Pure Storage has opened its arms to white box storage vendors after deciding that servicing the world’s biggest AI rigs is beyond any one storage company.
The all-flash storage pioneer’s FlashBlade//EXA platform, which it unwrapped today, sees the vendor’s inhouse technology focused specifically on handling metadata in hyperscale AI setups.
More mundane storage will be handed off to vanilla kit – although those white boxes could come from Pure itself.
International CTO Alex McMullan said the strategy “reflects a orthogonal change…in terms of making effectively a storage supercomputer for the next iteration of scale.”
Hyperscalers’ AI ambitions are so far-reaching that “the current product simply wasn’t intended to grow to the same size. So we’ve taken that decision to go where we are at this point in time.”
The incredibly expensive, and power hungry, GPUs that hyperscalers and model developers are spending so many dollars on are already struggling to hit 25 percent utilization, he said, because legacy storage architectures weren’t up to the job. Meanwhile, those GPUs are still sucking in the equivalent power of a typical UK household.
“The hyperscalers are trying to save every single watt they can, because every watt they save, they can turn on another GPU,” he said.
These organizations are looking at 50TB per second or better data rates, he added. “We just don’t think that’s going to be doable with our existing architecture.”
Pure’s solution is to take its FlashBlade architecture and use that “as a metadata engine for the next generation storage supercomputer.” That layer will connect to the GPU cluster over NFSv4.1 over TCP. Meta data has emerged as a key bottleneck in AI systems.
Then, he said, “If we start with something that can run a SuperPOD on its own and just use that for metadata, then we should be able to keep up with everything else that sits around it.”
The everything else comes in the shape of software-defined data nodes connecting into the cluster via NFSv3 over RDMA. The metadata engine will then direct the GPU nodes to the data nodes layer.
These could be white boxes, he said, or in time a Pure Storage node. “White boxes just run on Linux kernel with an NFS server,” said McMullan. “They are standard. Vanilla.” The only requirement on the white boxes will be to sustain two 400 GB NICs.
The metadata engine could grow to “hundreds of petabytes if we want it to. Honestly, if the metadata is 100 petabytes, we have a bigger problem on its own.” As well as tackling the GPU utilization problem, the architecture will make it easier to have AI datasets in one place, reducing the impact of fragmentation.
McMullan said the ultimate target was the problems hyperscalers and similar operators were expecting to tackle in two years’ time. “Our existing flash blade is good enough for what’s there today, but some of the new requirements customers are talking about are making us take that pause.”
He said the company had a system in the lab running around 300 nodes delivering roughly 30 TBps.
Some early access beta customers were already in place, with “three or four of those in each of the theaters.” General availability for FlashBlade//EXA is slated for the summer. But Pure itself will have to join the queue for Nvidia certification, McMullan said, which should be signed off in the second half of the year.
McMullan was clear that the product was targeted at a small section of the market, comprised of “AI natives, people are building models, who are tuning models, who are looking at RAG-based capabilities.”
So, effectively hyperscalers and the biggest of AI factories. It could also include some life sciences organizations and some telco customers. And it will likely include some high end government customers.
It’s unlikely that many enterprises will be tyre-kicking these systems though, McMullan said.
“We don’t think there’s a play. Although, given all the things that have happened in the last couple of years in this space, who knows?”
Druva’s relationship with Microsoft has been upgraded to “strategic,” meaning the data security vendor’s tech will be more tightly integrated into Azure.
The firms already work together to provide protection through the “Microsoft ecosystem,” spanning Redmond’s Windows and 365 platforms and “multiple Azure technologies.”
But Stephen Manley, Druva CTO, claimed the new relationship was “vastly different” and “encompasses deeper technical integration with Druva and Microsoft Azure cloud services.”
This will allow customers to “protect and secure cloud and on-premises workloads” and choose Azure as a storage target, the companies say.
The integration, we’re told, will enable “enhanced cyber resilience with cross-cloud protection, unified visibility across data environments, and more comprehensive data security strategies.”
Druva also has a tight relationship with AWS, under which it offers managed storage and security on the AWS platform. Which sounds a lot like the new Microsoft relationship.
Manley was at pains to point out “We’re not ‘porting’ features as much as delivering the same features in a different environment. Each cloud is unique and it requires customized integration to deliver the value and functionality the customers need.”
This means data could be stored on either platform. Manley said this makes it easier for enterprises to manage data and risks, even as new security threats arise and new compliance regimes sprout up.
“If most of their data resides in one cloud, they can use Druva to back it up across clouds. If they want copies of critical on-prem and cloud systems to be stored in both AWS and Azure, they can. If companies have sites in multiple locations, they can choose a separate cloud for each.”
Druva has long been tipped as an IPO candidate, without getting round to filing papers to kick off the process. But being able to run a bigger Azure logo in the future is unlikely to hurt its credibility with potential investors.
Manley insisted Druva’s immediate aim was to build a “long-lasting, durable brand.” And going public? “As we continue to scale, a public market offering could be in the cards, but that isn’t the goal — that’s an outcome driven by our growth.”
The worst excesses of NAND price volatility appear to be coming to an end, as Wall Street analysts throw their weight behind newly independent Sandisk.
According to reports over the weekend, Mizuho tech analyst Jordan Klein sent a note to investors last week saying that Sandisk had told customers it would be jacking up NAND product prices by at least 10 percent from the end of this month.
Apparently, Sandisk has told customers it expects demand to begin outstripping supply, with rising tariffs also feeding into its decision to markup the price list.
No one likes having to pay more for anything. At the same time, a 10 percent price bump would be a more palatable hike than the 30 percent price growth the industry was expecting earlier this year.
We’ve contacted Sandisk and Klein but have yet to hear back from either.
