Massive file transfer supplier MASV has an upcoming MASV Express product that can automate complex media ingest and delivery processes without coding or any complicated setup.
It can move files to the cloud or connected on-prem storage, MAMs (Media Asset Managers), and individual recipients with no coding, no desktop client, drag-and-drop workflows, two-way encryption, secure authentication, and privacy controls. The files can be up to 20 GB or more in size.
Majed Alhajry
Majed Alhajry, CTO and interim CEO of MASV, stated: “File transfer is no longer just a background process; it’s a critical function in the global media supply chain. When files move faster, more reliably, and with less friction, content gets to market sooner and that means more opportunities to monetize across broadcast, digital, and OTT (Over the Top) platforms.”
OTT refers to the delivery of audio, video, and other media content directly to viewers over the internet, bypassing traditional distribution channels like cable or satellite TV. MASV’s file transfer moves files at up to 10 Gbps across an AWS server and content delivery network. It uses TCP-based file delivery, with chunking – splitting files into smaller parts for parallel transfer – and AWS’s private network infrastructure to accelerate data movement.
A customer such as Troveo, which helps creators license unused video footage for AI training, uses MASV for fast, no-code ingestion of over 6,000 TB of video each month. Sarah Barrick, Head of Growth at Troveo, said: “If something like MASV did not exist, it likely would have precluded a lot of creators from monetizing their video assets with Troveo. The ultimate factor was the usability. It was just so easy for users of all kinds, from small creators to larger enterprises – they didn’t even have to download a desktop client. They didn’t need any technical expertise, and could use a private, secure portal to begin uploading right away.”
MASV partners with Qumulo, and has integrations with AWS, Backblaze, Frame.io, Iconik, Mystika Workflows, Wasabi, and other suppliers. Think of MASV as a dedicated large file transfer service, overlapping in that regard with Arcitecta, LucidLink, and PeerGFS, which all offer more functionality alongside basic file moving. The MASV Express offering will be showcased at IBC2025 in The Netherlands, Stand 1.B01, RAI Amsterdam, September 12-15, 2025.
Bootnote
MASV was a passion project within a small Canadian networking company before spinning off in late 2019. That company was LiveQoS where Majed Alhajry was a software engineer between 2014 and 2017. LiveQoS acquired him and his SuperBeam peer-to-peer file sharing app in June 2014. We understand that MASV started life as a LiveQoS VPN project to improve the reliability and performance of Office 365. MASV then evolved to move massive media files, such as rushes, across the internet, and became MASV Rush. It was in a preliminary development phase from late 2017 to late 2019 with Alhajry the technical team lead. Effectively, he was a MASV co-founder.
MASV was officially founded in December 2019 as a pay-as-you-go large file transfer service, with co-founder Gregg Wood as CEO. He is an ex-Corel Senior Director of Product Management who left in January 2017 and became a C-Level advisor at the Positive Venture Group in Ottawa. Early funding for MASV included investment from Peter Lalonde, a Positive Venture Group partner. Wood stepped down from the C-Level Advisor role in September 2018, initially becoming MASV’s VP Product Management and then, when it officially launched, its CEO.
MASV CFO Jeff Stoss is another co-founder, having been in at the start and CFO at LiveQoS from 2016. Stoss was also the co-founder, CFO, and C-Level Advisor at the Positive Venture Group. This was sold to Consero Global in July 2021. LiveQoS, with its residual SD-WAN IP, was sold to IP licensing business Adaptiv Networks in 2019.
We asked MASV some quesions about its recent history;
Blocks & Files: Is Majed Alhahry a co-founder of MASV? Same for CFO Jeff Stoss?
MASV: Yes, Majed and Jeff are co-founders along with Greg Wood who remains active in MASV via his vice-chairman role.
Blocks & Files: Why did CEO Greg Wood become vice-chairman at MASV?
MASV: Co-founder Greg has made the transition from CEO to Vice Chairman after his successful 5 year stint in leading MASV from its founding. This is a strategic succession to cofounding CTO Majed Alhajry, reflecting the company’s evolution into a deeply technical platform for secure enterprise data exchange. Greg did a fantastic job building the business and is now working on new initiatives to take MASV into new markets. Greg remains actively engaged supporting Majed and Team MASV.
Blocks & Files: Why is Majed Alhajry an interim CEO? I understand he’s been in that post since January when Greg became vice-chairman.
MASV: Majed brings a distinguished track record of technical and business leadership as the founder of SuperBeam and co-founder of MASV. His recent appointment as interim CEO, in addition to CTO, reflects a natural progression in his journey with MASV, leveraging his deep expertise to drive the company’s strategic vision. In this expanded role, Majed is spearheading MASV’s push into the enterprise market, bolstered by a growing team, strategic partnerships, and purpose-built solutions like MASV Express. These initiatives—combining cutting-edge features with enterprise-grade capabilities—position MASV to solidify its foothold and accelerate growth in the corporate sector.
Blocks & Files: What is cut-through delivery technology?
MASV: Cut-through delivery is no longer a term MASV will use going forward to refer to Express. Instead, they have explained how Express works:
MASV Express moves files from origin to the destination without first needing the complete file in MASV’s secure container. This way, Express dramatically accelerates turnaround time, on top of MASV’s already super-fast file transfer speed, preserving MASV’s operational agility. MASV Express keeps content flowing smoothly through any network environment, ensuring the fastest, most reliable, and most secure delivery experience available. Express builds on MASV’s current capability to overcome any point-to-point delivery issues—like propagating outages and network congestion—by preserving reliability of the file transfer while decreasing turnaround times.
Adeptia Connect is a business data exchange software moving data between applications, partners, and customers, supporting data integration, application integration, and B2B/EDI processes. It can be single or multi-tenant and deployed on-premises or in the cloud or both. It has AI-driven data mapping, pre-built connectors, and reusable templates, allowing non-technical users to create and manage data workflows. Adeptia has just announced self-managed and hybrid deployment options for Connect. “Hosted on Microsoft Azure and AWS and managed by Adeptia’s specialized Cloud Infrastructure Engineering team, Adeptia’s self-managed option combines the simplicity of SaaS with the enterprise security standards that customers demand. Adeptia handles complex operational tasks – such as deployment optimization, monitoring, backups, disaster recovery, and version upgrades – while maintaining rigorous security compliance, including SOC 2 certification.” More info here.
…
Data source connector supplier Airbyte has added new features to its Enterprise product: the ability to facilitate data sovereignty and compliance across different global regions; synchronize data with its metadata to assist AI models’ ability to improve reasoning and accuracy; and directly load data volumes to BigQuery and Snowflake, which we’re told increases data sync speeds and reduces compute costs. Direct loading can reduce compute costs by 50 to 70 percent, and increases speed by up to 33 percent depending on the use case.
Airbyte now enables synchronized unstructured files and structured records to be transferred in the same data pipeline, which is the most efficient way to preserve metadata and data relationships – enabling richer data context and significantly improving the performance of AI Large Language Models (LLMs) using that data. There are more details in the blog post and more about Airbyte Enterprise here.
…
Assured Data Protection (ADP) is partnering with technology services distributor Avant in North America to offer Rubrik and Nutanix-based managed backup, disaster recovery (DR), and cyber resiliency services. It will mean that Avant can offer Rubrik’s backup and DR technology to customers for the first time with Nutanix technology also being available, depending on needs.
…
Ataccama ONE v16.2 introduces AI-driven data lineage features designed to help business users audit, trust, and act on their data, without writing a single line of SQL. It says: “Business users can now trace a data point’s origin and understand how it was profiled or flagged without relying on IT. Ataccama shows how data flows through systems and provides plain-language descriptions of the steps behind every number. For example, in a financial services setting, a data steward can immediately see how a risk score was derived or how a flagged transaction passed through a series of enrichment and quality checks. That kind of visibility shortens reviews, streamlines audits, and gives business teams the confidence to act on the data in front of them.”
…
Broadcom’s Tomahawk Ultra is an Ethernet switch for HPC and AI workloads with 250 ns switch latency at full 51.2 Tbps throughput. It delivers line-rate switching performance even at minimum packet sizes of 64 bytes, supporting up to 77 billion packets per second. It reduces header overhead from 44 bytes down to as low as 6 bytes, while maintaining full Ethernet compliance – boosting network efficiency and enabling flexible, application-specific optimizations. The switch incorporates lossless fabric technology that eliminates packet drops during high-volume data transfer, and implements Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC) to eliminate packet loss and ensure reliability. It’s designed with topology-aware routing to support advanced HPC topologies including Dragonfly, Mesh, and Torus.
…
Cadence announced the tapeout of the industry’s first LPDDR6/5X memory IP system solution optimized to operate at 14.4 Gbps, up to 50 percent faster than the previous generation of LPDDR DRAM. The Cadence IP for the JEDEC LPDDR6/5X standard consists of an advanced PHY architecture and a high-performance controller designed to maximize power, performance and area (PPA) while supporting both LPDDR6 and LPDDR5X DRAM protocols for optimal flexibility. It says its new LPDDR6/5X memory IP system offering enables scaling up AI infrastructure to accommodate the memory bandwidth and capacity demands of next-generation AI LLMs, agentic AI and other compute-heavy workloads for various verticals. Multiple engagements are currently underway with leading AI, high-performance computing (HPC) and datacenter customers.
