LD3 drive and now we have the LD4 as a BICS 8 incarnation of the LD2 with a faster (PCIe 5) bus.
Competing E1.L PCIe 5 SSDs are Micron's 6550 ION, DapuStor's R6060 and the FADU Echo.
Kioxia has not revealed any performance or endurance numbers for the LD4. We’re told it supports NVMe 2.0, has support for OCP Datacenter NVMe SSD Specification 2.6, and end-to-end data loss protection.
The company is also demo’ing its liquid-cooled NX1 Series E1.S NVMe SSDs for data centers.
The OCP 2026 Global Summit takes place at the San Jose Convention Center, CA, from October 12 to 16.
]]>Arcserve Ransomware Protection and Recovery Suite – Core: Arcserve UDP Premium plus Arcserve Cloud Cyber Resilient Storage. AI-driven anomaly detection and encryption recognition flag mass encryption and bulk deletion inside recovery points before a restore decision is made. Immutable cloud storage holds an offsite copy with retention locks. Automated recovery testing measures restores against recovery point and recovery time objectives.
Arcserve Ransomware Protection and Recovery Suite – Pro: Includes all Core capabilities, plus Arcserve Disaster Recovery Runbooks, which deploy just-in-time temporary cloud resources and test hands-free on a schedule, surfacing configuration issues before an incident. It also adds the Rapid Ransomware Response Service: an Arcserve Technical Incident Commander takes ownership of the case within 30 minutes of a breach report, 24x7x365, and runs a full backup integrity audit to identify the last known clean recovery point.
Pricing: Both tiers cost less than buying the included products and services separately.
Upgrades: Existing UDP customers can upgrade without changing their UDP configuration.
Find out more here.
…
CrashPlan announced backup and recovery for Windows and Linux servers and Google Workspace, built for small businesses, extending its SMB offering to protect the full range of data their operations depend on.
CrashPlan Windows and Linux Server Backup with simple deployment and administration, minimal maintenance and infrastructure overhead, and secure, affordable cloud storage that scales as needed.
CrashPlan Backup for Google Workspace lets SMBs back up and recover their Google Workspace data and, as a result, they can meet compliance requirements, protect against ransomware, and get employees working again quickly after data loss.
CrashPlan Windows and Linux Server Backup and CrashPlan Backup for Google Workspace are generally available now, each with a 14-day free trial.
…
Cloud file services and data manager CTERA announced CTERA Forward Deployed Engineering, a new service that embeds CTERA engineers within customer teams to bring AI use cases into production using organizations’ own unstructured data. The service helps enterprises prepare governed data, build AI workflows around existing business processes and develop the skills to independently operate and extend those workflows.
Small, dedicated teams, pairing a Forward Deployed Engineer and Deployment Strategist, work as part of the customer’s project teams, with access to the environments and to the subject-matter experts who know how work actually gets done in an organization. Engagements follow a phased approach with discovery and data grounding, a first production use case, scaling across additional workflows, and handover for independent operation.
…
There will be a Future of Memory and Storage Europe conference event in the RAI center, Amsterdam on 19-20 October 2027. The organisers say they’re building a dedicated meeting place for the European memory and storage community, covering everything from HBM, CXL and NVMe through to the technologies underpinning the next generation of AI, cloud and high-performance computing. They hope to see 1,500+ attendees, 150+ speakers and 60+ exhibitors. Find out more by enquiring here.
…
Hitachi Vantara announced Common Criteria certification for its VSP One Block storage product range, providing independent validation of critical security capabilities, backed by a guarantee of a 100% clean recovery point after a cyberattack. Details include certification covers VSP One B20, validating authentication, encryption and trusted communications. The wider platform combines immutable data protection with ransomware detection.
…
Percona announced Valkey-proxy, a new open source project that helps enterprises adopt Valkey and migrate from Redis. Valkey is a high-performance, in-memory key-value data store used for caching and for database workloads. Valkey-proxy sits between the application and Valkey instances, manages communication between them, and presents a cluster to the application as if it were a single standalone deployment. It also pools backend connections, so connection counts stay bounded as applications scale. Commercial Redis offerings use a similar approach, but Valkey-proxy is fully open source. That matters for organizations concerned about vendor lock-in or about data and digital sovereignty. The code will be published in mid-October under the BSD-3-Clause license, followed by a public feedback and testing period, with general availability targeted for early 2027. More information here.
…
Rafay Systems provides the orchestration layer neoclouds use to carve shared GPU fleets into isolated, self-service environments for each customer, with provisioning, quotas, and metering handled through APIs. Sharon AI, for example, signed a five-year deal to run up to 150,000 GPUs on Rafay. compute-first. The tenant carving extends to storage, though as an orchestrator rather than a storage system. Rafay doesn't replace the storage layer. It plugs into what the operator already runs (DDN, WEKA, VAST, Dell PowerStore and others) and handles the storage side of each tenant automatically. When a neocloud spins up a new customer, Rafay sets up that tenant's GPUs, network and storage together, keeps each tenant's storage isolated from the others, and applies quotas and access controls. That way storage gets carved up per tenant in the same workflow as compute, rather than being a separate manual job for the storage team.
Rafay says operators want tenant isolation and automation across the whole stack. Storage platforms are adding multi-tenancy, but something still has to tie the storage tenant to the compute tenant, the billing and the customer's self-service portal. That's the layer it sits in. Learn more here.
…
A blog from Rubrik’s CTO shares how it “turned unverified vulnerability discovery with Anthropic's Mythos Preview into safe remediation at machine speed.”
…
Samsung Electronics announced preliminary Q3 earnings of ₩195 trillion ($145.3 billion), up 127% Y/Y, with its highest-ever operating profit of ₩107.4 trillion ($80 billion), up 782.5%. It's now had three successive quarters of booming revenues and the full year revenue number will likely exceed ₩700 trillion ($521 billion), a more than 100% increase Y/Y. AI GPU server demand for HBM and DRAM was most probably the main driver of the rise. Full results should be released at the end of October.
…
HA and DR supplier SIOS Technology appointed of Charlie Mantione as Head of Sales, Americas and promoted Director of Marketing Natalia Casey to head global marketing. Mantione has previous sales and marketing experience at NCR, Cushman & Wakefield, Avaya, Dimension Data, and AT&T.
…
Archive software and tape library systems vendor Spectra Logic announced a distribution partnership with ICOS S.p.A., a value-added distributor specializing in IT infrastructure and cybersecurity solutions in Italy. The agreement expands Spectra Logic’s presence in the Italian market and enables ICOS to offer Spectra’s data storage and management solutions to its network of channel partners.
…
UnicornIQ today announced GA of its flagship product, Vera, “the Verified Intelligence Platform that makes AI tell the truth. … Vera cleans a company’s unstructured data (the documents, slide decks, chats and emails where most knowledge actually lives), before any AI uses it. These stale assets are often riddled with dirty data, but Vera pulls out each individual claim, checks it against everything else within the company knowledge base, and resolves conflicts or outdated facts up front. The result is AI that answers from verified knowledge, with sources anyone can check, used as an operational backbone.”
Vera is aimed at sales reps and UnicornIQ says “Clean data is premium fuel for reps and partner teams to close complex deals, and Vera is the hypercar that gets across the finish line faster.” Read a frontier lab case study here.
…
SW RAID supplier Xinnor has released xiRAID Classic v4.4.1, introducing discard support designed to help read-intensive NVMe SSDs maintain more consistent write performance over time. The new capability allows discard information to flow from the file system through the RAID layer to compatible NVMe drives, helping them manage available flash more efficiently and reduce unnecessary garbage-collection overhead. The release is particularly relevant for high-capacity NVMe infrastructure used in HPC file systems, virtualized and containerized environments, databases, and backup repositories.
…
Xsight Labs, which produces an E1 800G DPU used by Hammerspace, announced its E1L, the next member of its E-Series DPU line. It's a fully programmable DPU with up to 200G of networking that starts at 25 watts TDP, built for edge, embedded and other power-constrained systems, and it runs the same software as the 800G E1. The SPDK is standard Linux and there is line-rate AES-XTS storage encryption, at a power budget that fits a storage appliance built around a fixed power supply. In many E1 deployments the DPU is the host, and Xsight expects that to be even more common with the E1L
…
]]>The payments involve Micron paying Netlist $30 million a quarter until the third quarter of 2031, and Micron has also bought $1 million of Netlist’s common stock. Netlist has now won IP disputes against Micron, Samsung and SK hynix over its memory technology, gaining fees valued at $1.54 billion.
Netlist founder, chairman, president and CEO Chun K. (Chuck) Hong said: “The agreement with Micron further validates the value of our AI memory technologies. We remain committed to protecting our IP while continuing to invest in the development of next-generation products.”
In August this year, Samsung agreed to pay Netlist $897 million over 5 years in licensing fees for Netlist server DIMM and High Bandwidth Memory technologies just 3 months ago, and became a Netlist shareholder, buying 10 million shares of Netlist common stock for $1 million. The licence fee payments involved an upfront patent license fee of $239 million (about $200 million net of Korean withholding tax), plus quarterly payments of up to $32.9 million gross through to the second quarter of 2031.