Sandisk only separated from erstwhile parent Western Digital at the end of last month. Over the last couple of weeks a number of other Wall Street analysts have begun coverage of the company with overweight recommendations, implying they believe the stock will outperform. Which in turn suggests they foresee a more predictable pricing environment and are not expecting prices to head south anytime soon.
Trendforce adds that other suppliers had already curbed production due to oversupply.
Philip Kaye, cofounder of Manchester-based datacenter infrastructure supplier Vespertec, said that the market can, at least, expect a lull in NAND price volatility. A post-Covid glut and the war in Ukraine had driven prices to record lows, he said, before giving way to a period of enormous fluctuations.
“Recently, however, the rate of price increases has levelled off to around 5-6 percent, a trend likely to continue into next year barring any external shocks.” Which should be reassuring. If you haven’t come to the conclusion that in the future, “external shocks” are the new normal.
SK Hynix will complete its $9 billion takeover of Intel’s NAND memory business in the coming weeks, setting the seal on the stumbling US giant’s exit from the flash business.
The news emerged as US president Donald Trump pontificated on the fate of Intel, declaring that its CEOs post Andy Grove (who vacated the role in 1998) had lost direction and allowed Taiwan to “steal” the chip business from the US.
Korean outlets reported this weekend that SK Hynix was poised to hand over the last $2.24 billion of the $9 billion price tag set on the business in 2020. The deal includes the unit’s Dalian China NAND business, which became part of SK Hynix. The SSD division became a separate subsidiary dubbed Solidigm.
The first phase of the sale saw SK Hynix take over the Dalian factory and the SSD operation, but not all the relevant intellectual property. The conclusion of the deal was slated for some time after March 2025.
With the final part of the payment made, SK Hynix will gain full control of the manufacturing IP and R&D at the Dalian unit. The buy will leave SK Hynix in a much stronger position against industry giant Samsung.
Meanwhile, SK Hynix is exiting the image sensor market to concentrate on AI focused memory, according to reports over the weekend. It is a key supplier of high bandwidth memory to GPU giant Nvidia.
As for Intel, the conclusion of the divestment represents another stage in its ongoing shrinkage. When it comes to flash, Intel had already shuttered its Optane memory business that was designed to offer faster data access than traditional SSDs. Optane was also pitched as being cheaper than DRAM but with slower access speed.
With the world of politics and technology interwoven these days, Donald Trump last week took time out from cutting intel-sharing with Ukraine to lambast Intel’s leadership since the CEO-ship of co-founder Andy Grove.
In a White House press conference centered on a $100 billion TSMC investment in the US, Trump repeated his previous claims that Taiwan “stole” the chip industry from its pioneers in the US. But he didn’t blame Taiwan for this. Rather, he said, he fingered previous incumbents of the White House for allowing that to happen.
As for Intel, he said, the company had lost its way after the Grove era. Trump described Grove, whose autobiography was titled “Only the Paranoid Survive” as a “tough smart guy” adding “I used to read about him when I was a young man.”
“He did an incredible job, he really dominated the chip businesses, and then he died and I guess they had a series of people that didn’t know what the hell they were doing and we gradually lost the chip business and now it’s almost exclusively in Taiwan.”
TSMC has pledged to build five plants in Arizona. There will be no government funding involved, and Trump took the opportunity to disparage predecessor Joe Biden’s plans to boost US chip manufacturing through subsidies. Trump claimed TSMC’s decision was a result of its fears over his plans to impose tariffs on pretty much everyone.
Kubernetes app data protector CloudCasa has integrated with SUSE Rancher Prime via a new Rancher Prime Extension. Users can install CloudCasa agents and manage and monitor backups directly from the Rancher Prime UI. They get multi-cluster management across clouds and on-premises, a single pane of glass for VM and container management, migration across any Kubernetes distribution, and “enterprise-grade security, compliance, and governance.”
…
Flexera announced that the Spot acquisition from NetApp is complete and “now has the most comprehensive FinOps offering in the market,” which “empowers organizations and MSPs to manage cloud financial commitments, automate billing and invoicing, reduce workload costs, and optimize containers.”
…
GPU RAID card startup Graid has raised $30 million in a Series B, bring its total funding to around $50 million. The round was led by HH-CTBC Partnership, a joint venture fund between Foxconn and CTBC, alongside Yuanta Ventures, and included participation from Delta Electronics Capital, Harbinger Venture Capital, and returning investors from Graid Technology’s 2022 Series A round. Graid will use the cash for “global expansion, product innovation, and strategic partnerships, strengthening Graid Technology’s presence in enterprise and OEM markets while meeting growing demand for AI, machine learning, and high-performance computing workloads.”
…
Data orchestrator Hammerspace has partnered with Yition.ai, a Chinese company aiming to democratize AI. Yition is little known outside China. Hammerspace says Yition has “cost-effective and scalable AI storage” offerings. The integration with Hammerspace will add “the high-performance data path to power large-scale compute clusters efficiently, the data orchestration to unify data sources, and standards-based approach to use the compute, storage, and networking infrastructure of the customer’s choice.”
…
HarperDB is rebranding as Harper, “removing the database notation from its corporate name and identity. This simplified name, new brand identity, and redesigned website reflects the company’s evolution from a database vendor focused on performance, to a full stack application delivery platform that enables data architects and development teams to transform the economics of their organization through the accelerated speed and scale of data-intense workloads.”