…
Research from data streamer Confluent says more than two-thirds (67 percent) of UK IT leaders see disconnected, siloed data as preventing them from fully adopting AI and machine learning (ML) – making fragmented data systems the largest obstacle to AI success. There was:
Ambiguity surrounding data lineage, timeliness, and quality assurance – 63 percent
Insufficient skills and expertise in managing AI projects and workflows – 61 percent
Limited ability to seamlessly integrate new data sources – 57 percent
Insufficient infrastructure for real-time data processing – 50 percent
Confluent announced that Confluent Cloud is now available in the new AI Agents and Tools category on AWS Marketplace. It says this enables AWS customers to easily discover, buy, and deploy AI agent solutions, including its fully managed data streaming platform, Confluent Cloud, using their AWS accounts, for accelerating AI agent and agentic workflow development. Confluent Cloud streams, connects, processes, and governs data as it flows throughout the business in real time, enabling customers to build and deploy intelligent, adaptable, and scalable AI applications faster. By replacing fragmented, batch-based systems with a continuously updating stream of business context, Confluent creates a shared, real-time data backbone that AI agents can subscribe to, interact with, and reason over. Confluent Cloud empowers customers to confidently scale their businesses for the future with AI at the forefront.
…
China’s CXMT, the parent of DRAM fabber ChangXin Memory, is preparing for an IPO in China. This is another sign, together with YMTC’s NAND plans below, of China’s growing self-sufficiency in memory technology and production. China could emerge as a DRAM and NAND supplier on the world stage.
…
Data management supplier Denodo’s cloud-based Agora offering is now available in the AWS Marketplace’s new AI Agents and Tools category. Denodo is offering both a SaaS version (Agora) and an Amazon Machine Image (Denodo Enterprise Plus), giving AWS customers a flexible path to build Agentic AI ecosystems using open standards. With support for the Model Context Protocol (MCP) built into the latest Denodo AI SDK, organizations can now unify, govern, and deliver real-time data to GenAI apps without having to replicate or re-engineer their entire data architecture.
…
Denodo has launched DeepQuery, a new capability that is bringing reasoning-based GenAI to enterprise systems. While most GenAI tools stop at retrieving facts, DeepQuery connects to live, governed enterprise data and delivers answers that explain the why behind the what. These answers are fully reasoned, cited, and ready for decision-making. Rather than pulling data into a silo, DeepQuery applies structured reasoning directly on real-time systems to support sales teams, finance, HR, and more. It’s the first enterprise-grade GenAI to offer this depth of synthesis with full explainability and traceability marking an evolution toward cognitive GenAI, where models don’t just answer – they investigate. Denodo is offering early access to select partners through its AI Accelerator Program. Denodo also announced the availability of Model Context Protocol (MCP) support as part of the Denodo AI SDK.
…
China’s Exascend offers products like PCIe NVMe SSDs, SATA SSDs, CFast, and CFexpress cards, tailored for demanding applications in aerospace, AI, IoT, and video surveillance. It has two new PE4 SSDs, 7 mm U.2 models: 23.04 TB and 30.72 TB. They “enable system integrators to maximize capacity in edge servers, rugged industrial computers, on-premises AI nodes, and space-constrained datacenters where every watt and every millimeter count.” The PE4 30.72 TB doubles the density of competing 7 mm U.2 SSDs and enables up to 737 TB of flash in a single 1U server, cutting rack space requirements by 50 percent compared to 15 mm drives and delivers a two-fold jump in TB-per-U efficiency.
The PE4 series have a PCIe Gen 4 x 4 interface and use 3D TLC NAND. They are rated for up to 1 drive write per day (DWPD) and offer a 2-million-hour mean time between failures (MTBF). Engineering samples of the PE4 Series U.2 SSD in 23.04 TB and 30.72 TB models are shipping now to select customers, with volume production scheduled for Q3 2025. The PE4 7 mm family is already available in 7.68 TB and 15.36 TB capacities.
…
Frontgrade Technologies, which provides high-reliability microelectronics for space and national security, has unveiled its embedded MultiMediaCard (eMMC) – a space-grade NAND memory product offering 32 GB density, the highest available in today’s market. Designed to deliver maximum capacity and resilience in space environments, this radiation-tolerant, non-volatile flash data storage device is optimized for low Earth orbit (LEO) missions and beyond. It has an integrated flash controller with built-in FTL for wear leveling, bad block handling, and ECC, and has radiation-tolerant construction suitable for many mission orbit requirements.There is a standard eMMC 5.1 interface for streamlined system integration. Frontgrade’s eMMC is currently available for sampling. Flight units are expected to be available in Q4 2025. Download a datasheet from this webpage.
…
IBM says integrating Storage Ceph with its Storage Deep Archive “presents a compelling solution for organizations looking to optimize their data archival strategies. The combination delivers substantial cost efficiency, operational simplicity, air gap resiliency, and policy-driven flexibility.” Ceph now supports policy-based transitions of data to external S3-compatible storage tiers, including those that present Glacier Retrieval classes. IBM’s Storage Deep Archive has “ultra-low-cost archival capabilities” and “IBM Storage Deep Archive uniquely complements Ceph’s capabilities by offering an S3 Glacier-compatible interface without the traditional overhead associated with tape systems,” although it is a tape library system. An IBM blog has more information.
Note: IBM Storage Deep Archive is an on-premises archival data storage system integrated within an IBM Diamondback server and a 5U top rack, utilizing tape storage with an S3 Glacier command interface. It offers up to 27 petabytes of capacity in a 19-inch rack.
…
William Blair analyst Jason Ader says of Nutanix: “With respect to the competitive fallout from Broadcom’s acquisition of VMware, Nutanix sees this as a multi-year displacement opportunity, less characterized by a specific tipping point and more by a continuum that will build on itself over time (putting more pieces in place right now to execute on the land grab, including channel enablement, OEM and external storage partnerships, and incremental sales reps). While Nutanix has already seen some benefit from VMware displacements (as evidenced by the recent uptick in new logos and the strong and growing seven-figure pipeline), inertia with VMware customers is real (especially with larger, complex deployments) and software/hardware refresh cycles often need to align for customers to shift off VMware.”
…
VergeIO is offering replacement storage with the launch of The AFA Replacement Kit. VergeIO says it’s “an offering designed to replace traditional all-flash arrays with a simpler, more cost-effective infrastructure solution. VergeIO customers have reported reducing storage costs by a factor of ten, in addition to the added savings from eliminating expensive VMware licensing and support agreements. This VergeIO kit brings together three Solidigm 4 TB enterprise SSDs and a VergeOS server license combined into one streamlined platform. Along with your servers, it’s a complete, ready-to-run infrastructure solution. All IT needs to do is insert the included flash drives into empty drive bays in existing servers, and they’re ready to deploy VergeOS.” VergeIO marketing head George Crump said: “It is 12 TB per node, typical customer has at least four nodes. If a customer needs more, we have flexibility.”
We don’t see customers choosing this as a general external AFA replacement, e.g. rejecting NeApp AFF A150, which is both a scale up (2RU x 24 slots) and scale-out external storage system (to 12 x A pairs) whereas the VergeIO AFF Replacement is a scale-out system with three-drive (slot) nodes. We see it specifically as a VMware vSAN replacement.
VergeIO was been asked about this point and a spokesperson said: “The VergeIO AFA Replacement Kit is not limited to three drives or three nodes—those were illustrative starting points for the press release. The solution can scale to any number of drives per server and any number of servers. It’s designed to give IT teams flexibility to size configurations based on workload requirements, not fixed appliance specs.”
Also the product is both a vSAN and general external AFA replacement, with VergeIO saying: “The AFA Replacement Kit serves both purposes. It functions as a general AFA replacement, suitable for customers considering alternatives to traditional SANs and filers from vendors like Dell, NetApp, Pure, HPE, and others.[And] It is a strong alternative for VMware customers who are moving off vSAN and want server-side flash with integrated compute and storage.”
VergeIO wants us to know “that research from ESG indicates that over 53% of SANs are used exclusively to support virtualized workloads, and over 90% support at least some virtualized workloads. VergeIO is designed to meet that need directly. While VergeOS is not designed to run non-virtualized workloads, many customers find that they can migrate bare-metal applications they previously believed too resource-intensive for virtualization. This is common in VergeOS environments thanks to its efficiency, where storage, compute, AI, and networking are integrated into a single software codebase.”
…
DigiTimes reports China’s NAND fabbing YMTC aims to pilot production of NAND chips using China-only tech this year. It was estimated to hold 8 percent of the global NAND market (a monthly capacity of 130,000 wafer starts per month) by the end of 2024, and wants to reach a 15 percent share by the end of 2026. This is based on ramping up WSPM to 150,000 and its latest gen 5 NAND memory with 294 layers built from bonded 150L and 144L components (Xtacking 4).