It also gave Netlist the right to purchase up to $1.5 billion of DRAM and NAND products over five years.
In April 2021 Netlist prevailed in a patent infringement lawsuit against SK hynix, winning a $40 million settlement, a cross-licensing deal and a supply arrangement.
Netlist has an outstanding legal case against Google since 2008 over U.S. Patent No. 7,289,386, covering rank multiplication in memory modules.
Netlist was founded in 2000 and is based in Irvine, CA. It provides high-performance modular memory subsystems to the world’s premier OEMs, manufactured in wholly owned, ISO- and OSHAS-certified manufacturing and testing facilities in Suzhou, China. Netlist also builds NVMe SSDs. Netlist holds a portfolio of patents, many seminal, in the areas of hybrid memory, storage class memory, rank multiplication and load reduction, among others.
It went public on Nasdaq in November 2006 and has been largely loss-making since then.
Revenue peaked near the IPO, collapsed after 2008, and bottomed at about $8 million in 2015. The 2021–22 jump included the SK hynix license payment plus stronger product sales; 2023 dropped again before recovering in 2024–25. Through the first half of 2026, quarterly sales were about $105 million and $110 million, so the full year is already ahead of 2025 but is not complete. The August and October 2026 Samsung and Micron license payments are not in the 2006–2025 totals.
Losses widened in 2022–2024 on weak product margins and heavy litigation costs, then narrowed in 2025. The first half of 2026 was profitable (about $8.6 million in the first quarter and $1.4 million in the second), before the larger Samsung and Micron license payments.
]]>The common thread between migration, DR, and backup is workload replication, and Matters pioneered a now-patented approach to replication that replicates workloads at the filesystem level rather than using a VMDK conversion approach or sector-level storage replication. There are two versions of the platform: one for physical and virtual servers and the other for Kubernetes clusters.
The company reckons that, at a high level, much of the DR market, including platforms such as Zerto, Veeam, Commvault, and native solutions like VMware Site Recovery, have traditionally relied on replication tied closely to the virtualization or hypervisor layer. RackWare’s replication operates at the OS and file-system layer, which means it isn't dependent on the source hypervisor or its underlying SDKs to move or protect a workload, freeing the software from any constraints imposed by source or target environments.
Workloads can be replicated between any two hypervisors, public or private clouds, or bare metal, and, with the Kubernetes version, between any two K8s distributions. Attempts by platform vendors to constrain workload mobility out of their platforms (most recently Broadcom’s removal of the VMware VDDK) are irrelevant at the filesystem level.
There are other technical advantages to filesystem-level replication, but cross-platform mobility is what’s driving the most business for RackWare lately.
We recently looked at SimplyBlock and its business continuity and disaster recovery functions for Red Hat OpenShift deployments to help with VMware-to-OpenShift migrations. How does it compare to Rackware?
They both target OpenShift continuity, but they protect different layers. SimplyBlock is about storage-level replication for clusters that already run on SimplyBlock. RackWare SWIFT is workload-level replication that can move or recover OpenShift applications onto almost any other Kubernetes platform, including a different storage system.
RackWare SWIFT, part of the RackWare Management Module, sits outside the clusters and replicates Kubernetes and OpenShift applications: namespaces, pods, configs, secrets, and persistent volumes, stateful and stateless. It is agentless, uses a delta-sync engine, and encrypts transfers. Supported targets include OpenShift, EKS, AKS, GKE, OKE, Rancher, and others, including cross-cloud and cross-version replication.
It’s suggested that you use SimplyBlock DR when the production and recovery OpenShift clusters will both run on SimplyBlock and the requirement is low-RPO volume protection without a per-node DR license. Use RackWare when the job is to replicate or fail over the applications themselves across clouds, Kubernetes distributions, or from VMware, and the target storage can be whatever that platform already provides.
Datadobi also offers migration facilities but it and RackWare are not substitutes. A VMware-to-cloud project and a NetApp-to-Isilon NAS refresh are different jobs. RackWare moves whole workloads; Datadobi moves unstructured files and objects.
RackWare was founded in 2009 in Santa Clara, California (later associated with the San Jose / Fremont area) by Sash Sunkara and Todd Matters. It started with angel funding and a narrower goal around storage-array management with LSI’s Engenio unit; after that partnership ended around 2011, it shifted to hybrid-cloud migration, disaster recovery, and workload mobility. Sunkara was the CEO until 2019 when he resigned. Todd Matters remains co-founder and CTO. Bryan Gobbett is the current CEO, appointed in 2019.
It’s had angel, seed, Series A and Series B funding rounds, accompanied by various loans, and has raised between $16 million and $23 million in total, with the last known funding event in 2021. Investors include Signal Peak Ventures, Oquirrh Ventures, Osage Venture Partners, Royal Street Ventures, Frazier VC, Park City Angels, Run Ventures and Kickstart Seed Fund.
]]>DBtune is headquartered in Malmö/Lund, Sweden, and was started up in 2020 by machine learning researcher Dr.Luigi Nardi as a spin-off from Stanford University and Lund University. It received startup grants from, we understand, the Wallengberg Foundation and Vinnova in the 2020-2021 period, and raised $2.4 million in a July 2023 seed round from VCs and angel investors.
The aim is to use agentic AI and optimize for a specific workload, use case and machine, using server parameter tuning, index optimization and autovacuum monitoring. DBtune says it tests PostgreSQL server parameter configurations against the workload and machine type and applies the best-performing settings automatically, or lets the user approve changes first. DBtune v 4.0 uses machine learning and AI to dynamically adapt server parameters, recommend optimal indexing strategies, and monitor autovacuum operations, based on specific workloads and hardware constraints.
Nardi said: “As … PostgreSQL environments grow, maintaining performance and stability requires increasing amounts of specialist attention, manual intervention and infrastructure. At the same time, organizations are under pressure to manage rising cloud costs while preparing their technology stacks for increasing AI adoption. We believe Autonomous PostgreSQL is the next step: using Agentic AI to reduce repetitive management work during database operations.”
He believes that “Achieving that requires a fundamentally different approach to database optimization — one DBtune has spent a decade researching and several years turning into a product. DBtune 4.0 advances that vision by expanding beyond server parameter tuning, bringing index optimization and autovacuum monitoring into a unified approach to Agentic AI for Autonomous PostgreSQL optimization.”
It identifies and ranks index opportunities based on the workload, showing the expected impact, cost and DDL (Data Definition Language). This includes recommendations for new indexes that can significantly speed up specific critical queries, as well as existing indexes that may require maintenance due to bloat or removal because they are obsolete or duplicates. DBtune also identifies where autovacuum is falling behind and helps diagnose what is causing the problem.
Midwest Tape, a DBtune customer, is a full-service media distributor that provides physical and digital content exclusively to public libraries across North America, Australia, and New Zealand. It uses a dedicated PostgreSQL replica to provide metadata to public libraries with its Hoopla digital content service. Midwest used DBtune for AWS RDS PostgreSQL performance tuning to resolve database query bottlenecks and achieved a 10.8x performance improvement, with average query runtime decreasing to 7 ms from 75.9 ms.
DBtune has benchmarking data following a benchmarking test with Intel Xeon processors to determine whether autonomous optimisation could uncover meaningful performance gains beyond extensive manual parameter tuning. It produced three main results:
• DBtune increased transaction throughput by 87 percent during the initial optimisation phase
• It delivered a further 17 percent improvement on an already heavily optimised system
• It reduced disk I/O activity by 43 percent and WAL generation by 23 percent.
DBtune autonomously identified the dominant bottlenecks, validated their root causes through experimentation and translated those findings into measurable performance and efficiency gains, without changes to the application, schema or hardware.
This autonomous optimisation process delivered a 17 percent throughput improvement on a system that had already undergone significant manual optimisation, which suggests that autonomous optimisation can uncover performance headroom even when traditional tuning efforts have already addressed the most obvious opportunities.
Fully autonomous Postgres is still some way off, but Luigi argues that spiralling cloud costs and the complexity of managing growing application and data demands is increasing the importance of using AI to optimise database instances. While PostgreSQL is already well established in the enterprise market, Luigi believes it will take significant adaptation of the source code to enable autonomous optimisation, but if the PostgreSQL community gets it right it will create an advantage over the likes of Oracle.
Download a product paper here, the MidWest Tape case study here, and access DBtune presentations here.
Autovacuum is a PostgreSQL background daemon (process) that aims to reduce worker-level IO by automatically cleaning up dead rows, reclaiming storage, and updating query planner statistics.
]]>Mainframe-to cloud and data integrity supplier Precisely says it’s built on a single metadata and semantic foundation, and provides data integration, quality, governance, location intelligence, master data management, and customer communications capabilities to create a trusted data layer for AI models and agents. Enterprise data is spread across cloud environments and data lakehouses, on-premises systems, and established platforms like mainframe and SAP. The Precisely Platform provides a unified way to help manage and improve data where it lives, without requiring migration to a single data ecosystem. Its integrated data management capabilities share services for semantics, data cataloguing, observability, data pipelines, workflow, and AI management.