…
Huawei’s Dr Peter Zhou, president of its Data Storage Product Line, said at MWC 2025 Barcelona that its “AI-Ready data lake breaks data silos, making data visible, manageable, and available.” Its “Data Storage provides the AI-Ready data lake solution, diverse data storage services, and the FlashEver business model, empowering carriers to turn their disordered data into high-quality assets to unlock the value of data.” Huawei has launched the “New-Gen OceanStor Dorado Converged All-Flash Storage and OceanStor A Series High-Performance AI Storage. These solutions boast 100 million-level IOPS, financial-grade reliability, and efficient AI training and inference, supporting tens of billions of daily charging services and robust mobile financial services.” The New-Gen OceanStor Pacific All-Flash Scale-Out Storage provides industry high density and low power consumption with exabyte-level scalability.
“Another new offering is the New-Gen OceanProtect All-Flash Backup Storage for data protection. The storage offers five times faster data recovery than industry alternatives. The Huawei DCS AI Solution provides a one-stop AI full-process toolchain and containerized environment, accelerating fine-tuning and large-scale deployment of AI models. The FlashEver business model provides an evolutionary, flexible architecture to enable seamless upgrades for live-network equipment.”
…
Public cloud backup storage service supplier Keepit announced the publication of its new report, “Intelligent data governance: Why taking control of your data is key for operational continuity and innovation,”which “covers the foundational importance of data control in the age of AI, with a focus on ensuring modern enterprises’ cyber resilience and compliance.” The “report finds that data governance is a key tool when striving to stay compliant as regulations such as NIS2, DORA, and GDPR are impacting organizations.” Download the report here.
…
Lightbits says a “massive online retailer, with multimillions of products in as many different categories, powers its eCommerce business with the Lightbits fast and flexible cloud data platform.” The multibillion-dollar company commands the largest share of the market in their region. The platform currently services millions of registered users, millions of daily visits, and delivers millions of shipments per month – and those numbers are increasing rapidly. “The Lightbits Cloud Data Platform forms the ideal combination to provide disaggregated, and composable storage and integrate natively with their Kubernetes environments.”
“The implementation consists of tens of clusters of 16 dual instance, high-performance nodes per datacenter – two instances of Lightbits are run on the same physical machine. This allows Lightbits to get 25-30 percent higher performance out of a single machine. Each Lightbits node consists of Lenovo ThinkSystem SR650 V2 servers with high-performance Intel 3rd Gen Xeon Scalable Processors, Intel Optane Pmem, and Intel Ethernet Adapters, and Micron NVMe SSDs. This implementation schema delivers performance and data replication advantages. This organization can leverage different replication scenarios for different workloads.”
…
Micron has enhanced its LPDDR5X memory with error correction code for the automotive market. The “LPDDR5X-optimized error correction code (ECC) scheme mitigates all system in-line ECC penalties delivering a 15 to 25 percent bandwidth increase. This new LPDDR5X-optimized ECC scheme, called direct link ECC protocol (DLEP), not only delivers increased performance but also helps LPDDR5X memory systems achieve the ISO 26262 ASIL-D hardware metric through reduced failures in time (FIT).” It “delivers approximately 10 percent lower power consumption on a pJ/b (picojoule-per-bit) perspective and a minimum 6 percent additional addressable memory space.”
…
Micron announced the world’s first G9-based UFS 4.1 and UFS 3.1 mobile storage products to accelerate AI on smartphones. The UFS 4.1 offering provides proprietary firmware features for flagship smartphones such as zoned UFS, data defragmentation, a pinned writebooster, and intelligent latency tracker. More information here.
…
Nexla has enhanced its Integration Platform, “expanding its no-code integration, RAG pipeline engineering, and data governance capabilities.” You can “integrate any data, create AI-ready data products, and deliver GenAI projects without coding, and up to 10x faster than the alternatives.” Nexla uses AI to connect, extract metadata, and transform source data into human-readable data products, called Nexsets, that enable data reuse and governance. Its agentic retrieval-augmented generation (RAG) framework lets companies implement RAG for agents and assistants without coding, and uses LLMs during each stage to improve accuracy. For example, Nexla can get context from multiple data products, use a unique algorithm to rank, prioritize, and eliminate data, and then combine the context with a rewritten query and submit it to just about any LLM.
…
OpenDrives CEO Sean Lee has reorganized the company to form two leadership teams, largely promoting existing execs. The Corporate Leadership Team (CLT) oversees overall company strategy and shareholder value, and the Senior Management Team (SMT) drives day-to-day operations and execution. Stefan Grycz joined OpenDrives as director of sales in January 2025. Joel Whitley, partner at IAG Capital Partners, said: “As lead investor, IAG is fully committed to driving OpenDrives’ success. We are confident that this leadership team will enable OpenDrives to continue its growth and out-innovate the competition. The OpenDrives software platform is poised to change how M&E and DevOps engage with data and storage in ways no one else can.”
…
Phison announced that its PCIe Gen4x4 PS5022 controller, designed for automotive applications, has become the world’s first SSD controller to receive ISO 26262 ASIL-B compliance certification. This ensures that electronic components within a vehicle can detect random failures in real-time and can transition into a safe state to support impact prevention for driving safety.
…
Data integrity supplier Precisely has a DataLink partner program to streamline the integration of the Precisely data portfolio with data from trusted providers via pre-linked datasets. Inaugural partners include Precisely, GeoX Analytics, Overture Maps Foundation, and Regrid, with additional partners coming soon. The “Data Link program transforms the traditionally complex and time-consuming process of mapping disparate third-party datasets into a seamless customer experience. Organizations no longer need to stitch together data from multiple vendors, navigate complex integrations, or evaluate dozens of disconnected datasets.”
…
Vector database supplier Qdrant has updated its Qdrant Cloud with “single sign-on (SSO), cloud role-based access control (RBAC), granular database API keys for granular RBAC, advanced monitoring and observability with Prometheus/OpenMetrics to connect external monitoring systems, and a cloud API for seamless automation.” Learn more here.