YMTC is on the US tech export sanctions list and cannot buy NAND fab tools using the latest technology from suppliers such as ASML. Tom’s Hardware has YMTC NAND die info:
X4-9070 – 1 TB TLC 3600 MT/s – in production
X4-6080 – ? TB QLC – H2 2025
X5-9080 – 2 TB TLC with 4800 MT/s – 2026
X5-6080 – ? TB QLC with 4800 MT/s – 2026
YMTC’s gen 6 die will likely feature 300+ layers built from 3 component stacks (Xtacking 5).
…
Zadara software is being used by East Africa services provider TouchNet at an iXAfrica co-location datacenter in Nairobi, Kenya, to provide a sovereign cloud, offering computer, networking, storage, and GPUaaS. The Zadara AI Cloud enables businesses, managed service providers (MSPs), and cloud service providers (CSPs) to optimise operations without the need to rely on offshore services. Zadara provides its own take on the public cloud with fully managed compute (EC2 compatible), storage (block, file, object), and network systems hosted in a global point-of-presence (POP) network, called edge clouds. It can also be run on AWS, Azure, GCP, and Oracle public clouds, or in a customer’s own datacenters – as in this iXAfrica case. This deployment enables sovereign cloud capabilities hosted locally in Kenya, and a public cloud-like experience (compute, storage, networking) using Zadara’s EC2- and S3-compatible services. There is full regulatory and data residency compliance for sectors requiring localized data control (e.g. government, finance, healthcare).
Analysis: There are many places within IT infrastructure that organizations can use to get a single proprietary data source for their large language models and AI agents doing RAG (retrieval-augmented generation) – backup stores, data lakes, data management catalogs, and storage arrays. Let’s sketch the background then take a look at how they compare to each other.
LLMs and agents can scan all kinds of unstructured information in text, numbers, audio, image, and video formats in response to natural language inputs and generate natural language outputs. However, they are initially trained on general data sources and will make better responses to an organization’s users if they have access to the organization’s own data. But that data is spread across many disparate systems and formats.
For example, datacenters (mainframes and x86 servers), edge sites, public cloud instances, SaaS app stores, databases, data warehouses, data lakes, backups on disk and tape on-premises and in the public cloud, and archives. The data can be accessed via structured (block) or unstructured (file and object) protocols with many subsidiary formats: Word, Excel, PDF, MPEG, etc.
There are four theoretical approaches to getting all these different data sources available for RAG. First, you could go to individual specific sources each time, which would require you to know what they are, where they are and what they contain. This approach operates at the level of individual storage arrays or cloud storage instances.
Secondly, you could go to existing agglomerated data sources, meaning databases, data warehouses, and data lakes. Third, you could go to the backup stores and, fourthly, use a data management application that knows about your data.
These second, third, and fourth options already have metadata catalogs describing their content types and locations which makes life a lot easier for the RAG data source hunter. All else being equal, accessing one location to find, filter, select, extract, and move RAG data is better than going to myriad places. It makes AI data pipeline construction much simpler.
A couple of other observations. The more of your proprietary data the single place contains the better as you have to make fewer exception arrangements. Secondly, the LLMs and agents need vectorized unstructured data for their semantic searches, and this needs producing and storing. Any central RAG data source facility needs to support vectorization and vector storage.
That’s the background. Let’s take a gander at the four main RAG data sources.
Storage arrays
Virtually every storage array vendor we know, both hardware and software, software-defined, file and object, is building in some kind of Gen AI support. Cloudian, DDN, Dell, HPE, NetApp, Scality, StorONE, VAST Data, WEKA, and others are all piling in.
Arrays (or SW-defined storage) which have a fabric connecting on-premises arrays, and also public cloud instances with a global namespace will have an advantage; their reach is obviously greater than non-fabric arrays. This is true for cloud file services suppliers, such as Box, CTERA, Egnyte, Nasuni, and Panzura as well.
However, such vendors can only supply data for RAG stored on their systems and nowhere else, unless they have connectors giving them wider data access. An agent granted access to a Dell, NetApp, DDN, HPE, whatever array won’t be able to see data on another supplier’s array in that organization though the NetApp, DDN, HPE, whatever array RAG lens.
Database/warehouse/lake/lakehouses
A database, data warehouse, data lake and lakehouse store more information respectively and, generally, in more formats as we head from databases to data warehouses, then data lakes and lakehouses. At the database end of this spectrum, specialized vector databases are appearing, from suppliers such as Pinecone and Zilliz (Milvus). They say they offer the best of breed vector storage, filtering, extract, and support for AI pipelines.
Other databases aim to be multi-modal, like SingleStore. Its SingleStoreDB is a performant, distributed, relational, SQL database with operational, analytical, and vector data support, integration with Apache Iceberg, Snowflake, BigQuery, Databricks, and Redshift, and a Flow ingest feature.
Data warehouses are rapidly becoming AI data feed warehouses, witness Snowflake and its AI Data Cloud. Data lakes and lakehouses are also very much GenAI-aware and rapidly developing features to support it. As an example consider Databricks which recently added a Lakebase Postgres database layer to its lakehouse, enabling AI apps and agents to run analytics on operational data within the Databricks environment. It also introduced Agent Bricks, a tool for automated AI agent development.
These are all excellent products, and connectors exist to bring them data from multiple different sources or allow them to access external table software data.
Their operational and admin environment becomes more and more complicated as they extend their reach to external tables or use connectors to bring in data from distant sources
Backup stores
A backup vault stores data copied from multiple places within your organization. It can function as a single source for LLMs and agents needing access to your proprietary information. It’s not real-time information but it can be very close to that and it, obviously, has troves of historical information.
The backup vault storage media is an important consideration. Tape is obviously a no-no due to lengthy data access times. Disk is better while an all-flash vault is probably the fastest.
Cohesity, Commvault, Rubrik, and Veeam are well aware of this strength they possess, as a vast potential RAG data store, and building out features to capitalize on it. Cohesity has its Gaia initiative. Rubrik recently acquired Predibase for agentic AI development functionality.
An obvious caveat is that the backup vendors can only serve data to LLMs and agents that they have backed up. Anything else is invisible to them. This could encourage their customers to standardize on one backup supplier across their organization.
Data Managers
Data managers such as Arcitecta, Datadobi, Data Dynamics, Hammerspace, and Komprise almost never act as data sources. We say almost never because Hammerspace is edging towards that functionality with the Open Flash Platform initiative. The data managers manage data stored on multiple storage arrays and in public cloud storage instances. This partially reflects a hierarchical life cycle management background common to several of them.
The data managers don’t store data, but they do index and catalog it; they know what data their customers have, and they often have data moving capability, coming from a migration background in some cases. That means they can construct interface processes to AI data pipelines and feed data up them. Komprise is highly active here. Datadobi with its StorageMap, policy-driven workflows and data moving engine has a great base on which to build AI model-specific functionality.
Arcitecta’s Mediaflux product can be used to curate high-quality datasets from vast volumes of unstructured data, making it easier to feed data into large language models (LLMs) and other AI systems. We expect more functionality specific to AI to emerge here.
To expand their reach and give them more relevance here the data managers could usefully partner with the database/warehouse, lake and lakehouse suppliers. Hammerspace and Snowflake is an example of this.
Other data silos
SaaS apps such as Salesforce are another data source for RAG. They cannot see outside their own environment and, absent explicit connectors, other data sources, with one exception, cannot see into theirs. Their own environments – Microsoft’s, for example – can of course be large.
The exception is backup. If a SaaS app has its customer data backed up, by a Cohesity, Commvault, Druva, HYCI, etc., then that data is inside the backup vendor’s purview.
Summing up
There is no silver bullet here, no instant fix, no one all-embracing Gen AI RAG data source. You can only search out the best available Gen AI data source or sources for your organization considering your on-prem and public cloud IT data estate. All the suppliers mentioned are developing features, building partnerships and even acquiring companies that own relevant technology. They all have their merits and will have more next month, next year, and so on.
Wherever you are starting from you can perhaps place candidate suppliers into one of the four categories: storage, database/warehouse/lake/lakehouse, backup store and data management. That will give you a quick overall view of their scope. You can then, hopefully, decide which route or routes you might take to arrive at the most efficient way to give the LLMs and agents you use access to the proprietary data they need
The Competitive Corner provides the best and most detailed analysis of vendor ratings in the 2025 Gartner Magic Quadrant (MQ) for Backup and Data Protection Platforms that we have encountered.
We covered this MQ in a storage ticker on June 30, noting Druva got promoted from Visionary to Leader, Huawei entered as a Challenger, the separate Cohesity and Veritas entries have been combined, and Microsoft exits as a Niche Player. Otherwise it was pretty stable with small vendor position changes. Competitive Corner analyst Matt Tyrer applied his microscope to the MQ, plotting the changes in detail, looking at why they took place, and comparing the positioned vendors to a depth we have not seen before.