Precisely Chief Product Officer Matt Waxman, ex-Arctera and Veritas, said: “Agents don't stop to question the data they're given. They act on it at scale, far faster than people can catch and correct errors. The challenge is knowing what data is being used, whether it’s fit for purpose, and whether the right governance and accountability are in place. The Precisely Platform gives business leaders the visibility and evidence they need to scale AI responsibly and with the confidence to prove it."
Precisely has history, 58 years of it, going back to Whitlow Computer Systems which was started up in 1968 in New Jersey, and its SyncSort product, faster than IBM’s in-house mainframe data sorting software. It became a popular independent sort product. Whitlow rebranded as Syncsort (lower-case second S) in 1981. It bought Trillium in 2016, and Pitney Bowes’ software and data business in 2019 to expand from its core mainframe and IBM i/AS400 sorting and data moving-to-analytics-platforms. Pitney Bowes gave it the ability to clean data, match it, enrich it with 3rd party information, and add location context. It rebranded again to Precisely in 2020, because the company pitch was more bout data integrity across integration, quality, enrichment, and location, than mainframe and i/AS400 connectivity.
The company is owned by private equity, including Clearlake Capital and TA Associates and its core pitch is now about unlocking mainframe data for analytics and AI while leaving transactional workloads in place.
It’s now unveiled three offerings aimed at making the data it presents to AI models and agents better quality. The Precisely Platform has built-in data enrichment and location intelligence, and provides hundreds of pre-linked datasets including location and business attributes. Address verification and geocoding validate, cleanse, and standardise addresses and assign coordinates to each record. Spatial analytics help users discover and visualise patterns, trends, and relationships between location-based datasets. These capabilities create a connected, machine-readable model of the world that AI can use to make more informed decisions and take more effective action.
User staff can access and apply Platform capabilities directly within the tools, workflows, and AI models they already use, without changing how they work. They can do so using Gio, the Precisely Platform AI assistant, as well as APIs, SDKs, and a command-line interface.
New Model Context Protocol (MCP) servers, also announced today, extend connectivity to multiple products across the Precisely portfolio, allowing teams to connect the Precisely Platform and their existing Precisely investments directly to MCP-compatible AI clients like Claude, Microsoft Copilot, and ChatGPT.
Initial MCP server support enables MCP-compatible AI agents and assistants to work across the Precisely Platform suite, and includes:
Precisely Automate initiates and interacts with SAP automation workflows while maintaining the organisation's own governed business processes.
Precisely Data360 prepares and analyses data using Data360 Analyze to create, inspect, update, and run data flows.
Precisely EngageOne connects customer communications workflows using EngageOne RapidCX, including triggering new communications, inspecting communication activity, and customer engagement history, with all requests scoped to the user and recorded in an audit log.
Precisely EnterWorks enables natural-language search across master data, helping users find and understand data without navigating MDM structures.
Precisely Spectrum enables trusted address verification, geocoding, data enrichment, and job and dataflow management.
Precisely Trillium helps cleanse, match, and validate records using trusted business rules.
Precisely Syncsort optimises mainframe sorting; explaining messages, troubleshooting jobs, and generating sort, copy, and merge logic grounded in Syncsort MFX documentation and knowledge base articles. These capabilities will be available in the coming months.
Precisely also announced the availability of Precisely AI Studio, a collection of ready-made apps, agents, and skills that help builders develop AI offerings. The assets in AI Studio work with AI tools, including Claude, Microsoft Copilot, and ChatGPT. Builders can avoid starting from scratch, with working examples that show how to make data AI-ready, using capabilities available in the Precisely Platform, such as data quality and enrichment.
The Precisely Platform is in preview and will be generally available in early 2027. For more information go here.
Precisely’s AI Studio is available immediately. New users can sign up for a free trial.
Bootnote
Precisely has more than 12,000 customers in over 100 countries, including 92 percent of the Fortune 500,
]]>The capabilities of its Operational File Data Platform for the AI era are largely based on two acquisitions it made earlier this year. Nasuni bought DryvIQ in August to add AI-powered content intelligence and governance capabilities to its File Data Platform (FDP). Earlier it had bought Resilio in April to add high-performance file synchronisation and edge acceleration technology to its FDP. At that time, it declared it had significantly evolved from its original collaboration and cloud file services era and was now an unstructured data platform for enterprise teams and AI. Its updated software can manage, protect, and activate unstructured data at scale from a single global namespace. People and AI work from the same current data.
CEO Sam King pointed out: “Many organisations have achieved broad adoption of LLMs, but turning that usage into measurable business value is proving much harder. Rising token costs, limited production-scale deployments, and data governance that has not kept pace are exposing the real challenge: your advantage is not the model, but your proprietary data, your business context, your enterprise intelligence.”
She said Nasuni is now about “helping enterprises understand the data they have and [putting] their own intelligence to work securely. Today, we are delivering on that promise with capabilities built on our proven file data platform architecture.”
The new and generally available capabilities are:
Content Intelligence and Governance: Via DryvIQ, Nasuni has the ability to understand, manage, and protect file data across more than 40 repositories, including Microsoft 365, Google Drive, Box, Amazon S3, and on-premises NAS. DryvIQ classifies content across 550 file formats and 75 languages, identifies PII, PHI, and PCI data, and scans hundreds of TB per day. Enterprises can understand what they have, identify and classify what is valuable, risky, redundant, obsolete, or trivial (ROT), and apply automated governance policies across their data estate.
Governed AI Access: AI Activate connects AI directly to enterprise file data through the open Model Context Protocol (MCP). It’s powered by a metadata and content index, with permissions enforced before AI sees any content, allowing third-party MCP-compatible AI platforms and agents to work from live operational files rather than giving them a specially curated separate copy.
Enterprise Search: This adds fully hosted, permission-aware, full-text search across the operational file environment, making the same enterprise knowledge discoverable by people and AI agents. It will be made available soon.
High-Speed Edge Delivery: Resilio 6.0 extends the Nasuni platform to the far edge, using peer-to-peer distribution to move and deliver governed file data at speeds of up to 200 Gbps on commodity hardware. There is no need for any VPN, dedicated infrastructure or caching appliances. Staff at construction sites, studios, field offices, and survey vessels can work with globally-managed file data at local speed. (McKim and Creed consolidated 27 file servers onto Nasuni, with remote engineers reporting a 50 percent productivity improvement in early Resilio trials.)
Chief Product Officer Nick Burling said: ”A global namespace, immutable versioning, native permission enforcement, and customer-owned object storage let enterprises meet sovereignty and residency requirements without fragmenting their environment into regional silos. It also means security teams are not approving a new data pipeline every time the business adopts an AI tool. The permissions and controls already in place carry forward automatically, to people and to agents alike."
As for Nasuni's so-called PSYCHIC framework to help define what AI-ready data should mean, it includes:
Proprietary: Can AI access the unique data, knowledge, and context of your business?
Secure: Can AI use the unstructured data safely?
Yours: Do you maintain control over what happens with the data?
Current: Is AI working with the latest version of what is happening now?
Hybrid: Can you bring your AI to your data wherever it resides?
Indexed: Can AI find the right information, quickly and efficiently?
Continuous: Can AI access and use the accumulated knowledge embedded in years of files and projects?
If you are not psychic (Sorry. Couldn’t resist it.) you can learn about PSYCHIC here.
All the cloud file services businesses, such as Box, CTERA, Egnyte, Nasuni and Panzura are enabling their customers to make the data they store available to AI models and agents with their own versions of content intelligence, governance and controlled permissions, and the common factor of MCP. It has become table stakes. AI models and agents are becoming universal accessors of data.
This could eventually drive a consolidation of cloud file services-fed edge storage suppliers, on-premises data center storage providers and distributed data management suppliers. All three kinds of suppliers are wrestling with the problem of how to give AI models and agents access to the widest possible distributed data estate without losing their own relevance or facing massive costs to extend their reach to parts of the data estate in which they have no presence.
]]>TDK is the only independent supplier of HDD recording heads and sells to all three drive makers—Seagate, Western Digital, and Toshiba. The latter depends on TDK for all of its heads, including technology needed for its HAMR plans. But Seagate and Western Digital use TDK for surge capacity as they mostly make heads in-house.
With a single main customer, which has a low share of the disk drive market, 10 - 11 percent, TDK’s disk drive head business is viewed as a small operation with low profitability, compared to its other divisions producing batteries, passive components, and sensors. Overall, the unit’s business performance is volatile and capital-intensive, and secondary reporting puts its return on invested capital around 4 percent.
TDK’s July 2026 most recent business performance report said “Sales of HDD heads and HDD suspension assemblies increased due to continuous robust demand in the HDD market, resulting in an increase in profitability.”The demand comes from AI data centers and sales are expected to increase.