…
The Qumulo Cloud Data Fabric is available now with Qumulo 7.4.1 and “enables organizations to make any data, in any location, instantly available with total control. Customers can create data portals in under 10 seconds that enable remote users to connect to and access data anywhere in the world, as if it were local. Predictive Caching enhancements improve application performance up to 2X compared to existing solutions by inspecting data access patterns and prefetching data into the cache before a client requests it.” Customers can “Read and Write data with strict data consistency across the Cloud Data Fabric, instantly, with assured data and application correctness.” Also, “loading only folder and file metadata at the remote site improves application responsiveness while preserving WAN bandwidth.”
It’s available globally through major IT infrastructure resellers and system vendors, including Hewlett Packard Enterprise and Supermicro, distributors, and most major public clouds, with prepay and pay-as-you-go options. Pricing is based on the data stored and shared across Qumulo’s data core infrastructure. Learn more here.
…
Qumulo has partnered with large file transfer business MASV to create a high-performance, enterprise-grade data ingest and access pipeline enabling seamless petabyte-scale data ingest into Cloud Native Qumulo (CNQ). This supports large-scale AI model training and data analytics with scalable cloud-native infrastructure. Find out more here.
…
Application HA and DR supplier SIOS announced that SIOS LifeKeeper and SIOS DataKeeper clustering software have been validated for use with Cimcor’s cybersecurity offering, the CimTrak Integrity Suite. This allows Cimcor customers to “seamlessly integrate high availability and disaster recovery into their CimTrak environments, ensuring continuous protection against cyber threats and minimizing downtime in critical cybersecurity operations.”
…
View Systems, in collaboration with Alpha3 Cloud, announced the launch of Managed View AI, a comprehensive turnkey GenAI platform designed for enterprises and government organizations that need AI-driven insights while retaining full control of their data. “We are pleased to work with Alpha3, Ampere and other AI Platform Alliance members to advance enterprise AI deployment standards and practices while enabling deeper integration of our technologies,” said Keith Barto, chief product and revenue officer at View Systems. “Managed View AI allows organizations to deploy AI workloads with confidence and optimize their deployment based on specific workload requirements, with options to add GPU or other accelerator support as needed.”
…
ReRAM developer Weebit Nano has a partnership with ultra-low-power AI computing business Embedded AI Systems (EMASS). The EMASS SoC “delivers unparalleled energy efficiency and cost advantages, with best-in-class AI capacity.” Weebit ReRAM “delivers lower power, faster read and write speeds, and cost advantages compared to traditional non-volatile memory (NVMe) technologies like flash.” The combined system “can deliver new levels of performance and power efficiency for edge AI.” Attendees at the embedded world 2025 Conference & Exhibition can see a live demonstration of the integrated technologies.
…
Software RAID supplier Xinnor has partnered with Advanced HPC to implement an improved backup and restore system for a US federal agency focused on scientific research. The agency sought to address performance bottlenecks and enhance data throughput to its Spectra Logic T950 tape library and now uses a system using Xinnor’s xiRAID with 12x 30 TB and 2x 6.4 TB Gen 4 NVMe SSDs “to meet the agency’s needs for high-speed read and write capabilities, seamless integration with existing infrastructure, and scalability for future growth.” Read the case study here.
Agentic AI is coming to startup Weaviate as it develops three agents to perform tasks using its open-source vector database.
The Weaviate agents use large language models (LLMs) pre-trained on its APIs to perform tasks in the Weaviate vector database environment. The LLM pre-training makes them “experts in performing Weaviate-specific data tasks based on natural language commands.” Three agents are being launched in preview mode this month.
Alvin Richards
Weaviate VP of Product Alvin Richards stated: “Weaviate’s development tools come with batteries included. By unifying data management, agentic workflows and vector storage and search on our enterprise-class infrastructure, we empower development teams to quickly create applications that bring intelligent AI to the masses.”
The first one, the Query Agent, accepts queries in natural language, decides which data in Weaviate is relevant, formulates the necessary searches, retrieves the data, correlates and ranks the answers, and then returns the results. It can chain commands together, taking the results of a previous query and extending it with a new prompt. This Query Agent can simplify complex query workflows and accelerate Retrieval Augmented Generation (RAG) pipelines.
The company says the Transformation Agent can, based on the natural language instructions passed to it, automatically update data, create new properties, add new data, and more. This agent, according to Weaviate, can be used for cleaning and organizing raw data for AI, generating and enriching metadata, automatically categorizing, labeling and preprocessing data, or translating an entire dataset.
The Personalization Agent “can go beyond static, rules-based recommendations and deliver smart, LLM-based personalization on the fly,” making “it easy to curate search results tailored to each user’s preferences and interactions.” We’re told it can “be used in conjunction with Weaviate Query and Transformation Agents to deliver hyper-personalized user experiences in real-time.”
Bob van Luijt
Weaviate Agents will become available in Weaviate Cloud as they are released, including the free Developer Sandbox. Query Agent is available now, and Transformation and Personalization Agents are coming later this month.
Bob van Luijt, Weaviate CEO and co-founder, said: “Vector embeddings have been at the core of AI’s development—from early deep learning models to transformers and today’s large language models. What started as a linear process—data to vector, to database, to model, to results—evolved into dynamic feedback loops, giving rise to agentic architectures. This milestone is a natural next step in a journey we saw beginning a decade ago. … And what’s most exciting is that this is just the beginning.”
Cyber-resilience dominates the latest Rubrik features, with a dozen new protection points in its latest rollout that it says will help detect, repel, and recover from cyberattacks.
The company is moving its protection product line across more environments, including the public cloud, SaaS and on-prem apps, and enhancing its ability to detect threats and verify user identities.