Matt Tyrer
Tyrer is an ex-head of competitive intelligence at Druva and Commvault.
He notes: “We did see some significant changes in the Leaders quadrant. First off, Druva makes their debut as Leader this year – making them the only 100 percent SaaS platform in the Leaders quadrant and first new entrant to that space since 2020. Veritas is no longer separately listed now that their acquisition by Cohesity is complete, and Dell Technologies continues their inexorable slide out of the Leaders space – something I’m expecting to see in the 2026 MQ.”
Tyrer comments: “Technically, Veeam barely holds onto their position as ‘Highest in Ability to Execute’ (by a sliver) and Rubrik stays just ahead of Cohesity to be a repeat ‘Furthest in Vision.’ We’ll explore why in the next section when we delve into each Leader in more detail.”
“Outside of the Leaders there were a few things of note. As previously mentioned, Microsoft dropped out of the Magic Quadrant this year after a brief appearance in last year’s report. Huawei is the only new vendor this year, serving as the only representative in the Challenger’s space. HYCU lost some ground in terms of Vision, while Arcserve gained position on that axis. Lastly, IBMslightly moved closer to that Leader space.”
Tyrer looks at the leader vendors in some detail and issues verdicts:
Cohesity: With the addition of NetBackup, Cohesity’sworkload support across the board is unmatched. Only Commvaultcan compare to the sheer diversity of workloads and data sources that Cohesitycan protect. While there is indeed some expected chaos associated with the Cohesity+ Veritaspost-merger initiatives, those won’t be a distraction for long and Cohesityhas the very real potential to be THE platform to beat in the enterprise with their combined technology stack – if executed properly.
Commvault has expanded their portfolio faster than ever over the past 12 months, which has introduced the previously mentioned challenges around complexity and usability, but those are issues Commvaulthas successfully addressed in the past and can likely correct moving forward. Despite these concerns, they remain one of the most feature-rich technology stacks in the industry and can address many use cases and data sources others cannot.
For customers that are invested in the broader Dell technology portfolio, the Delldata protection suite (DPS) should be adequate for their needs. The solutions are proven, but do lag behind the market leaders in terms of innovation and features. Customers with mixed heterogeneous vendor environments, particularly those with investments in cloud, will need to properly evaluate which of the many Dellsolutions they will need to deploy – Dell PowerProtect Backup Service (OEM’d from Druva) is by far the best option for those types of Dellcustomers needing on-prem, hybrid, cloud, and SaaS workload coverage.
Druva’s SaaS solution delivers an easy to deploy, simple to manage solution that truly hits the “secure by design” cyber security goals by not having any infrastructure for bad actors to directly attack – something customers would otherwise have to design/architect themselves if using competing solutions in the MQ. This is a great fit for security-minded customers and those investing into hybrid and cloud.
Rubrik’s portfolio continued to grow, and with a pivot back to more of a data protection focus they appear to be finding balance between expanding their core backup capabilities and enhancing their cyber security products. Rubrik doesn’t yet have the broad workload coverage available from Commvault or Cohesity, but it is quickly catching up. Rubrik remains simple to deploy and operate, but that simplicity also translates into Rubrik’s platform being more limited in terms of more advanced features and functionality many enterprise customers need such as in depth reporting and robust cross-platform disaster recovery.
Veeam’s well established market presence and customer base are quick to adopt new Veeam features and products, but Veeam very often shares their footprint with other backup solutions – seeing multiple backup products on the customer floor to address different needs. The Veeam roadmap is ambitious, striving to deliver on 34 new products and enhancements, including significant expansion of their VDC offering and a new release of their core platform. So, look for Veeam to continue to impact the data protection market with a host of new capabilities over the next year.
He also has a look at Arcserve, Huawei, HYCU, and IBM, saying:
Arcserve’s simple, turnkey UDP appliance offering is well proven in the market and a great fit for midmarket customers or enterprise customers with multiple sites they need to easily manage for protection. Their pricing is flexible and with these latest investments into their product, the Arcserve solution supported by these new resilience features is one to watch.
Huawei is an emerging player in the enterprise market, and while their support for multi-cloud and availability is mainly limited to their own cloud they are slowly expanding. For now, it’s likely that customers that are invested in the Huawei cloud ecosystem and those with many sites in APAC will be better suited for their solution, but as workloads and geographic availability/support grow so too will their suitability for a broader market.
HYCU’s focus on Nutanix and SaaS applications gives them a solid advantage for customers heavily invested in either. There is no other vendor close to providing the SaaS coverage available from HYCU R-Cloud. Customers with larger multi-cloud and on-prem environments may need to look at other solutions outside of HYCU to complete their backup coverage, but the differentiated approach HYCU is taking to the backup market continues to see some interesting innovations. If they continue to expand their cyber resilience features and workload support they will be moving in the right direction again on the MQ.
IBM’s technology is proven and steadily catching up with today’s business needs for data protection. Their innovations in cyber security integrations and broad channel support make them a reliable partner for enterprises. With some cloud expansion and continued development of their first party IBM Storage Defender solution, we could see them eventually punch into the Leaders quadrant.
Tyrer’s analysis is available as a blog at his Competitive Corner website, with a 14-minute read time. We’d recommend it as the single best Gartner enterprise backup MQ analysis we have ever seen.
Storage array customers, like those using Rubrik and Nutanix, increasingly require as-a-service backup and disaster recovery. ADP is looking to expand into this area, and is also exploring how it might use AI with the customer data it collects.
Assured Data Protection (ADP) is a global managed services provider, offering Backup-as-a-Service and DR-as-a-Service. It is Rubrik’s largest MSP and, as of this year, also supports Nutanix. It has main offices in both the UK and US, and operates 24/7 in more than 40 countries where it has deployments, with datacenter infrastructure in six worldwide locations. The UK operation is helmed by co-founder and CEO Simon Chappell and his US counterpart is co-founder Stacy Hayes in Washington, DC.
Simon Chappell
We met Simon Chappell in London and discussed how ADP is doing and where it might be going in the future.
ADP launched operations in the Middle East in October last year via a strategic partnership with local value-added distributor Mindware. Mindware will establish local datacenters to help clients manage data sovereignty issues and minimize latency in data transfer. In February, ADP partnered with Wavenet, the UK’s largest independent IT MSP, to provide enterprise-grade backup and disaster recovery solutions to UK customers.
ADP has set up an Innovation Team aimed at expanding the company’s DR, backup, and cyber resiliency services with the addition of new technologies that complement current data protection services.
Rubrik recently acquired Predibase to help with AI agent adoption. The agents would be helped to generate responses by accessing Rubrik-stored data, a Rubrik data lake. You can use AI to clean the data but you can also use AI to look into it and analyze it.
What would Chappell think about providing AI support services for Rubrik customers who are using Rubrik as a data source for their AI? Chappell said: “There you go. That’s the $64 million question. The thing we are really good at is disaster recovery and cyber resilience. If we start to think, oh we’ve got all this data, are we as Assured going to do some service analytics around that? That is a big leap because we’ve gone from being pre-eminent at cyber resilience and disaster recovery to something else.”
Our understanding is that ADP is the custodian of the customer data it has backed up using Rubrik, and it could provide the data to large language models. In fact, it could go so far as to build its own LLM. Such an LLM could be used to clean recovered data in a clean room environment. We’re familiar with Opswat and Predatar malware presence scanning capabilities and wonder if ADP might be thinking along these lines.
This is all speculation on our part, based on the conversation with ADP and how it could potentially evolve. We think that there would need to be some kind of metadata filtering as the dataset size could be enormous, in the 500 TB+ area, leading to prolonged full scan time. You would use metadata to reduce the scan set size.
Another potential direction for ADP’s evolution could be to look for adjacent services possibilities, such as a strategic storage array partner. It could provide BaaS and DRaaS capabilities using the Rubrik and Nutanix-focused infrastructure it has built up in North and South America, the Middle East and Europe. Our understanding is that such a vendor would not be a legacy incumbent but rather a storage array supplier similar in its field to Rubrik and Nutanix in theirs. That would mean a relatively new-ish supplier with a strong and growing upper-mid-market-to-lower-enterprise customer base and a long runway ahead of it; an incumbent-to-be so to speak.
We would not be surprised to find out that, in 6 to 12 months time, ADP has a third strategic partner alongside Rubrik and Nutanix, and that it would be a storage supplier. Similarly, we wouldn’t be surprised to discover that ADP is offering malware cleansing services for recovered data.
The SNIA’s DNA Storage Alliance has published a 52-page technology review looking at the data encode/decode tech, commercial readiness metrics, and the challenges ahead.
DNA data storage relies on the encoding of digital information using sequences of the four nucleotides in strands of DNA. The four nucleotides are adenine (A), guanine (G), cytosine (C), and thymine (T), and they are found in the double helix formation of the DNA biopolymer molecule, located in the cells of all living organisms. Synthetic DNA can store data in a form orders of magnitude smaller than other storage media and can endure for centuries. It relies on chemical reactions at the molecular level and these are slower than electrical operations in semiconductors. Therein lies its core challenge.