Full year sales of TDKs Magnetic Application Products segment (HDD heads, suspension assemblies, and magnets) in the fiscal year ended March 2026 were ¥262.9 billion ($1.74 billion), up 17.6 percent Y/Y, with operating profit increasing strongly, about 8x, ¥27.0 billion ($179 million). The HDD head portion is not known but will be smaller.
Nearline head volume was up roughly 14 percent and suspension volume about 35 percent.
A TDK statement said “Certain media reports have been published regarding HDD head business of the Company. However, these reports are not based on any announcement made by the Company. The Company is pursuing initiatives, including strengthening its business portfolio management, to further enhance corporate value, but no decisions have been made at this time.”
Neither WD nor Seagate responded to Bloomberg requests for comment.
Were Seagate to succeed in buying TDK’s head business then Toshiba would find itself buying heads from one of its two main and much larger competitors. WD would also effectively find its surge capacity buying option would be closed off. Authorities in the competition area might well have doubts about allowing this to go ahead on that dual basis.
]]>The company says AI models and agents need access to the right data from a disparate and distributed data estate with multiple different silos and data formats widely spread across an enterprise. Its improving the PowerScale file storage applicability to that data estate by increasing its tenant capability to 500 tenants per cluster and also providing a fully-managed, cloud-native version of PowerScale for Microsoft’s Azure public cloud. The multi-tenancy features mTLS over NFS to encrypt and authenticate file traffic and granular role-based access control for managing each tenant.
Dell Infrastructure Solutions Group President Arthur Lewis stated: "Data without context is just noise. Most enterprises have spent years making their data accessible. That's not the same as making it usable. An agent that can find a customer record but doesn't know what it means, how it connects to everything else, or whether it can be trusted isn't intelligent. It’s just fast. The companies that solve it won't just deploy AI faster or run more agents. They'll get more out of the data they already have.”
The company is announcing a Unified Semantic Layer, Enterprise Knowledge Graphs and subject matter-specific Enterprise AI Agents to identify and collect he most appropriate and consistent data for an AI task. Specifically;
Unified Semantic Layer - gives structured and unstructured information consistent business meaning with rules, definitions and a searchable glossary, so a term means the same thing everywhere it appears. If one system calls it a client and another calls it an account, AI now knows they're the same thing. Users can also import existing ontologies and classification taxonomies, reusing industry-standard or enterprise-specific taxonomy assets. Dell is also enabling Nvidia Auto-Ontology, an open-source library that builds knowledge graphs from enterprise data, to extend the Unified Semantic Layer.
Enterprise Knowledge Graph - maps how structured and unstructured data is related, so an application or an agent can find the right context. It uses metadata, lineage and query history to keep tuning the graph as activity changes. When an agent asks a question, the platform pulls in every related piece it's allowed to see: the right tables, data products, multimodal data and vector indexes, wherever they live. For instance, a manufacturer chasing a production line problem can trace one odd sensor reading to the machine, its repair history, the supplier batch and the orders at risk.
Knowledge Agents - act as a trusted advisor on a single topic, grounded in a defined slice of the Enterprise Knowledge Graph. Customers set the rules: what guidance it follows, what data it can see, what quality bar it must clear and how much it's allowed to spend. Nvidia Nemotron Retriever models provide reasoning and visual understanding to Knowledge Agents.
Once data has been initially selected for AI processing then Dell is now offering cuDF accelerated data preparation with Apache Arrow. Nvidia cuDF is an open-source, GPU-accelerated library for processing tabular data, built on Apache Arrow’s columnar memory format and optimized CUDA kernels. Operations such as filtering, joins, group-bys, sorts, and aggregations can run across thousands of GPU cores at once, with less data movement between CPU and GPU.
Dell’s SVP for Product Marketing, Varun Chhabra said: ”Nvidia cuDF accelerates data processing, GPU -accelerated Apache Spark prepares structured and unstructured data at scale, and Apache Arrow moves data efficiently between data storage and processing.”
Dell says its Data Processing Engine, powered by cuDF on Nvidia RTX PRO 4500 Blackwell Server Edition GPUs, processes data nearly 4 times faster on average than CPUs alone across a range of workloads, and up to 20 times faster on batch processing workloads.
Jason Hardy, VP storage technology at Nvidia, said: "By bringing Nvidia cuDF acceleration directly into the Dell Data Processing Engine, Dell helps shorten the path from stored data to GPU-accelerated AI-ready data their agents can use to drive real world innovation.”
A new Dell Storage Performance Tool helps customers size AI infrastructure, compare vendor solutions and keep GPUs fed by testing the performance of their S3-compatible object storage across training, inference and checkpointing. Chhabra said: “It is an open-source tool for Dell Object Scale and Dell PowerScale that measures object storage under realistic repeatable AI workload capabilities. Organisations can define the workload they want to benchmark our storage platforms with. And then [test] checkpoint-style writes, high concurrency reads, mixed read write environments, even Iceberg queries.”
“They can all run these benchmarks on their own infrastructure with the same tool that Dell Engineering is using internally for benchmarking our products. It measures throughput and latency live and verifies persistent data end-to-end, and every result records versions and provenance, so it's a repeatable process.”
Dell Professional Services are being expanded with AI Data Platform implementation services across its data and storage engines. The aim is to establish a production-ready foundation, activating the platform’s analytics, processing, search and orchestration capabilities while keeping the AI Data Platform system tuned for the best performance.
The Unified Semantic Layer, Enterprise Knowledge Graph and Knowledge Agents will be released in 1H 2027.
Dell Data Processing Engine enhancements with the Nvidia acceleration stack will be available in December 2026, with further acceleration using Apache Arrow expected 1H 2027.
Dell PowerScale security and multi-tenancy enhancements will be available in November 2026.
The cloud-native Azure PowerScale service is available now, starting with a US East region.
The Dell Storage Performance Tool is available from GitHub.
Dell says its fully-managed, cloud-native PowerScale on Azure “provides 4x faster performance than our closest competitor, 4x larger namespace, as well as a 2x higher cluster resiliency than our closest competitor.” We understand the closest competitor is Qumulo’s Azure Native Qumulo (ANQ).
]]>Toshiba Information Equipment (Philippines), Inc. (TIP), has three manufacturing sites;
Laguna Technopark (LTI), Biñan - historically internal and specialty HDDs (historically client drives for notebook PCs, external HDDs, multifunction printers, and set-top boxes), and recently high-capacity, nearline disk drives on a newly-expanded production line.
Laguna Technopark Annex (LTI–Annex) - enterprise SSDs, client SSDs, and printed circuit boards (not HDD assembly).
Carmelray Industrial Park (CIP), Calamba - data-center and enterprise HDDs (including nearline and helium-filled drives) for high-end and mid-range servers, storage systems, and data centers.
The 30th anniversary of TIP’s manufacturing operations this month saw Toshiba say it will invest roughly 60 billion yen ($380 million) to expand facilities in the Philippines, according to Japan’s Nikkei media news site. This will enable LTI to build new hard disk drive (HDD) products that increase per-unit capacity by as much as 40 percent, with Toshiba continuing capital spending toward mass production of 65 TB-class drives in 2030 and eventually 100 TB-class drives. Toshiba currently has 11 percent share of the world-wide HDD market, behind Seagate and WD which each have round 44 percent, and intends to achieve a 30 percent share in the medium term; understood to mean 3 to 5 years.
That's ambitious. However, HDD demand currently is greater than supply, with Seagate's 2026 manufacturing capacity sold out and both Seagate and WD having signed long-term supply agreements with major customers to ensure deliveries and pricing of the nearline, high-capacity drives needed to store their bulk, warm data. Theoretically, Toshiba could increase its manufacturing capacity output and capture market share without either Seagate or WD selling less capacity.
The road to capacity increases depends upon recording technologies and a drive's internal platter count. Seagate's drives have 10 platters and it increases capacity by improving the areal density of its Heat-Assisted Magnetic Recording (HAMR) technology, with 4 TB-class platters qualified with some hyperscalers now, 5 TB-class platters in development, and a roadmap out to10 TB-class platters, and hence 100 TB drives, by 2030.
WD and Toshiba both use 11-platter drives, giving them an easier task areal density-wise, and WD has been able to push its enhanced perpendicular magnetic recording (ePMR) technology out to 26 TB with this CMR (conventional magnetic recording) method and 32 TB using shingled magnetic recording (SMR) with partially overlapped write tracks increasing track density, and thereby overall capacity, on a platter without having to increase overall density. It's talked about increasing the number of platters further, with 14 mentioned in February.
In February it also predicted it would deliver 40 TB drives this year, with ePMR and HAMR variants. A 44 TB HAMR drive will follow later this year, with 60 TB ePMR and HAMR drives in 2028 and 100 TB HAMR drives in 2029, before Seagate. It sees ePMR technology being unable to advance much further past 60 TB drives.