Arvind Nithrakashyap, Rubrik CTO and co-founder, stated: “We are seamlessly integrating new technologies across the world’s major cloud platforms, SaaS offerings, and on-premises so our customers can better detect compromised data, enhance the speed of identifying affected data, and accelerate the discovery of clean entry points.”
The new capabilities in the public cloud include:
Cloud Posture Risk Management (CPR), which automatically discovers and inventories cloud data assets, identifying unprotected or sensitive data en route.
Oracle Cloud Protection: Rubrik Security Cloud (RSC) will support data protection for Oracle Cloud Infrastructure (OCI) beginning with Oracle databases and Oracle Cloud VMware Solution (OCVS).
Azure DevOps and GitHub Backup: Rubrik now protects Azure DevOps and GitHub with automated backups, granular recovery, extended retention, and compliance coverage.
Rubrik Cloud Vault (RCV) for AWS provides a secure off-site archival location, with flexible policies and/or regions, and immutable, isolated, logically air-gapped off-site backups, role-based access controls, encryption, and retention locks.
The SaaS area has two items, enhanced protection for Microsoft Dynamics 365 and Sandbox Seeding for Salesforce, which is planned for later this year. Users can select objects and records depending on specific criteria to prevent seeding errors by analyzing data selection size versus destination size availability before moving data to the sandbox environment. Users will be able to save queries for future repetitive use.
The on-prem world gets Identity Recovery across Entra ID and Active Directory (AD). It includes orchestrated Active Directory Forest Recovery to restore entire identity environments without reintroducing malware or misconfigurations.
Rubrik now protects PostgreSQL with data backup, availability, and recoverability. It has also added Red Hat OpenShift support with automated, and immutable backups and fast recovery.
The company has extended its anti-malware functionality:
New security features for Azure and AWS, which use machine learning and automation, include Anomaly Detection, Data Discovery, and Classification, and soon, Threat Hunting and Threat Monitoring. They are all designed to work together “to proactively detect and mitigate cyber threats, accelerate recovery, and ensure sensitive data remains protected and compliant.”
Rubrik is planning to extend its Orchestrated Recovery capabilities to the cloud beginning with Azure VM and featuring automated recovery sequences, regular test recovery scheduling, and recovery reports to reduce human error.
Turbo Threat Hunting scans at scale by using pre-computed hashes in Rubrik’s metadata, with no need for file-by-file scanning. It claims clean recovery points can be found in seconds. Testing found Turbo Threat Hunting scans 75,000 backups in up to 60 seconds.
Enterprise Edition for Microsoft 365 is covered with Sensitive Data Discovery, to identify and protect high-risk data before an attack happens, and Prioritized Recovery, which restores critical data first. Coming soon are Anomaly Detection, Threat Monitoring, Threat Hunting, and Self-Service Recovery capabilities.
Nithrakashyap says: “Cybercriminals won’t stop innovating, and neither will we. Our utmost priority is the security, safety, and appropriate accessibility of our customers’ data, regardless of where the data lives.”
As long as cybercriminals invent new methods and attacks, Rubrik can respond with new features to keep its subscription-paying customers feeling safe and ready to repel attacks.
Bootnote
Rubrik itself suffered a security intrusion last month. A February note by Nithrakashyap says the company “recently discovered anomalous activity on a server that contained log files.” It took the server offline. “An unauthorized actor accessed a small number of log files, most of which contained non-sensitive information. One file contained some limited access information … We have rotated keys to mitigate any residual risk, even though we found no evidence that access information was misused.”
He emphasizes that “after a detailed analysis with the third party partner, we have found no evidence of unauthorized access to any data we secure on behalf of our customers or our internal code.”
PEAK:AIO has a new 2RU 1.5 PB AI Data Server product, using Dell hardware, that ships data at 120 GBps.
PEAK:AIO is an AI-focused UK storage startup that supplies software-defined storage on third-party hardware, which it manages and controls closely to cut latency and increase throughput. Its 2RU servers have been delivering 40 GBps to mid-size GPU clusters, using NFS and NVMe-oF, and can now go three times faster.
Mark Klarzynski
The new hardware incorporates Solidigm’s 61.44 TB QLC SSDs and data is pumped across Nvidia’s CX7 Ethernet NICs. PEAK:AIO says specialized fields such as healthcare and research need pretty much the same high-speed GenAI model capabilities as those produced by massive datacenter infrastructure systems. It says its systems are less costly but still very fast and energy-efficient. PEAK says it has a new NVMe software stack that “eliminates legacy Linux bottlenecks” to provide better performance for data-intensive AI apps.
Co-founder and Chief Strategy Officer Mark Klarzynski stated: ”Our approach addresses both ends of the AI infrastructure spectrum. For smaller, innovative projects, this 2U solution delivers unmatched performance and energy efficiency in a compact form. For large-scale deployments, it becomes the foundation of a new generation of scale-out file systems, purpose-built to meet the energy and performance demands of modern GPUs.”
The CX7 requires a PCIe Gen 5 bus and it maxes out at 50 GBps. We understand that Dell’s PowerEdge R7625 rack server has a 2RU chassis and supports up to eight PCIe gen5 slots. Three single-port CX7s, each using a PCIe gen5 x 16 lane slot, will deliver 120 GBps. The R760 supports up to 24 NVMe SSDs, and filling these slots with the 61.44 TB Solidigm SSDs produces 1.5 PB of raw capacity.
Klarzynski tells us: “Dell Technologies validated Solidigm’s 61TB NVMe drives and NVIDIA’s 400Gb crypto interfaces within the R7625 exclusively for us, outpacing the mainstream 15TB solutions. … This move highlights our scale and global ambitions. Our 2U AI Data Server (1.5PB, 120GB/sec) sets a new industry benchmark for efficiency.”