DNA data storage has been demonstrated in research projects, but the data write/read speeds, as well as equipment size, complexity, and cost, are all far from any successful commercial product. This technical document reviews the field, looking at the DNA codec, synthesis, storage and retrieval, sequencing, and commercialization challenges.
DNA Storage Alliance tech review paper diagram showing an example of a DNA Channel: In step 1, the source bitstream is randomly scrambled, mitigating problematic sequences, packetized into large blocks which are then coded with ECC (outer code), and encoded from bits-to-bases (inner code), which divides the large blocks into small DNA sequences that are compatible with the properties of the Physical Layer chemistry. Also, object tags (primers) may be added that can be used to retrieve all DNA segments in a pool associated with a particular digital object. Next, the now fully “line coded” DNA sequences are passed to the DNA Physical Layer for writing (synthesis), storing, retrieval, and reading (sequencing). Lastly, the recovered DNA sequences are passed back to the codec where they are converted back to bits and decoded, reversing all the transformations, error correction, packetization etc. done on the encoding side.
A look at the challenges reveals five: data throughput, total ownership costs, media endurance and data retention metrics, bio-security and data security, and standardization. On the throughput topic, the paper reveals: ”The requirements for moving data into and out of traditional storage far exceed the current capabilities of writing and reading DNA in biotechnology use cases.”
It says: “The most fundamental challenge for DDS (DNA data storage) systems is to increase the throughput of DNA write and read operations. The underlying write and read operations for DNA are relatively slow chemical reactions (high latency), so the emphasis for increasing throughput involves enabling parallelism.”
It recommends: “DDS technology providers must increase the throughput of the underlying write and read operations, as well as reducing the time required to move molecules between operations, all while maintaining a competitive TCO for the use case at hand.”
The paper ends on an optimistic note: “While DNA data storage is still quite nascent and there remain significant challenges to commercialization, the foundations of writing, storing, retrieving, and reading data using DNA have been shown to work on scalable technology platforms. Moreover, the ongoing investment in DNA technology, driven by biological and scientific applications, will continue to drive innovations that enhance DNA data storage capabilities.”
DNA data storage will augment, not replace, existing archival storage technologies, “resolving the ‘save/discard’ dilemma with a viable TCO for zettabyte scale and data preservation.”
Use cases, it says, will emerge over the next three to five years for DNA archival data storage.
The paper has lots of terminology that will be unfamiliar to people working with electricity-based digital storage, such as homopolymer, oligonucleotide, ligation, and polymerases, but that’s because it’s molecular organic chemistry. The document is freely downloadable and an excellent introduction to DNA data storage.
Bootnote
The DNA Storage Alliance is an SNIA community, with around 35 members and a six-member board:
Esther Singer, Director, DNA Data Storage. Twist Bioscience
Stephane Lemaire, Co-founder and Senior Innovation Officer, Biomemory
David Landsman, Director Industry Standards, Western Digital
Marthe Volette, Director of Technology, Imagene (an AI biotech company in Israel)
Julien Muzar, Technologist, Life Science, Entegris (8,000 employee supplier of advanced materials and process solutions for the semiconductor and other high-tech industries)
Twist Bioscience is a member and has board representation and recently changed its stance towards DNA data storage. It spun off its DNA business as Atlas Data Storage, a commercializing startup led by Varun Mehta, co-founder and CEO of HPE-acquired Nimble Storage. Twist retains an ownership stake and Atlas raised a $155 million seed funding round in May. We expect Atlas will take over the Bioscience membership and possibly its board position.
Esther Singer is still a Twist employee, being Director of Product and Market Development. In our opinion, the three most important DNA storage technology companies are Biomemory, Catalog and Atlas Data Storage.
DDN has released performance benchmarks showing it can can speed up AI processing time by 27x because of the way it handles intermediate KV caching.
An AI LLM or agent, when being trained on GPUs or doing inference work on GPUs and possibly CPUs, stores existing and freshly computed vectors as key-value items in a memory cache, the KV cache. This can have two memory tiers in a GPU server; the GPUs’ HBM and the CPUs’ DRAM. If more data enters the KVCache, existing data is evicted. If needed later, it has to be recomputed or, if moved out to external storage, such as locally attached SSDs or network-attached storage, retrieved, which can be faster than recomputing the vector. Avoiding KV cache eviction and recomputation of vectors is becoming table stakes for AI training storage vendors, with DDN, Hammerspace, VAST, and WEKA as examples.
Sven Oehme
Sven Oehme, CTO at DDN, states: “Every time your AI system recomputes context instead of caching it, you’re paying a GPU tax – wasting cycles that could be accelerating outcomes or serving more users. With DDN Infinia, we’re turning that cost center into a performance advantage.”
Infinia is DDN’s multi-year, ground-up redesigned object storage. It provides sub-millisecond latency, supports more than 100,000 AI calls per second, and is purpose-built for Nvidia’s H100s, GB200s, and Bluefield DPUs. DDN reminds us that Nvidia has said that agentic AI workloads require 100x more compute than traditional models. As context windows expand from 128,000 tokens to over 1 million, the burden on GPU infrastructure skyrockets – unless KV cache strategies are deployed effectively.
The company says that the traditional recompute approach with a 112,000-token task takes 57 seconds of processing time. Tokens are vector precursors, and their counts indicate the scope of an AI processing job. When the same job was run with DDN’s Infinia storage, the processing time dropped to 2.1 seconds, a 27-fold speedup. It says Infinia can cut “input token costs by up to 75 percent. For enterprises running 1,000 concurrent AI inference pipelines, this translates to as much as $80,000 in daily GPU savings – a staggering amount when multiplied across thousands of interactions and 24/7 operations.”
Alex Bouzari, CEO and co-founder of DDN, says: “In AI, speed isn’t just about performance – it’s about economics. DDN enables organizations to operate faster, smarter, and more cost-effectively at every step of the AI pipeline.”
It is unclear how DDN’s implementation compares to those from Hammerspace, VAST Data, and WEKA, as comparative benchmarks have not been made public. We would suppose that, as KV caching is becoming table stakes, suppliers such as Cloudian, Dell, IBM, HPE, Hitachi Vantara, NetApp, PEAK:AIO, and Pure Storage will add KV cache support using Nvidia’s Dynamo offload engine.
Bootnote
The open source LMCache software also provides KV cache functionality, as does the Infinigen framework.
AWS said customers can boot Amazon Elastic Compute Cloud (Amazon EC2) instances on AWS Outposts using boot volumes backed by NetApp on-premises enterprise storage arrays and Pure Storage FlashArray, including authenticated and encrypted volumes. This enhancement supports both iSCSI SAN boot and LocalBoot options, with LocalBoot supporting both iSCSI and NVMe-over-TCP protocols. Complementing fully managed Amazon EBS and Local Instance Store volumes, this capability extends existing support for external data volumes to now include boot volumes from third-party storage arrays, providing customers with greater flexibility in how they leverage their storage investments with Outposts.
…
Rai Way, part of the RAI group (Italy’s state TV broadcaster), and a specialist in digital infrastructure and media, has signed a Business Alliance Partnership with decentralized storage and compute provider Cubbit, adopting its DS3 Composer technology to integrate cloud storage services into its portfolio. Rai Way now offers a fully Italian edge-to-cloud service for storing customer data, enabled by Cubbit. The joint system ensures maximum resilience, data sovereignty, cost-effectiveness, and performance – even in low-connectivity areas – thanks to Rai Way’s distributed datacenter network, proprietary fibre infrastructure, and Cubbit’s geo-distributed technology. The initial capacity is 5 petabytes, with data hosted across Rai Way’s first five newly built edge datacenters in Italy.
…
DapuStor announced the launch of its ultra-high capacity J5060 QLC SSD, delivering 22.88 TB of capacity. Delivering up to 7,300 MBps read speeds at just 13 W, and maxing out under 25 W, it offers a best-in-class data-per-watt ratio. The J5060 is engineered for large, sequential data workloads. It uses coarse-grained (large-granularity) mapping and a dual-PCB hardware design to minimize DRAM usage and overcome capacity limitations.
– Sequential Read up to 7.3 GBps – Sequential Write up to 2.8 GBps – 4K Random Read 1.5M IOPS – 32 KB Random Write 15K IOPS – 4K Random Read Latency as low as 105 μs
…
Research house DCIG is researching the unstructured data management space. It says that, in the last several years, DCIG has published more than ten TOP 5 reports covering various aspects of unstructured data management. However, DCIG discovered a significant gap in the marketplace. “There’s really no good mental model for people to understand the overall unstructured data management space,” Ken Clipperton explains. After meeting with nearly 20 solution providers, he found widespread confusion about what constitutes unstructured data management. DCIG has devised a “7 Pillars of Unstructured Data Management” framework which provides a cohesive, vendor-neutral, actionable model for understanding the data management challenges faced by organizations today. It is creating a technology report on the seven pillars framework. This report will be available for licensing by solution providers in the second half of 2025.