Toshiba is currently at the 11 platter level and has no HAMR drives in production, using Flux-Controlled Microwave Assisted Magnetic Recording (FC-MAMR), a variation on PMR, instead. It began sampling 30 to 34 TB, 11-platter M12 SMR disc drives aimed at hyper scales and cloud service providers in March, with sample shipping of 28 TB M12 CMR drives planned for the third 2026 quarter. It has not said if this took place. HAMR technolog is on its roadmap.
It could double the LTI plant's output capacity from 2025 to 2027 by increasing the number of drives it manufactures, and also increasing drive capacity by adding platters and again by upping the areal density per platter.
We think Toshiba will shortly announce that its higher capacity M12 drives are shipping.
]]>Traditional business intelligence (BI) requires analysts to define questions and build dashboards in advance. Newer BI has generic AI assistants infer business meaning from raw tables and schema. Jedify is an AI context graph startup and reckons that enterprise data is fragmented across on-prem, public cloud, and SaaS app silos and systems, structured and unstructured datasources, definitions, permissions, and workflows. Its software reads these distributed. disparate knowledge fragments and uses them to populate a live context graph that AI agents can use to produce more accurate request responses. The graph embodies the semantic context, the relationships between the business entities and items an AI agent deals with in its processing for an organization.
Jedify co-founder and CTO Adi Elimelech said: "Enterprises are moving analytics from answering questions to doing meaningful work. With Data Agents, teams can build an analytical workforce tailored to how their business operates and give employees access to those agents wherever they already work."
The Data Agents product features:
Plain-language agent building. Describe the analytical outcome you want, and Jedify creates and refines the agent conversationally, with no code or workflow builder.
Flexible triggers. Agents run on demand, on a schedule or when a defined condition is met, triggering recurring analysis without someone asking for it each time.
Root-cause investigation. Grounded in Jedify's context graph and research engine, Data Agents monitor metrics, investigate anomalies and identify likely drivers behind a change.
Proactive, not reactive. Data Agents surface a gap, an anomaly, a market that's underperforming, the moment it emerges, instead of waiting to be discovered in a report.
Permission-aware by default. Agents inherit the requesting user's existing data permissions, so an agent cannot reveal or access a row, column or metric that the user isn't entitled to see.
Built for efficiency at scale. Because agents reason from Jedify's pre-resolved context graph instead of exploring raw schema at query time, they use far fewer tokens per question than a schema-inferring agent.
Works where teams already work. Through MCP, a Data Agent is exposed as a tool teams can call from Claude, ChatGPT, Slack or any agent platform they run. Results can also be delivered directly through Slack and email.
Building AI agents using natural language would seem like a common sense idea and is table stakes, just not in the storage and data protection world. Think Microsoft Copilot Studio, Salesforce Agentforce, ServiceNowSI Agent Studio and others. Commvault is one of the few data protection companies with a product here, as its AI Studio has an AI agent builder.
Data Agents is available now, accessible through Jedify's Agents Hub. Learn more here.
Jedify raised $24 million in an A-round in June this year. It was started up in 2013 and took in an $8.5 million seed round later that year.
]]>The enormous rise in purchases by hyperscalers, neoclouds and large enterprises of AI servers, with HBM-intensive GPUs, DRAM-loaded CPUs, and all-flash storage has not just sent Nvidia revenues to the near $100 billion quarter level, it has near quadrupled revenues for US, Japanese and Korean memory and NAND suppliers. The really big winners have been high-bandwidth, memory (HBM) suppliers but NAND-only suppliers have benefitted as well. So much so that its has made a total of $483.7 billion dollars in the past year.
If we tabulate the last four quarter’s dollar revenues by Kioxia, Micron, Samsung, Sandisk and SK Hynix then we see how their revenues have, in total, nearly risen 4x over four quarters.
Charting their revenues shows this even more clearly. The bulk of the money taken in is accounted for by HBM, then by DRAM and, thirdly, by NAND. This is shown by the comparatively low revenues of the NAND-only suppliers Kioxia and Sandisk and a DRAM+HBM vs NAND revenue split for Micron.
The beneficiaries of this humungous dollar transfer to the memory+NAND suppliers are the businesses themselves, of course, but they are spending the money on uprating and expanding their fab cleanrooms and packaging plants, the equipment inside them, building new fabs, bonus payments for their staff, dividends and stock buybacks for their share holders and taxes. The money is put to use.
If the memory plus NAND supply imbalance continues out to 2030, then we face three more years of this and a likely transfer of around another $1.5 trillion to these five companies. Bringing in China with its indigenous memory and DRAM suppliers, CXMT and YMTC for example, and we're likely looking at the $2 trillion area.
Crude oil sales in 2025 were about $2.6 trillion with up to 100 major world-wide operators taking most of the revenues. The worldwide memory and NAND business is smaller but what is unprecedented, we think, about the memory and NAND dollar glut is that it is concentrated across so few companies; less than 10 worldwide.
]]>FlashCore Modules (FCMs) are IBM’s proprietary SSDs, with the gen 5 version being computational SSDs as they feature hardware compression, deduplication, die striping, and cyber-attack detection. The FCM5 uses QLC (4 bits/cell) NAND with an SLC (1 bit/cell) incoming write zone. There are five FCM5 raw capacity points; 6.6, 12.2, 26.4, 52.8 and 105.6 TB, with the maximum effective compressed capacities being 19.8, 36.6, 79.2, 158.4 and 316.8 TB.
Magneto-resistive RAM (MRAM) stores bits in magnetic tunnel junctions rather than as charge levels, and is non-volatile like NAND, has DRAM-class access speed (10s of nanoseconds), is byte-addressable, and has very high endurance, unlike NAND. Everspin supplies a range of MRAM chips and an Everspin MRAM presentation at the SNIA Developer Conference in San Jose in September (SDC26) by Steffen Hellmold included this slide;
Hellmold said MRAM is an ideal write buffer or cache for small written in data storage applications, and a slide showed it being used in this way in IBM’s FCM5 drives. Because it’s non-volatile, the drives don’t need to use super-capacitors to write consistency data to NAND if there’s a power loss. Instead this metadata is held in MRAM and persists through any power loss.
IBM can reclaim the space that would have been needed for the SuperCaps and populate it with more NAND chips, increasing capacity.
There was an indication the AIBM was using Everspin MRAM in an FCM5 board illustration on page 21 of the Redbook entitled IBM FlashCore Module (FCM) Product Guide. This shows an Everspin chip on the board's top right-hand side above the big AMD processor chip.
We understand that IBM has used a few hundred megabytes of Everspin MRAM inside FCMs for several years as a small, non-volatile write cache and journal. In April 2024 Everspin announced that IBM chose its PERSYST EMD4E001G 1 Gb STT-MRAM chip product in its FlashCore Module 4 drives, specifically for data protection on power loss.
Prior FCM3s used up to 128 MB of Toggle MRAM for that purpose.
Everspin has more than 2,000 customers for its MRAM products which have been in production for 20 years with more than 200 million chips shipped. We understand Dell and LSI/Avago use Everspin in some RAID-controllers.
Steffen Hellmold is a fellow at Everspin and also director at advanced archival technology developer Cerabyte. He was previously Everspin's VP for marketing in the 2011-2013 period.
]]>This offering is part of AgentRecover, Commvault’s offering for enterprise AI environment recoverability. Employees are running these agents on their Windows and macOS PCs and laptops. When a laptop pr PC is lost, damaged, replaced, or otherwise disrupted, organisations and employees – developers, marketers, and others – may not only lose files, but the agent configurations, accumulated AI context, and workflows they rely on. Commvault can restore that agent metadata.
Pranay Ahlawat, Commvault’s Chief Technology and AI Officer, said: “Most organisations cannot tell you which AI agents are running on their employees’ machines, let alone recover the work inside them. We give IT and Security visibility into where these AI tools are installed, and a recovery path for the context employees have built inside them.”
IT teams can identify instances of certain agentic AI tools installed across managed devices, protect agent configurations stored locally, and recover that protected information if a disruption occurs. Security teams can get a centralised view of supported AI tools installed across managed employee devices. If agentic AI tools are installed outside approved IT processes, teams can see where the tool is present and investigate.
Commvault's competitors are looking at agent protection as well:
Cohesity Agent Resilience (CAR) protects agent memory and configuration and is a Cohesity Data Cloud capability, supporting Amazon Bedrock AgentCore and Bedrock Agents. It does not support the Anthropic agents or OpenClaw.
Druva provides Claude Backup under its AI Resilience / Endpoints offering. It backs up locally stored Claude Code and Cowork data on Windows and macOS: project files, sessions, tasks, prompts, notes, skills, plugins, connectors, settings, and configuration. Snapshots go to Druva’s air-gapped cloud; admins manage policy centrally. Claude.ai cloud chats are out of scope because they are not on the device.
Rubrik has its Rubrik Agent Cloud for Claude Code and Claude Cowork. It inventories agents, offers Agent Rewind to reverse unintended actions, keeps immutable off-repo snapshots of GitHub and Azure DevOps, and backs up and version-tracks agent configuration (system prompts, tool permissions, skills, settings) at organization, repository, and user levels, with drift detection. That is closer to agent governance and codebase recovery than laptop file backup.