The AI Data Server can be connected directly to an Nvidia GPU server for small or new-start projects, with a switch being added to hook it up to multiple GPU servers, up to 10 of them. The data servers can be scaled out to support more GPUs.
Dell PowerEdge R7625
PEAK says it enjoyed 400 percent growth in 2023 and 2024, much in the USA, and plans to launch a new range of products later this year to complement its new AI Data Server that will, “with Dell Technologies’ backing … disrupt energy, cooling, and density challenges.” It also has an upcoming line-rate cryptographic offload capability integrated with the CX7.
Interview: We had the opportunity to discuss storage servers with Supermicro and find out how it designs and customizes them for its customers, and how interfaces are changing. Wendell Wenjen, director of Storage Market Development, answered our questions.
Blocks & Files: How does Supermicro design a storage server, other than stuffing either SSDs or HDDs in a standard 24-slot chassis?
Wendell Wenjen
Wendell Wenjen: Supermicro designs a very wide array of servers ranging from enterprise, embedded, multi-node, GPU servers to cloud optimized servers. While many of these servers are used for storage applications, we have a dedicated product management and development team focused on servers specifically designed for storage applications.
The key features of these servers are capacity of SSD, HDD, or both, use of EDSFF SSD form-factors, mainly E3.S and E1lS but also including U.2, and validation of these systems with leading storage ISV software. We also make specialized storage servers which are unique for storage including SBB-type two-node high-availability systems, JBODs, JBOFs and, most recently, a 1U 16-bay storage server using the Nvidia Grace CPU Superchip, specifically designed for scale-out storage applications.
Blocks & Files: Does Supermicro customize its storage servers for particular customers, such as hyperscalers? How does this process work?
Wendell Wenjen: While Supermicro has many of the capabilities of an ODM such as custom design and manufacturing, we’re not really an ODM. Instead of making a custom design for hyperscalers, for example, we use a unique Building Block design strategy which provides many options to combine chassis, motherboards, and I/O expansion panels to create semi-customized designs for large scale customers. This provides all of the benefits of a unique design but is faster to market and is without the added cost of custom-designed boards and tooling.
Supermicro also provides all of the compatibility testing, ISV and OS certification, and support services that OEMs provide. Because of this customization capability through the Building Block strategy, Supermicro differentiates itself both from ODMs and traditional OEMs.
Blocks & Files: What channels exist for Supermicro’s storage servers, such as OEM, distribution, retail or direct sales, and what proportion of business goes through each?
Wendell Wenjen: Supermicro has different sales models for storage servers and our datacenter products, which include all of these channels. We do not disclose sales by channel but will note that we have substantial direct business, channel business and OEM fulfillment.
For customers with an appliance delivery model, we will OEM our storage and server products and provide integration services. We sell in the US and internationally through a network of distributors and value-added resellers. We also operate a direct-to-consumer web sales business (eStore) and sell through retailers. We have a direct sales force which is focused on CSPs, OEM business and selling to large enterprises internationally.
Blocks & Files: What’s Supermicro’s view of the needs for NVMe, Fibre Channel, and SAS infrastructure in its products?
Wendell Wenjen: NVMe-over-Fabric and the variations including RDMA, RoCE, and GPUDirect storage are increasingly becoming important as storage system performance and media access performance increases. We support these protocols through both compatible NICs and through ISV software and work with key technology providers such as Nvidia, Broadcom, and many ISVs to plan, validate, and deliver these technologies. As a systems and solutions provider, Supermicro takes the responsibility for testing, integrating and supporting these technologies.
Fibre Channel as a storage network has been declining for some years due to high cost, specialized administrative expertise and has been largely replaced with Ethernet except in certain legacy environments. We do support Fibre Channel HBAs, but it is not a large part of our network business.
SAS continues to be the most popular interface for disks and JBODs which we see continuing in the future with the STA roadmap. We continue to support SAS in our disk-based products through on-board or add-in card controllers and resell HDDs and some SSDs with SAS interfaces from leading suppliers.
Blocks & Files: Does Supermicro supply storage server software as well as hardware?
Wendell Wenjen: Supermicro has reseller relationships with many storage software partners including but not limited to WEKA, VAST, DDN, Hammerspace, Cloudian, Quantum, Qumulo, OSNexus, GRAID, Xinnor, Scality, Minio. In addition, we resell OS and other software from Red Hat, SUSE, Canonical, Oracle, Cloudera, and a number of other companies.
Supermicro also works with many other software companies through meet-in-the-channel business models, working together with resellers and integrators. Because of the availability, the large number of designs and options, and the reliability of our systems, we often meet new (to us) storage software companies that have been using Supermicro storage systems for many years through the channel without working directly with us. We are working directly with more of these companies.
Unlike some OEMs that have a portfolio of acquired software solutions, Supermicro is free to work with all of the leading ISVs and ensure that our customers are receiving the best solution for their needs.
Supermicro validates specific hardware configurations with the ISV’s software and can factory install and test the software and perform on-site installation, cabling and testing if requested by the customer. Large scale customers ordering many racks of servers and storage often find these services to be very critical in enabling fast on-site bring-up. This type of service where Supermicro designs the rack architecture, installs the software and networking in the factory and stress tests it, and then sends a team responsible for on-site cabling, installation and bring-up is an area which is unique to Supermicro from most of our competitors.
Blocks & Files: What do customers need to consider when enhancing their storage systems?
Wendell Wenjen: In our experience, customers often look at these criteria in either expanding their storage capacity or adding a new storage system as a separately managed system:
Type of storage access – Block, File, or Object, or some combination; how these access methods work with existing or new applications consuming the storage capacity.