…
Fast file and object storage array supplier DDN announced a strategic partnership with Polarise to deliver high-performance, resource-efficient AI infrastructure designed for the next generation of sovereign European AI workloads – at scale, with speed, and with sustainability in mind. Polarise specializes in building turnkey AI factories – end-to-end, AI-centric datacenters based on the Nvidia reference architecture. With locations in Germany and Norway, Polarise offers customers a sovereign alternative for consuming AI computing power through colocation, dedicated/private cloud, or direct API access via its own cloud platform. The companies will initially focus on joint deployments across Germany and Norway, with further European expansion planned in 2025 and beyond.
…
Deduping and fast restore backup target supplier ExaGrid announced its Tiered Backup Storage appliances can now be used as a target for Rubrik backup software. With advanced data deduplication, it will lower the cost of storing backups using the Rubrik Archive Tier or Rubrik Archive Tier with Instant Archive enabled, as compared to storing the data in the cloud or to an on-premises traditional storage. ExaGrid can achieve an additional reduction of 3:1 to 10:1 in addition to Rubrik’s compression and encryption, further reducing storage by as much as 90 percent. The combined deduplication is between 6:1 and 20:1 depending on length of retention and data types.
…
Streaming log data company Hydrolix announced support for AWS Elemental MediaLive, MediaPackage, and MediaTailor, as well as client-side analytics from Datazoom. The new integrations provide media and entertainment companies with real-time and historical insights into video streaming performance and advertising delivery, helping optimize viewer experience and ad revenue while significantly reducing the cost and complexity of data storage and analysis. The AWS Elemental and Datazoom integrations complement existing integrations with AWS CloudFront and AWS WAF, as well as other data sources.
…
HighPoint Technologies’ Rocket 1628A and 1528D NVMe Switch Adapters empower enterprise IT and solution providers to deploy up to 32 NVMe drives and 8 PCIe devices in a single x16 slot. Read more here.
…
Lexar NM990
Chinese memory supplier Longsys now owns the Lexar brand. Lexar has launched the NM990 PCIe 5.0 SSD with up to 14,000 MBps read and up to 11,000 MBps write (4 TB model), engineered for high-end gamers, professional creators and AI developers. It features:
1-4 TB capacity range
Thermal Defender Technology: Ensures 23 percent more efficient power consumption and a smoother experience
HMB and SLC Dynamic Cache: Delivers random read/write speeds up to 2000K/1500K IOPS for faster load times and reduced latency
3,000 TBW (4 TB model): Built for long-term durability and heavy workloads
Microsoft DirectStorage Compatible: Optimized for next-gen gaming performance
…
F5 and MinIO are partnering. The use cases addressed by F5’s Application Delivery and Security Platform (ADSP) combined with MinIO AIStor are:
Traffic Management for AI Data Ingestion (Batch and Real-Time): Enabling high-throughput, secure ingestion pipelines for data training and inference while eliminating pipeline bottlenecks and hotspots.
Data Repatriation from Public Clouds: Supporting organizations that are bringing data back to on-prem or hybrid environments to manage costs and comply with regulations.
Data Replication, Secure Multicloud Backup, and Disaster Recovery: Ensuring data availability, reliability, and security across geographies.
AI Model Training and Retrieval-Augmented Generation (RAG): Accelerating model development through real-time access to distributed data.
…
Cloud file services supplier Nasuni says it achieved top honors in the NorthFace ScoreBoard awards for the fifth year running. The NorthFace ScoreBoard Award is one of the industry’s most respected benchmarks for customer satisfaction. Nasuni achieved a Net Promoter Score (NPS) of 87, a Customer Satisfaction (CSAT) score of 98 percent and an overall ScoreBoard Index rating of 4.8 out of 5 – all above the industry average. Nasuni has also been awarded seven badges in G2’s Summer 2025 Reports, which are awarded based on real-user reviews and market presence.
…
Memory and storage semiconductor developer Netlist has expanded recently filed actions against Samsung and Micron in the U.S. District Court for the Eastern District of Texas. The two actions were originally filed on May 20, 2025, to assert Netlist’s new U.S. Patent No. 12,308,087 (“the ‘087 patent”) against Samsung’s and Micron’s High-Bandwidth Memory (HBM) products. The amended complaints add another Netlist patent, U.S. Patent No. 10,025,731 (“the ‘731 Patent”), against the two defendants’ DDR5 DIMM products, as well as their distributor Avnet, Inc. to these actions.
Netlist’s ‘087 patent entitled “Memory Package Having Stacked Array Dies and Reduced Driver Load” covers HBM for current and future AI applications. Analysts currently project U.S. HBM revenues to exceed $25 billion annually for each Samsung and Micron by 2031. The ‘087 patent will expire in November 2031. Netlist’s ‘731 patent reads on DDR5 DIMM with DFE and ODT/RTT circuits. Analysts currently project U.S. DDR5 DIMM revenue to exceed $65 billion annually in 2029. The ‘731 patent will expire in July 2029.
During the past two years, Netlist has obtained jury verdicts awarding combined total damages of $866 million for the willful infringement of its patents by Samsung and Micron.
…
Patriot Memory announced the launch of its latest flagship PCIe Gen5 SSD – the PV593 PCIe Gen 5 x4 M.2 2280 SSD, delivering industry-leading speeds of up to 14,000 MBps read and 13,000 MBps write. Powered by the advanced SMI SM2508 controller and built on TSMC’s 6nm process, the PV593 redefines next-gen storage for demanding workloads such as AI training, 4K/8K video editing, AAA gaming, and high-throughput multitasking. It features DRAM cache and dynamic SLC caching technology to reduce latency, accelerate system startup, and improve application launch times. With 4K random read/write speeds reaching up to 2000K/1650K IOPS, users can experience lightning-fast responsiveness across all tasks.
…
Vector search company Qdrant announced the launch of Qdrant Cloud Inference. This fully managed service allows developers to generate text and image embeddings using integrated models directly within its managed vector search engine offering Qdrant Cloud. Users can generate, store and index embeddings in a single API call, turning unstructured text and images into search-ready vectors in a single environment. Directly integrating model inference into Qdrant Cloud removes the need for separate inference infrastructure, manual pipelines, and redundant data transfers. This simplifies workflows, accelerates development cycles, and eliminates unnecessary network hops for developers.
Qdrant Cloud Inference is the only managed vector database offering multimodal inference (using separate image and text embedding models), natively integrated in its cloud. Supported models include MiniLM, SPLADE, BM25, Mixedbread Embed-Large, and CLIP for both text and image. The new offering includes up to 5 million free tokens per model each month, with unlimited tokens for BM25. This enables teams to build and iterate on real AI features from day one.
…
We’re told the CEO of Redis halted all product development for a week, just to ensure the entire team knew how to code with AI. Rowan Trollope, a developer-turned-CEO, was initially sceptical of vibe coding – a new phenomenon where developers use AI to help code. Now he’s pushing it across Redis after seeing first-hand how much vibe coding can accelerate delivery, improve code reviews, and reshape workflows. He vibe codes in his spare time, mostly for side projects, and while he knows AI won’t give him perfect code, he sees it as an accelerant. He’s one of the few CEOs with first-hand experience of vibe coding who has both the technical chops and the authority to enforce AI adoption from the top down.
…
Cyber-resilience supplier Rubrik unveiled upcoming support for Amazon DynamoDB, AWS’s flagship serverless, distributed NoSQL database service, and launched a proprietary cyber resilience offering for relational databases, beginning with Amazon RDS for PostgreSQL. Rubrik enables storage-efficient, incremental-forever backups and provides the flexibility to choose from a full range of Amazon Simple Storage Service (Amazon S3) storage classes, including Amazon S3 Standard, S3 Standard-Infrequent Access, S3 One Zone-Infrequent Access, S3 Glacier Instant Retrieval, S3 Glacier Flexible Retrieval, and S3 Glacier Deep Archive. Read a blog to find out more.
…
AI Data Cloud company Snowflake announced the appointment of Chris Niederman as Senior Vice President of Alliances & Channels. He joins Snowflake with more than 30 years of technology experience and previously spent 11 years at Amazon Web Services (AWS), most recently serving as Managing Director of the AWS Industries and Solutions team, leading the organization responsible for AWS’s worldwide partner strategy and industry transformation initiatives. Niederman will be responsible for leading the Snowflake global channel and partner ecosystem, and driving growth and collaboration through the Snowflake AI Data Cloud to empower every enterprise to achieve its full potential through data and AI.
…
DDN subsidiary Tintri announced “a game-changing and award-winning refresh to its global Partner Programme – introducing powerful new tools, exclusive benefits, and a reinvigorated partner experience designed to ignite growth, accelerate sales, and fuel innovation across the channel.” It includes technical certifications, pre-sales experts, marketing development funds, rewarding incentives and no barriers to entry to join the partner program. Click here to find out more.