Veeam, with its recent focus on agentic AI cyber-threats will probably follow closely behind Cohesity, Commvault, Druva and Rubrik.
Discovery of OpenAI’s ChatGPT desktop app is available today. Discovery, protection, and recovery for supported local data associated with Claude Cowork, Claude Code, and OpenClaw will be generally available in the coming months. Customers can access these capabilities through Commvault’s existing endpoint protection offerings without a separate SKU or additional purchase. Support for additional AI tools is planned for future releases.
Anthropic’s Claude Cowork is a knowledge work AI agent that can organize a folder, turn notes and screenshots into a spreadsheet, or draft a report from scattered sources, and is available for macOS and Windows systems. Claude Code is Anthropic’s agentic coding tool, which can read a codebase, edit files, run commands, and work with development tools to build features, fix bugs, and automate tasks, rather than only suggesting the next line of code.
OpenClaw is a free, open-source autonomous and general personal or team AI agent you run on your own system. It connects a large language model to messaging apps and local tools so it can act on tasks, not only reply in a chat window. You talk to OpenClaw through channels such as WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and others, or through a web UI, CLI, or companion apps. It keeps session state and memory on your hardware. Models are plug-ins: Claude, GPT, Gemini, DeepSeek, OpenRouter, or local models via Ollama and similar tools.
Commvault has achieved the AWS Data Competency in the Relational Database category by delivering cyber-resilient data protection for relational database workloads on AWS, helping customers protect, recover, and manage business-critical data across Amazon Relational Database Service (RDS) environment.
]]>…
Backblaze introduced updates to Business Computer Backup that change how Jamf admins deploy and manage backup across their environments. The release builds on Backblaze’s existing Jamf integration, adding a more flexible way to onboard devices and users - especially in environments that aren’t perfectly centralized - so protection can be applied reliably across every endpoint.
…
Denmark’s GenomeDK is an HPC facility that supports researchers, students, SMEs, and hospitals with the infrastructure to store, manage, and analyze large-scale datasets, as well as run large scale simulations. GenomeDK’s largest storage system is a 33 PB BeeGFS deployment used as a single mount point for everything, active analysis as well as long-term retention. Unlike a typical HPC environment where a fast parallel file system is reserved purely for scratch space, GenomeDK treats the same BeeGFS instance as both working storage and permanent archive; some datasets on the system are 15 years old. On a fixed annual public-sector budget, the team can't simply buy one platform and run it unchanged for five years, instead, storage is expanded roughly every two to three years, with data migrated onto new capacity and the oldest hardware retired, an ongoing, incremental process rather than a single large refresh cycle. This process is performed in the background with minimal impact on services and is aided by internally developed tooling. Read more here.
…
Belgium-based Corbenic AI’s Galahad is a long term memory layer for AI. A model reads a text once, and Galahad keeps what is computed. When the same text is needed again, Galahad gives the computation back, so the GPU never reads the same text twice. Galahad has a C++ core and runs inside vLLM, SGLang and llama.cpp. The model itself is not changed. It is free for non-commercial use on one GPU; commercial use needs a paid licence per GPU. Testing on messy data from Hugging Face showed:
14× faster on vLLM: 0.59 seconds per question
100 against 77 for RAGFlow on the same questions
99% of the work comes back from memory
91–95% correct answers on messy, real-world data
Galahad software is available at GitHub: github.com/corbenicai/galahad.
…
Neocloud CoreWeave has a 25+ member partner ecosystem with several storage-related suppliers present, not just prominently featured VAST Data. We find LanceDB, Pinecone (vector database) and WEKA (fast parallel file services) also present.
…
Dr. Luigi Nardi, founder and CEO at DBtune has a vision for autonomous Postgres. While he believes database technology for AI agents is still in its early stages, Nardi has undertaken extensive academic research into AI which he is now applying to the database optimisation technology DBtune is developing. Fully autonomous Postgres is still some way off, but Luigi argues that spiralling cloud costs and the complexity of managing growing application and data demands is increasing the importance of using AI to optimise database instances. While PostgreSQL is already well established in the enterprise market, Luigi believes it will take significant adaptation of the source code to enable autonomous optimisation, but if the PostgreSQL community gets it right it will create an advantage over the likes of Oracle.
…
Knowledge graph supplier Graphwise announced major updates to its AI Context Platform, Graphwise Platform, which include new Adobe integration, built-in evaluation tools, and broader workflow automation features. The new release optimizes the creation of a comprehensive context layer for enterprises, which accelerates the integration of AI into multilingual content and data. Enhancements include:
AI Trust & Rapid Onboarding: A new evaluation framework to measure, audit, and benchmark AI response accuracy, paired with automated unstructured data ingestion pipelines that convert raw enterprise documents into grounded context faster.
Enterprise Multilingual Governance: Native Adobe Experience Manager (AEM) integration for automated multilingual content tagging and cross-channel taxonomy synchronization to guarantee term integrity globally.
Unified Administration & Automation: A centralized administrative control center for enterprise user and license management, alongside a rebranded, n8n-powered workflow engine to simplify complex data pipelines and GraphRAG orchestration.
Mission-Critical Reliability: Comprehensive core performance and stability enhancements across the underlying graph database architecture, built to support high-concurrency enterprise deployments.
…
Cybersecurity supplier Index Engines announced new research from its own CyberSense Research Lab showing that ransomware is increasingly moving beyond full encryption and using techniques designed to make corrupted data appear unaffected. The lab's findings highlight a broader challenge for cyber resilience as attackers target the structures that make data usable while suppressing traditional signs of corruption. Among 1,064 ransomware strains acquired and detonated during the first half of 2026, 47.8% exhibited directory-entry destruction, more than twice the 18.3% that exhibited full encryption. The lab also documented new ransomware designed to suppress familiar signs of corruption by preserving file extensions and timestamps, maintaining low entropy, encrypting slowly, and targeting only selected sections of files.
“Bad actors know what scanning tools look for, and the variants we detonated this year are built to hide it," said Jim McGann, CMO of Index Engines. "Encoder is the clearest example. It destroyed files while leaving their names, sizes, timestamps, and entropy unchanged. A surface scan would report that data as clean. CyberSense caught it by analyzing the content and structure of each file. The Research Lab exists to find techniques like this early and build them into the model our customers rely on for recovery.”
Access the full CyberSense Research Lab report here.
…
InfluxData announced InfluxDB 3.12, saying InfluxDB 3 Enterprise tackles a scaling problem as write volumes grow, the background work that organizes data for fast queries can fall behind. With 3.12, teams can add nodes dedicated to compaction, so it scales the same way ingest and query already do. Bulk Parquet imports are faster too, with import times dropping 64-70% across workloads tested. This distributed compaction is in beta.
…
Spanish cloud storage provider Internxt wants to take share from AWS and Google with flat, S3-compatible storage pricing. We said: “Many CSPs operate under AWS' S3 price umbrella with no egress fees - Backblaze and Wasabi for two. So now we have a Spanish startup emulating them and using a US PR agency. What's special here?” Internxt told us its distinction is the combination of predictable object-storage pricing with privacy, security, and European data sovereignty. Internxt S3 costs €7 per TB stored each month, with no ingress, egress or API fees. It is hot storage and compatible with the AWS S3 and IAM APIs. [An] “up to 80%” saving compares with hyperscalers such as AWS, Azure, and Google Cloud, rather than with Wasabi or Backblaze. For S3 customers, data is encrypted in transit and at rest, and the platform supports server-side encryption with customer-provided keys. Customers can also choose whether to store their data in U.S. or European regions. That makes Internxt particularly relevant to organizations looking for an S3-compatible service from a European privacy company, with more control over jurisdiction, encryption, and where their data is held.
We’re told Internxt was founded in Valencia in 2020 by Fran Villalba Segarra. Net revenue grew by more than 100% year over year in 2025, and the company closed the year EBITDA-positive. Internxt is now targeting $10 million in net revenue next year. Its most recent funding round was €3.3 million in July 2025, with participation from Prosegur Tech Ventures, Andorra Telecom and existing investors. That followed a €3 million round in 2024 and an earlier €250,000 round led by Angels Capital. Internxt has also received €1.4 million in public funding through Spain’s CDTI, supported by the Ministry of Science and European funds. Internxt already has a substantial presence in Europe, but the technology and the need for greater control over company data are not limited to Europe. The U.S. is an important growth market, particularly for the enterprise business.
…
Lightbits Labs said its ultra-fast, disaggregated block storage Lightbits SDS software is certified for Red Hat OpenStack services on OpenShift. Lightbits serves as the Cinder storage backend on the OpenShift control plane and also serves as a direct NVMe/TCP target for External Data Plane Management compute nodes. Download a white paper here.
…
The MariaDB Foundation announced its MariaDB Server Ecosystem Hub has exceeded 200 entries, six months after launch. The Hub is a curated, open directory of MariaDB Server approved products, services and open-source projects. The ecosystem spans cloud and hosting platforms, monitoring and observability, backup and recovery, security, migration, connectivity, DevOps, analytics, AI and application development, content management, e-commerce and complete business applications. The 200 entries include database connectors and management tools, managed MariaDB services, hosting providers, monitoring platforms, migration technologies, enterprise applications and open-source projects.