Performance Requirements – in IOPS and throughput, performance metrics and benchmark tests, storage traffic transfer pattern (size, read vs writes, random vs sequential). How to meet the performance requirements (hardware – Flash vs. Flash + Disk + mostly disk), storage networking performance. Note that Supermicro will work with customers to fully performance test a clustered storage system which enables our customers to optimize the storage design during the design phase and have confidence in the delivered system’s performance.
Cost – Total acquisition cost, cost per TB, lifecycle cost.
Greenfield vs Brownfield, compatibility with existing systems, expansion versus new system build.
HPC and AI parallel file system storage supplier VDURA has added a high-capacity all-flash storage node to its V5000 hardware architecture platform.
The V5000 was introduced just over three months ago and featured central slim (1RU) director nodes controlling hybrid flash+disk storage nodes. These storage nodes consisted of a 1RU server with a 4RU JBOD. The overall system runs the VDURA Data Platform (VDP) v11 storage operating system with its PFS parallel file system. The new all-flash F Node is a 1RU server chassis containing up to 12 x 128 TB NVMe QLC SSDs providing 1.536 PB of raw capacity.
Ken Claffey
VDURA CEO Ken Claffey stated: “AI workloads demand sustained high performance and unwavering reliability. That’s why we’ve engineered the V5000 to not just hit top speeds, but to sustain them—even in the face of hardware failures.”
VDURA says “the system delivers GPU-saturating throughput while ensuring the durability and availability of data for 24x7x365 operating conditions.”
An F Node is powered by an MD EPYC 9005 Series CPU with 384GB of memory. There are NVIDIA ConnectX-7 Ethernet SmartNICs for low latency data transfer, plus three PCIe and one OCP Gen 5 slots for high-speed front-end and back-end expansion connectivity. An F-Node system can grow “from a few nodes to thousands with zero downtime.”
A combined V5000 system with all-flash F Nodes and hybrid flash+disk nodes delivers, VDURA says, a unified, high-performance data infrastructure supporting every stage of the AI pipeline from model training to inference and long-term retention. The VDP use of client-side erasure coding lightens the V5000’s compute overhead, with VDURA claiming VDP eliminates “bottlenecks caused by high-frequency checkpointing.”
The minimum F-Node configuration is three Director nodes and three Flash nodes. Both can be scaled independently to meet performance and/or capacity requirements. A 42U rack can accommodate three director nodes and 39 Flash nodes: 59.9 PB of raw capacity.
Customers can deploy a mix of V5000 Hybrid and All-Flash (F Node) storage within the same namespace or configure them separately as discrete namespaces, depending on their workload requirements.
Nvidia Cloud Partner Radium is implementing a V5000-based GPU cloud system with full-bandwidth data access for H100 and GH200 GPUs and modular scaling – it says this means storage can grow in sync with AI compute needs, “eliminating overprovisioning.”
The VDURA V5000 All-Flash Appliance is available now for customer evaluation and qualification, with early deployments underway in AI data centers. General availability is planned for later this year, with RDMA & GPU Direct optimizations also planned for rollout in 2025.
PARTNER CONTENT: As IT leaders move away from VMware, they face a critical decision: do they stick with traditional storage architectures, or is now the time to finally unlock the full potential of an infrastructure that converges virtualization, storage, and networking technologies?
Early convergence efforts centered on hyperconverged infrastructure (HCI), where storage ran as a virtual machine under the hypervisor, commonly called a vSAN. While adoption has lagged behind traditional three-tier architectures, recent advancements have significantly improved vSAN, making it worth reconsidering by addressing past shortcomings.
When vSANs were first introduced, they came with bold promises:
– Lower storage cost utilizing server hardware: By eliminating the need for expensive proprietary storage arrays, vSAN aimed to reduce capital expenses by running on standard x86 servers already used by the hypervisor.
– Automatic scale with each node addition: Every time an organization adds a node to the cluster, it expands both the compute resources for the hypervisor and the storage resources for the vSAN, creating a straightforward path for seamless growth.
– Virtualization awareness for higher efficiency: Because vSAN was running as a VM of the hypervisor, it promised optimized data placement and workload efficiency, ensuring storage performed per virtualization needs.
– Simplified Storage Management: By eliminating dedicated storage infrastructure, vSAN solutions aimed to reduce administrative overhead and remove the need for complex LUN or volume management.
– High-performance, comparable to dedicated storage arrays: Vendors promised that vSAN could match or exceed the performance of all-flash arrays while providing greater flexibility and scalability.
While these promises were appealing, the real-world implementation of vSAN has not met these expectations. The high licensing costs, resource inefficiencies, and complex failure scenarios have made IT teams question whether vSAN aligns with modern infrastructure goals.
The challenges facing vSAN technologies
As IT teams explore alternatives to VMware, they often reconsider vSANs or hyperconverged infrastructure. However, traditional implementations introduce several key limitations:
– Performance bottlenecks: Traditional vSAN architectures struggle with latency-sensitive workloads, particularly when managing high IOPS applications across multiple nodes.
– Storage overhead: Many existing vSAN architectures require excessive CPU and memory resources, impacting overall workload efficiency.
– Scalability constraints: vSAN solutions often scale in fixed increments, requiring simultaneous addition of compute and storage rather than allowing independent scaling.
– Failure recovery and data resilience: Conventional vSAN implementations, like dedicated storage arrays, require long rebuild times after drive failures, impacting business continuity.
– Vendor lock-in: Most vSAN architectures, similar to dedicated storage arrays, lock customers into specific hardware configurations, limiting flexibility in multi-cloud and hybrid deployments.
For many IT teams, these limitations have led them to stick with traditional three-tier storage architectures. However, recent vSAN advancements are changing the equation.