…
Glenn Lockwood
VAST Data has hired Glenn Lockwood as Principal Technical Strategist. He joins from Microsoft, where he helped design and operate the Azure supercomputers used to train leading LLMs. Prior to that, he led the development of several large-scale storage systems at NERSC – including the world’s first 30+ PB all-NVMe Lustre file system for the Perlmutter supercomputer. Glenn holds a PhD in Materials Science. On AI inference, he says: “The hardest data challenges at scale went from ‘read a big file really fast’ to ‘search a vector index, rank and filter documents, load cached KV activations, cache new KV activations, and repeat.’ Not only are access sizes and patterns different across memory and storage, but application developers are expecting data access modalities that are much richer than the simplistic bit streams offered by files.” Read more in a VAST blog.
…
Veeam ran a survey into six months after the EU’s Digital Operational Resilience Act (DORA) came into effect. It found organizations are lagging: 96 percent of EMEA FS organizations believe they need to improve their resilience to meet DORA requirements.
DORA is a priority: 94 percent of organizations surveyed now rank DORA higher in their organizational priorities than they did the month before the deadline (Dec 2024), with 40 percent calling it a current “top digital resilience priority.”
The mental load of mounting regulations: 41 percent report increased stress and pressure on IT and security teams. Furthermore, 22 percent believe the volume of digital regulation is becoming a barrier to innovation or competition.
Barriers to compliance: Third-party risk oversight is cited by 34 percent as the hardest requirement to implement. Meanwhile, 37 percent are dealing with higher costs passed on by ICT vendors, and 20 percent of respondents have yet to secure the necessary budget to meet requirements.
…
Wedbush analysis about the AI phenomenon: “Our bullish view is that investors are still not fully appreciating the tidal wave of growth on the horizon from the $2 trillion of spending over the next 3 years coming from enterprise and government spending around AI technology and use cases. We have barely scratched the surface of this 4th Industrial Revolution now playing out around the world led by the Big Tech stalwarts such as Nvidia, Microsoft, Palantir, Meta, Alphabet, and Amazon.”
“Now the time has come for the broader software space to get in on the AI Revolution as we believe the use cases are exploding, enterprise consumption phase is ahead of us for 2H25, launch of LLM models across the board, and the true adoption of generative AI will be a major catalyst for the software sector and key players to benefit from this once in a generation 4th Industrial Revolution set to benefit the tech space. 2025 so far has been an inflection year within enterprise generative AI as true adoption has begun by going from idea to scale as more companies are looking to invest into AI to decrease costs/increase productivity. Looking forward its all about the use cases exploding which is driving this tech transformation being led by software and chips into the rest of 2025 and beyond and thus speaks to our tech bull and AI Revolution thesis further playing out over the next 12 to 18 months.”
It has “identified the 30 tech companies in our ‘IVES AI 30’ … that define the future of the AI theme over the coming years as we believe these tech stocks are closely tied to the AI Revolution ranging from hyperscalers, cybersecurity, software, semis, internet, and autonomous/robotics.”
…
WEKA co-founder and sometime COO Omri Palmon is departing for something, possibly a startup, that’s not public yet. A LinkedIn post and comments discuss this.
South Korea’s CXL memory-focused Panmnesia believes that AI clusters need both GPU node memory sharing and fast inter-GPU networking with a combined CXL and UALink/NVLink architecture.
Panmnesia has released a 56-page technical report titled “Compute Can’t Handle the Truth: Why Communication Tax Prioritizes Memory and Interconnects in Modern AI Infrastructure,” written by CEO Dr Myoungsoo Jung. The report outlines the trends in modern AI models, the limitations of current AI infrastructure in handling them, and how emerging memory and interconnect technologies – including Compute Express Link (CXL), NVLink, Ultra Accelerator Link (UALink), and High Bandwidth Memory (HBM) – can be used to overcome the limitations.
Dr Myoungsoo Jung
Jung stated: “This technical report was written to more clearly and accessibly share the ideas on AI infrastructure that we presented during a keynote last August. We aimed to explain AI and large language models (LLMs) in a way that even readers without deep technical backgrounds could understand. We also explored how AI infrastructure may evolve in the future, considering the unique characteristics of AI services.”
The technical report is divided into three main parts:
Trends in AI and Modern Data Center Architectures for AI Workloads
CXL Composable Architectures: Improving Data Center Architecture using CXL and Acceleration Case Studies
Beyond CXL: Optimizing AI Resource Connectivity in Data Center via Hybrid Link Architectures (CXL-over-XLink Supercluster)
The trends section looks at how AI applications based on sequence models – such as chatbots, image generation, and video processing – are now widely integrated into everyday life. It has an overview of sequence models, their underlying mechanisms, and the evolution from recurrent neural networks (RNNs) to LLMs. It then explains how current AI infrastructures handle these models and discusses their limitations:
Communication overhead during synchronization
Low resource utilization resulting from rigid, GPU-centric architectures
Jung writes in the report that no single fixed architecture can fully satisfy all the compute, memory, and networking performance demands for LLM training, inference prefill and decode, and retrieval-augmented generation(RAG). He suggests the best way to address the limitations is to use CXL, and specifically CXL 3.0 with its multi-level switch cascading, advanced routing mechanisms, and comprehensive system-wide memory coherence capabilities.
Panmnesia has developed a CXL 3.0-compliant real-system prototype using its core technologies, including CXL intellectual property blocks and CXL switches. This prototype has been applied to accelerate real-world AI applications – such as RAG and deep learning recommendation models (DLRMs) – and has proven practical and effective.
Jung then proposes methods to build more advanced AI infrastructure through the integration of diverse interconnect technologies alongside CXL, including UALink, NVLink, and NVLink Fusion, collectively called XLink.
He says “CXL addresses critical memory-capacity expansion and coherent data-sharing challenges.” But there are “specific accelerator-centric workloads requiring efficient intra-accelerator communications” such as “Ultra Accelerator Link (UALink) and Nvidia’s NVLink, collectively termed Accelerator-Centric Interconnect Link (XLink) in this technical report.”
Both CXL and XLink are needed to optimize AI super-clusters: “XLink technologies provide direct, point-to-point connections explicitly optimized for accelerator-to-accelerator data exchanges, enhancing performance within tightly integrated accelerator clusters. In contrast to CXL, these XLink technologies do not support protocol-level cache coherence or memory pooling; instead, their focus is efficient, low-latency data transfers among accelerators with a single-hop Clos topology interconnect architecture.”
He notes: “UALink employs Ethernet-based communication optimized primarily for large-sized data transfers, whereas NVLink utilizes Nvidia’s proprietary electrical signaling, tailored for small-to-medium-sized data exchanges, such as tensor transfers and gradient synchronization between GPUs.”
Panmnesia Technical Report diagram
So “integrating CXL and XLink into a unified data center architecture, termed CXL over XLink, including CXL over NVLink and CXL over UALink, leverages their complementary strengths to optimize overall system performance. … this integration adopts two architectural proposals: i) ‘accelerator-centric clusters,’ optimized specifically for rapid intra-cluster accelerator communication, and ii) ‘tiered memory architectures,’ employing disaggregated memory pools to handle large-scale data.”
Jung then proposes “an extended, scalable architecture that integrates a tiered memory hierarchy within supercluster configurations, explicitly designed to address the diverse memory-performance demands of contemporary AI workloads. This structure comprises two distinct memory tiers: i) high-performance local memory managed via XLink and coherence-centric CXL, and ii) scalable, composable memory pools enabled through capacity-oriented CXL.” The report discusses how these would be deployed in an AI data center, with notes on hierarchical data placement and management.
Composable server infrastructure startup GigaIO has raised $21 million in what it calls “the first tranche of its Series B financing.”
GigaIO produces SuperNODE, a 32-GPU, single-node AI supercomputer, and Gryf, a mini supercomputer-on-wheels with patented FabreX memory fabric architecture. This enables the scale-up and dynamic composition of compute, GPU, storage, and networking resources. It claims this unlocks “performance and cost efficiencies that traditional architectures are unable to deliver. As AI models grow larger and more complex, FabreX provides the flexibility needed to scale infrastructure on demand, at the rack level and beyond.”
CEO Alan Benjamin said in a statement: “Our vendor-agnostic platform uniquely frees customers from dependency on single-source AI chips and architectures. Whether it’s GPUs from Nvidia and AMD or new AI chips from innovators like Tenstorrent and d-Matrix, GigaIO enables customers to leverage the best technologies without vendor lock-in. This funding gives us the fuel to move faster and meet the surging demand.”
Jack Crawford, founding general partner at Impact Venture Capital, added: “As enterprises and cloud providers race to deploy AI at scale, GigaIO delivers a uniquely flexible, cost-effective, and energy-efficient solution that accelerates time to insight. We believe GigaIO has assembled a world-class team and is poised to become a foundational pillar of tomorrow’s AI-powered infrastructure, and we’re proud to back their vision.”