The Hub includes Solution Stacks: curated, production-oriented architectures showing how MariaDB Server can be combined with complementary technologies to address a particular use case. The first published stack, the Privacy-First Collaboration Stack, brings together MariaDB Server, Nextcloud and Passbolt to provide an open-source foundation for file collaboration, password management and data control. Further stacks are being developed around areas including sovereign infrastructure, e-commerce, multicloud independence, high availability and open web performance.
…
Micron has appointed Dr Scott DeBoer as a President and its Chief Technology and Products Officer and also Manish Bhatia as another President and Chief Operating Officer. Bhatia will lead Micron’s global operations and business units, with accountability for the company’s operating P&L, spanning capital investment, manufacturing execution, and customer demand, pricing and delivery. DeBoer will be responsible for advancing the company's memory and storage roadmap, accelerating innovation to meet customers' rapidly evolving requirements, and overseeing Micron Research Labs, a global flagship research hub dedicated to breakthrough memory and compute technologies.
…
Tom’s Hardware reports Micron this month filed a lawsuit against China's 3D NAND champion YMTC, accusing it of poaching its leading engineers, illegally obtaining its NAND memory know-how from them, and then using that information to build its own non-volatile memory as well as its patent portfolio. The Boise, Idaho-based company further alleges that YMTC uses patents granted to its former employees to accuse Micron of patent infringements in various jurisdictions. Read (a lot) more here.
…
The OpenStack community released OpenStack 2026.2 Hibiscus, its 34th release, which includes enhancements across OpenStack services that strengthen security and trust, expand support for AI and other compute-intensive workloads, and help operators use infrastructure more efficiently at scale. It features:
Nova expands confidential computing support with AMD SEV-SNP and Intel TDX, providing hardware-backed memory encryption, stronger workload isolation and attestation for sensitive workloads.
Designate strengthens DNS security with enhanced cross-tenant isolation, stronger authentication and support for TLSA/DANE, while adding new tooling to help operators prepare for post-quantum cryptography.
Cyborg strengthens accelerator security and tenant isolation with tighter access controls and hardware management protections designed for multi-tenant environments.
Nova and Cyborg expand support for accelerated workloads by enabling OpenStack to manage mediated devices such as vGPUs, giving operators greater flexibility in how accelerator resources are provided and managed.
Manila expands high-performance storage options with support for Weka and Lustre parallel filesystems for data-intensive AI, machine learning and HPC workloads.
Ironic and Nova improve bare-metal resource tracking, helping operators more efficiently manage the specialized and heterogeneous compute increasingly used for AI, HPC and other demanding workloads.
Watcher improves infrastructure utilization by making workload consolidation more responsive to both real-time usage and reserved capacity, while expanding automation options for managing unused resources.
Neutron significantly reduces networking overhead at scale, including a more than 15x reduction in the memory footprint of HA router monitoring.
Glance makes image distribution and management more efficient with parallel image imports, preferred-store downloads and expanded cache management capabilities.
Learn more here.
…
Quantum has appointed now ex-board member James Clancy as its COO, as CEO Hugues Meyrath continues strengthening his executive team. Clancy will guide the Company’s execution of its operational and strategic objectives, with the goal of strengthening organizational performance, supporting revenue growth and customer success, and advancing Quantum’s ongoing transformation initiatives.
He has a decent background to say the least as, while President of Global Storage Sales at Dell, he drove significant global sales growth worldwide. As SVP Global Sales for Dell EMC’s Data Protection Solutions, he enhanced sales and data efficiency for the industry market leader. Clancy also served as SVP at Dell Technologies, where he provided strategic leadership to improve customer relationships, streamline operations, and drive profitable revenue growth. Earlier roles also include Divisional VP of Americas for EMC’s Backup Recovery Systems Division and Specialty Sales Management at EMC Corporation (now Dell EMC, DPS). Clancy has also served as a business advisor to HYCU, Black Kite, Index Engines, and Nexus Advisory Partners.
Quantum has also ppointed Mike Heuer as its CIO to lead Quantum’s global information technology organization, including enterprise applications, data, AI adoption, infrastructure, and cybersecurity.
…
SerialTek has announced the Kodiak PCIe 7.0 Protocol Test and Analysis System, which captures, analyzes and generates PCIe 7.0 traffic at 128 GT/s across all 16 lanes. It is the third generation built on the Kodiak architecture and is scheduled to ship in Q1 2027. Learn more here.
…
Snowflake customer LSEG is increasing its Snowflake spend. LSEG is the London Stock Exchange Group, a UK-based financial markets infrastructure and data company listed on the London Stock Exchange. It and Snowflake announced an expanded five-year enterprise-wide collaboration bringing together LSEG’s trusted financial data and resilient financial market infrastructure with Snowflake's cloud-based data and AI technology. LSEG is increasing its commitment with Snowflake to support planned growth, migration programmes and customer-facing data services across its Markets, Data & Analytics, AI and Risk Intelligence. Snowflake will also become a licensed customer of LSEG’s Risk Intelligence World-Check services, supporting its own customer onboarding and third-party risk management processes.
…
Customer intelligence supplier Stravito launched its Model Context Protocol (MCP) Server to bring each company's market and consumer research into the everyday AI tools its teams use, so decisions get made on evidence the business owns and trusts. Even connected to a company's files, generic AI often reads a research report like a text document, missing the charts, graphs, and visuals where key findings live, or falling back on unverified answers from the web. Stravito’s MCP Server changes that, connecting to AI tools like ChatGPT, Claude, Copilot, and the custom agents enterprises have built, through a single, governed connection. It draws on the full body of a company's research, reading not just text but also visual findings from reports and presentations to videos, audio, and spreadsheets.
…
Veeam announced that it has upgraded Aberdeen Investments’ data resilience strategy. It has achieved 100% protection of its Salesforce files and metadata with zero failures, while improving total cost of ownership (TCO) by 50% using Veeam Data Cloud for Salesforce.
…
VergeIO has announced its expansion into Europe and named Rémi Bargoing VP, EMEA Sales. Rémi will build a local sales and partner team for Europe, the Middle East and Africa from a new office in Milan, Italy. He brings more than 20 years of infrastructure sales leadership in Europe, including 13 years at DataCore Software, where he ran the Italian subsidiary and most recently served as VP Sales.
…
Walrus announced Matterhorn is making Walrus Memory the default backend for agent memory. Project context, conversations, docs, deployment history, contract artifacts, audit reports, and workflow state are written to Walrus instead of living only in a chat session. Matterhorn (matterhorn.so) is an AI coworker for crypto and Web3. You describe what you want in natural language; it helps you use protocols, generate and audit smart contracts, and prepare transactions across 20+ chains. Signing and approval stay in your own wallet—it does not hold funds or sign for you. Walrus Memory is a portable memory layer for AI agents, built on Walrus (a decentralized, verifiable data platform from the team behind Sui / Mysten Labs). It launched in June 2026. LLM calls are stateless, so agents normally forget everything between sessions, and when memory exists it is usually locked inside one vendor. Walrus Memory stores encrypted memories as blobs on Walrus.
…
]]>Blogger Daniel Valdivia, a MinIO architect, compares a CoreWeave LOTA (Local Object Transport Accelerator) caching proxy benchmark with MinIO AIStor object storage performance. CoreWeave ran Warp, an open benchmark test, loading object data on on 20 GPU nodes, each with 8 GPUs, 1 TiB of LOTA cache per node, and a dedicated network adapter with dual 100 Gbps links for storage access, and achieved S3 GET performance of 18.4 GiB/s per node (19.8 GB/s) . MiniO’s AIStor, using Solidigm QLC SSDS, with no caching achieved c33.5 GiB/s per node (36 GB/s) in a separate benchmark.
Valdiva says: “An erasure-coded object store on capacity-optimized QLC drives, running over plain TCP, delivers more throughput per node than a warm cache on local NVMe (about 33.5 GiB/s versus 18.4 GiB/s), and more than 11 times the cluster throughput of CoreWeave's own cold path. A cache should be faster than what's behind it. When a durable store with parity beats the cache, the architecture is telling you something.”
He explains: “A cache exists because the thing behind it is slow. When your object store is fast, you don't need to make complicated trade-offs to keep GPUs fed. …you shouldn't need a cache on every GPU node to make your object store fast in the first place.”
MinIO is not saying you don't need KV caching systems: “An object cache speeds up S3 GETs of datasets and checkpoints. An inference KV cache stores attention state so the engine doesn't recompute a long prompt every time a user comes back. They have different access patterns, block sizes, and success metrics.” They are two different caching situations.
MinIO supports KV caches and has developed its own MemKV, a shared-nothing KV store on raw NVMe, wired into inference engines like vLLM.
In effect, CoreWeave wouldn't need LOTA if it used MinIO's AIStor instead.