Modern innovations in vSAN technology
The core principles of vSAN – aggregating local storage, distributing data across nodes, and enforcing redundancy – have remained the same for years. However, recent advancements have significantly improved performance, resilience, and efficiency.
1. High performance that rivals dedicated all-flash arrays Traditional vSAN implementations run storage as a VM on the hypervisor, creating unnecessary and inefficient I/O hops:
– The application VM sends an I/O request to the hypervisor. – The hypervisor forwards the request to the storage VM. – The storage VM processes the request and communicates with physical storage. – The storage VM sends the result back to the hypervisor. – The hypervisor finally sends the result back to the application VM.
Each step introduces latency and consumes additional CPU resources.
Modern converged solutions integrate storage (the vSAN) directly with the hypervisor and networking, creating a single code base for a data center operating system. This technique, known as Ultraconverged Infrastructure (UCI), eliminates redundant I/O paths and drastically reduces storage latency.
Additionally, most legacy vSANs rely on iSCSI or NFS for internode communication, which introduces overhead. Modern vSANs leverage custom protocols optimized for ultra-converged environments, improving data transfer efficiency, eliminating bottlenecks, and improving scalability.
The result? Performance comparable to high-end all-flash arrays – but with the cost advantages and flexibility part of the original vSAN promise.
2. True storage and compute disaggregation Older vSAN solutions force IT teams to scale computing and storage together. However, newer disaggregated architectures allow organizations to:
– Scale storage independently of compute nodes. – Add GPUs without affecting storage expansion. – Optimize storage for AI, virtualization, and analytics workloads.
By separating computing from storage, IT can match the specific needs of their organization and infrastructure growth to actual demand rather than being forced into inefficient upgrades.
3. Zero-impact failure recovery Legacy vSAN solutions often require lengthy RAID rebuilds, significantly impacting system performance. In contrast, modern vSAN architectures leverage distributed, self-healing data placement, allowing storage workloads to rebalance without degrading overall performance.
Instead of requiring complex RAID rebuilds or erasure coding reconstructions, these next-gen vSAN solutions offer inline recovery mechanisms, reducing recovery time from hours to minutes – a significant upgrade over traditional approaches and can enable you to also rethink using backup for storage failure.
4. Inline storage optimization and real-time data management Newer vSAN solutions implement inline data deduplication, reducing overhead while improving storage efficiency. Unlike legacy vSANs that deduplicate after data is written, some modern solutions perform these optimizations during ingest, reducing I/O amplification and improving write performance.
Additionally, advanced vSANs support multiple storage tiers and allow live migration of VMs between those tiers without downtime. This capability ensures that high-priority workloads remain on the most performant storage, while less critical workloads can reside on cost-effective storage tiers, maximizing resource utilization.
Is it time to reconsider vSAN in a VMware alternative strategy?
As organizations migrate away from VMware, they should reassess their storage architectures. Opting for a VMware alternative isn’t just about replacing the hypervisor – it’s an opportunity to evaluate whether a modern vSAN implementation can deliver higher performance, better resilience, and greater flexibility than dedicated storage arrays.
Not all vSAN implementations are created equal. Some still suffer from legacy limitations, while others offer breakthrough innovations that eliminate storage inefficiencies.
To make an informed decision and get the full comparison of vSAN alternatives download our white paper comparing VMware vSAN, Nutanix AOS Storage and VergeIO VergeOS.
Open source object code developer MinIO has appointed its first chief business and marketing officers as AI training and inferencing take hold of enterprise customers.
MinIO didn’t need the classic marketing and business integration functions because it was developer-led, with more than 2 billion Docker pulls and over 50,000 GitHub stars. Its rise since it was founded in 2014 has been led by word of mouth and it provided intense developer support and education, at one time having three simultaneous CTOs. But AI is changing all that, accelerating its penetration of enterprises and requiring what we might call a more gr49own-up stance towards business customers and their CIO and line-of-business IT buyers.
One of the initial signs of this was the November launch of its AIStor product, which encourages the use of object stores to feed AI training and inferencing with a GPUDirect-like facility for S3-stored data. This is bringing it into more direct competition with other AI object data suppliers who talk to business customers via a focused marketing operation and have their internal business-related functions integrated to provide cost-effective, coordinated, and efficient internal operations. This helps bring relevant products to informed customers through competent and skilled sales channels with responsive sales and billing and supply operations.
Another sign of this need has been larger orders. MinIO recorded multiple eight-figure, exabyte-scale customer deals in 2024, with record ARR growth. Garima Kapoor, co-CEO and co-founder, stated: “We are at a pivotal moment in the AI revolution where data infrastructure must support exascale demands that the industry has never seen before.” She sees potential for “aggressive business growth for MinIO in the future” and the company needs a stronger go-to-market function.
From left, Mahesh Patel and Erik Frieberg
So MinIO has appointed Mahesh Patel as its first chief business officer, with Patel having previously been COO at SaaS data protection business Druva, which he joined as CFO in September 2014. It has also hired Erik Frieberg as its first chief marketing officer, coming from stints at Pantheon Platform, Tetrate, solo.io, and Puppet.
MinIO says it has increasing momentum, a strengthened GTM leadership team, “rampant innovation that is raising the bar for AI storage infrastructure, and key collaborations with top AI ecosystem players.” Patel and Frieberg will help it convince enterprises to use MinIO object storage for AI data needs rather than offerings from Cloudian, DataCore, Dell, IBM, HPE, Hitachi Vantara, NetApp, Pure, Scality, StorONE, and VAST Data.
MinIO claims it is “the world’s fastest growing object store” and does not want to relinquish that title to object storage competitors with stronger go-to-market operations for AI buyers.