Alan Benjamin
GigaIO says it has a clear focus on AI inference. The new funding will be used to:
Ramp up production of SuperNODE, which it calls “the most cost-effective and energy-efficient infrastructure designed for AI inferencing at scale.”
Accelerate the deployment of Gryf, “the world’s first carry-on suitcase-sized AI inferencing supercomputer, which brings datacenter-class computing power directly to the edge.”
Invest in new product development to broaden GigaIO’s technology offerings.
Expand the sales and marketing teams to serve the increasing demand for vendor-agnostic AI infrastructure.
The round naming is unusual, as back in 2021, it raised $14.7 million in “the completion of a Series B round of funding.” Oddity aside, this latest “B-round” was led by Impact Venture Capital, with participation from CerraCap Ventures, G Vision Capital, Mark IV Capital, and SourceCode Cerberus.
Total public funding is now $40.25 million, and GigaIO plans “a second close of the Series B in the coming months, citing continued strong interest from strategic and financial investors.”
Yesterday we reported on Liqid’s refresh of its composable GPU server infrastructure to accommodate PCIe Gen 5 and CXL. It supports up to 30 GPUs versus GigaIO’s 32.
Liqid has announced products enabling host server apps to access dynamically orchestrated GPU server systems built from pools of GPU, memory, and storage, focused on AI inferencing and agents.
Liqid originally composed systems made up from CPUs and DRAM, GPUs and other accelerators, Optane, and NVMe SSDs. It offered dynamic system setup to provide precisely the amount of such resources needed for a workload, enabling better overall resource efficiency. Its design was based on PCIe Gen 4 in 2023. Since then, Optane has gone away, PCIe Gen 5 has arrived, and CXL memory sharing has developed. Liqid has now moved on to a PCIe 5 architecture, which has enabled it to support CXL 2.0 memory pooling.
Edgar Masri
CEO Edgar Masri said of this: “With generative AI moving on-premises for inference, reasoning, and agentic use cases, it’s pushing datacenter and edge infrastructure to its limits. Enterprises need a new approach to meet the demands and be future-ready in terms of supporting new GPUs, new LLMs, and workload uncertainty, without blowing past power budgets.”
Liqid’s latest development enables customers to get “the performance, agility, and efficiency needed to maximize every watt and dollar as enterprises scale up and scale out to meet unprecedented demand.” They can, Liqid says, achieve balanced, 100 percent resource utilization across on-prem datacenter and edge environments.
Its products are the Matrix software, composable GPU servers, composable memory, and I/O accelerators, which are souped-up SSDs such as the 64 TB AX-5500 in PCIe Add-in Card format. The Matrix software runs the composability show and talks via Fabric-A to a PCIe switch for composing storage and accelerators, and Fabric-B to a CXL switch for composing memory, DRAM, not HBM. Both fabrics access a compute (GPU server) pool to deliver their resources to GPUs and thus composed chunks of GPU+memory+storage capacity to hosts.
This latest Liqid setup uses v3.6 of the Matrix software and has six hardware elements:
EX-5410P 10-slot GPU box supporting 600 W GPUs (Nvidia H200, RTX Pro 6000, and Intel Gaudi 3), FPGAs, DPUs, TPUs, NVMe drives, and more.
EX-5410C with up to 100 TB of disaggregated composable memory using CXL 2.0.
LQD-5500 accelerated SSD now supporting up to 128 TB capacity, delivering <6 million IOPS and 50 GBps bandwidth in a FHFL PCIe Gen 5×16 card that hosts 8 x E1.S standard form factor NVMe Gen5 NAND drives. These drives are permanently affixed to the card.
PCIe Gen 5 switch.
CXL 2.0 switch.
Host bus adapters (HBAs).
There are two configuration options:
UltraStack dedicates up to 30 GPUs and 100 TB of DRAM to a single server.
SmartStack pools up to 30 GPUs across up to 20 server nodes and 100 TB of DRAM across up to 32 servers.
Matrix v3.6 has native compatibility with Kubernetes, Slurm, OpenShift, VMware, Ansible, and more. The fabric support will include Nvidia Fusion, Ultra Accelerator Link (UAL), or Ultra Ethernet Consortium (UEC) as they become available.
Updated datasheets were not available when we looked on Liqid’s website but they should be there soon.
Bootnote
In Liqid’s fourth generation, GPUs were installed in a 30-slot EX-4410 Expansion Chassis and connected to hosts across a PCIe Gen 4 network. Physical GPUs are provisioned from this pool to bare metal hosts with the Matrix software, which runs in a 1RU Liqid Director server.
A SmartStack 4410 system supports heterogeneous GPUs, up to 30 per host, and up to 16 host servers, PCIe Gen 4 with a 48-port PCIe switch, and up to three EX-4410s. A smaller SmartStack 4408 supports up to 24 GPUs per host with up to 16 hosts.
The UltraStack system supports bringing up to 30 Nvidia L40S GPUs to a single Dell PowerEdge server, either an R760 or R7625.
Vector buckets have been added to AWS S3 object storage to lower the cost of having infrequently accessed vectors stored in Amazon’s OpenSearch Service.
Unstructured data is being seen as increasingly important to organizations adopting large language model-based AI, with retrieval-augmented generation (RAG) and agents being part of that. RAG brings proprietary information to the LLMs and agents. The bulk of it is unstructured and needs vectorizing so that the LLMs and agents can bring their vector-based analysis to bear. Special LLMs do the vectorizing and then the vectors need storing. AWS is already doing that with its OpenSearch Service. However, it has a disadvantage compared to on-prem object stores and databases that store vectors. An AWS database instance costs in terms of storage capacity and also compute. By storing infrequently accessed vectors in S3, AWS can reduce compute costs for vector databases and save customers money.
Channy Yun, AWS Principal Developer Advocate for AWS cloud, writes in a blog: “We’re announcing the preview of Amazon S3 Vectors, a purpose-built durable vector storage solution that can reduce the total cost of uploading, storing, and querying vectors by up to 90 percent.”
This makes it cost-effective to create and use large vector datasets to improve the memory and context of AI agents as well as semantic search results of customers’ S3 data. The S3 Vectors feature is, an AWS announcement says, “natively integrated with Amazon Bedrock Knowledge Bases so that you can reduce the cost of using large vector datasets for retrieval-augmented generation (RAG). You can also use S3 Vectors with Amazon OpenSearch Service to lower storage costs for infrequent queried vectors, and then quickly move them to OpenSearch as demands increase or to enhance search capabilities.”
S3 vector support comes via a new vector bucket, which Yun says has “a dedicated set of APIs to store, access, and query vector data without provisioning any infrastructure.” The bucket has two types of content: vectors and a vector index used to organize the vectors. “Each vector bucket can have up to 10,000 vector indexes, and each vector index can hold tens of millions of vectors,” Yun says.
There’s a neat extra twist: “You can also attach metadata as key-value pairs to each vector to filter future queries based on a set of conditions, for example, dates, categories, or user preferences.” This cuts down vector selection and scan time.
S3 Vectors automatically optimizes the vector data to achieve the best possible price-performance for vector storage as bucket contents change.
The S3 Vectors feature is integrated with Amazon Bedrock Knowledge Bases, including within Amazon SageMaker Unified Studio, for building RAG apps. When creating a knowledge base in the Amazon Bedrock console, you can choose the S3 vector bucket as your vector store option.
AWS S3 Vectors integrations
S3 Vectors also has an integration with the Amazon OpenSearch Service. If you do lower storage costs by keeping low access rate vectors in S3 Vectors, you can move them to OpenSearch if needed and get real-time, low-latency search operations.
Andrew Warfield, a VP and Distinguished Storage Engineer at AWS, said: “S3 vectors anchors cost in storage and assumes that query demand fluctuates over time, meaning you don’t need to reserve maximum resources 100 percent of the time. So when we look at cost, we assume you pay for storage most of the time, while paying for query/insertion costs only when you interact with your data.
“When we looked across customer workloads, we found that the vast majority of vector indexes did not need provisioned compute, RAM or SSD 100 percent of the time. For example, running a conventional vector database with a ten million vector data set can cost over $300 per month on a dedicated r7g.2xlarge instance even before vector database management costs, regardless of how many queries it’s serving.
“Hosting that same data set in S3 will cost just over $30 per month with 250 thousand queries and overwriting 50 percent of vectors monthly. For customers who have workloads that heat up, they can also move their vector index up to a traditional vector store like Open Search, paying instance-style cost for the time that the database is running hot.”
Yun’s blog post explains how to set up and use S3 Vectors.
AWS is not alone in adding vector support to object storage. Cloudian extended its HyperStore object storage with vector database support earlier this month, using the Milvus database.
The Amazon S3 Vectors preview is now available in the US East (N. Virginia), US East (Ohio), US West (Oregon), Asia Pacific (Sydney), and Europe (Frankfurt) Regions. To learn more, visit the product page, S3 pricing page, documentation, and AWS News Blog.