]]>MRAM (Magneto-Resistive RAM) is a semiconductor memory technology. It’s non-volatile and byte-addressable memory using spin-polarised currents to change the magnetic orientation, and hence bit value, of a cell. It offers DRAM-class access speed, NAND persistence and virtually unlimited endurance. CXL (Computer eXPress Link) is technology to extend a server’s DRAM by linking external memory across the PCIe bus. The initial and basic CXL v1standard enables server CPUs to access shared memory on local accelerator devices with a cache coherent protocol. The external memory is addressed using memory semantics and is not an external block-addressed storage device
Everspin CEO and president Sanjeev Aggarwal, said: “Everspin has been working with data center customers for 15 years, and we continue to innovate as AI changes the memory needs of systems. CXL gives us a standard way to bring persistent MRAM to a much broader range of compute platforms. We now have a working system so customers can run their own workloads and measure the benefit for their specific applications.”
The demo system created a fast persistent buffer and cache tier with cache-line-level, nanosecond-class access and writes up to 100x faster than NAND-based solid-state storage.
Everspin wants to extend MRAM technology into new memory architectures, including persistent caching, write buffering and metadata applications. Up until now and like Optane, MRAM has not been able to make much progress outside specialist low volume markets where its relatively high price is not a buying barrier.
The demo system was a Supermicro AS-1116CS-TN server with an AMD EPYC 9355 (32-core) CPU plus an AMD Alveo U250 FPGA with a Wolley CXL controller hooked up to Everspin PERSYST 1GB DDR4 MRAM UDIMMs (Unbuffered Dual In-line Memory Modules).
Wolley CEO DR.Bernard Shung said: “Wolley enables memory and storage connectivity through CXL, and this demonstration with Everspin shows how persistent MRAM can be brought into that environment with memory semantics. Together, we are demonstrating a practical way to connect persistent memory through CXL and make it available to today’s compute platforms.”
The AMD Alveo U250 card’s host interface is PCIe Gen3 x16 (8 GT/s), per AMD’s U200/U250 datasheet. It is not a PCIe Gen4 or Gen5 endpoint.
We understand that Everspin used its PERSYST ST-DDR4 family, specifically the 1 Gb device: EMD4E001G. This is a 1 Gb Spin-Transfer Torque (STT-MRAM) part.
The demo system involved the Alveo U250 FPGA card, with the Wolley CXL controller, sitting atop a plastic crate, with 4 Everspin UDIMM Cards mounted on it.
The idea is, as we understand it, that system developers could offer an x86 edge and small data center AI inferencing server with expanded memory capacity using CXL-connected MRAM to hold AI data such as KV cache blocks, with the ability to read them faster than from external SSDs, even where they are connected with RDMA-based KV Cache software, and not lose the data when power is switched off or fails. Data remains available through a power interruption or system restart without relying on battery-backed DRAM or having to first flush data to an SSD. This would surely require CXL running across PCIe gen 5 or even 6 links, far faster than the lumbering (by today’s standards) PCIe gen 3.
A subsequently obtained Q&A interchange and demo block diagram shows 8-lane PCIe gen 3 was the CXL carrier;
Blocks & Files: Can you confirm that the Alveo U250 card and Wolley controller linked to the Supermicro host server using PCIe gen 3?
Everspin: Yes, the block diagram shows the PCIe Gen3 x8 interface.
Blocks & Files: Can you say what level of CXL was used in the demo: CXL 1.0, 2.0, 3.0 or 3.1/2?
Everspin: Yes, the demo is based on Wolley CXL 3.2 Controller IP.
Blocks & Files: Can you say if the Everspin MRAM UDIMMs were mounted on the Alveo U250 card which has 4 on-board DIMM slots?
Everspin: Yes, the demo is based on 4 1GB DDR4 MRAM DIMMs inserted into the FPGA card on-board DIMM slots.
Everspin was started up in 2003 and has raised a total of $80.3 million in funding, with the last round in 2015. It IPO’d in October 2017 and its revenues have see-sawed since then but with an overall lowish growth trajectory.
We subsequently found out that the Wolley Compute Express Link (CXL) Controller IP is compatible with CXL v1.1, 2.0 and 3.2, with max data rates of 32 GT/s for CXL 1.1 and 2.0, and 64 GT/s for CXL 3.2.
]]>Diskover’s technology combines general file metadata inspection and specific application knowledge, in the media, EDA, and life sciences about workflow temporary files and the like, to better identify existing files and their relevance. Its software scans and indexes data, builds a metadata catalog, and uses that to support searches, analysis, actions, and automation. The agreement with NetApp joins existing ones with Dell and Snowflake.
NetApp’s Sandeep Singh, SVP and GM, Platform, said: “NetApp believes in customer choice, and we’re excited to collaborate with Diskover to bring its metadata discovery and cataloging to our media and entertainment customers and other industries with file-intensive workloads. These customers need specialized ways to discover and organize data across complex environments, and Diskover extends the NetApp ecosystem with exactly that. Together, NetApp’s own metadata catalog and our collaboration with Diskover serve the full range of our customers’ metadata needs, building on the intelligent data infrastructure they already trust.”
The Diskover Platform deploys within the customer’s own environment, on-premises or in their virtual private cloud, and connects to ONTAP and StorageGRID using native scanning. Its DataDiscovery function connects file metadata to the productions, chip designs, and research programs it belongs to, and fits into the media, EDA, and life sciences workflows teams already use. It reveals cost by project, user, group, age, and storage tier, supporting chargeback and identifies redundant and obsolete and trivial data (ROT). Diskover says its customers typically reclaim 10 to 30 percent of their storage footprint, exceeding $3 million dollars in larger environments, with return on investment (ROI) inside 30 to 60 days.
The Diskover catalog also gives governance teams a place to search, trace, and act on sensitive or regulated content across every share and tier, helping to enforcing retention and access policies against files that actually exist.
NetApp invested in Diskover in its June 2025 $7.5 million seed round, and has a media customer, DreamWorks, using ONTAP with Diskover indexing.
Scot Miller, Technology Fellow, DreamWorks Animation, said: "NetApp ONTAP systems are the storage foundation of high-demanding environments where reliability and performance are mandatory. The true power unfolds when combined with Diskover’s Data ability to provide a unified, queryable view across NetApp storage, enabling efficient data hygiene, change tracking, and deep governance insights."
Snowflake Ventures also invested in Diskover's seed round.
At the time of the NetApp investment, Gagan Gulati, SVP and GM of Data Services at NetApp, said: “Customers who use Diskover alongside NetApp solutions to build an Intelligent Data Infrastructure will be better equipped to surface and activate their unstructured data at scale, wherever it lives. This collaboration helps accelerate cyber resiliency, AI readiness, and storage efficiency to deliver outcomes that drive business value.”
Regarding the latest resell-type NetApp deal, Diskover CEO Will Hall said: “We’ve spent nearly a decade building the infrastructure enterprises now run on to make unstructured data visible and actionable. Collaborating with NetApp means that work extends directly into environments already running at enterprise scale, through a partner customers already trust, without asking a customer to change how they store their data to get there.”
He added: “NetApp is the foundation enterprises build their data infrastructure on. What we do is extend that foundation’s value into the specialized workflows our customers in media, semiconductor design, and life sciences run every day, across NetApp storage and the environments around it. A NetApp customer gets a governed, searchable view of the data behind their projects, on NetApp and around it. NetApp customers get more from the infrastructure they already trust, not a reason to look elsewhere for visibility.”
The Diskover Platform offering is available now for NetApp ONTAP and StorageGRID environments and will be available to purchase through NetApp beginning in November 2026.
NetApp’s AIDE (AI Data Engine) can discover and characterize data spread across a customer’s NetApp data estate, both on-premises and in the three main public clouds. As fresh data comes in and/or old data is deleted the metadata is updated. The catalog covers NetApp ONTAP and StorageGRID, ONTAP-based first-party cloud services in AWS, Azure and GCP (with OCI coming), and third-party NAS and object storage over NFS, SMB/CIFS, and S3, on premises and in the cloud.
So far this year, NetApp has acquired Data Pelago, JetStream Software, and is buying Peak:AIO as well. Why not buy Diskover as well? We suppose it is not set on acquiring Diskover as, perhaps, its ROT-finding and data placement technology is only applicable in the, for NetApp, relatively small media, EDA, and life sciences markets.
Still, Diskover looks like a worthwhile technology tuck-in possibility.
Diskover was founded by initial CEO and current CTO Chris Park in 2016 and developed a web application, diskover-web to manage files and provide storage analytic functions. It used Elasticsearch for indexing and crawling supported on-premises and public cloud NFS/SMB accessed file systems. The aim was to provide visibility into a heterogeneous data estate, identifying unused files and reducing storage waste.
Diskover has free community edition software and developed supported paid editions – Professional, Enterprise, Media, Life Science – for enterprises. It has also developed a plug-in architecture to make its SW extensible.
Will Hall was appointed CEO in July 2024.
]]>