Interview. Blocks and Files is a storage-focussed outlet that has progressed, with the industry, into data access, management and AI pipeline feeding. We know, as does everyone else, that organizations need a storage team as soon as the storage infrastructure grows beyond a certain point. The storage team provides stored data access and protection to applications that need it.
Analytics and AI is making data access far, far more complex. An analytics function needs its data located, filtered and selected, extracted, transformed, loaded and then processed with appropriately coded functions, SQL jobs for example. AI has only made this even more complicated with its initial reliance on files for training evolving to object data and retrieval-augmented generation for AI inferencing, vector databases and semantic search.
Data engineering and science teams have sprung up and they use different data concepts than storage teams. File and object data management supplier Komprise sits above the storage infrastructure and is involved in making stored data available and accessible to data teams. It says the two team types have dissonant concepts and need to talk.
We explored its ideas in an interview with Komprise President and COO Krishna Subramanian.
B&F: Isn’t it the case that, historically, customer organizations have always had a storage admin function responsible for ensuring the storage infrastructure operates properly and efficiently to store and provide access to the data the organization’s IT processes and users require? Do you think that a separate data platform team function has come into being? Why, and what are its responsibilities?
Krishna Subramanian.
Krishna Subramanian: Yes, the storage team delivers the infrastructure to store and provide access to the data. But, “access to data” is a loaded word – and increasingly with AI, storage teams are simply ensuring the data can be accessed from the device whereas data engineering teams are the ones servicing AI teams, providing them with the data needed for their use cases. We believe storage teams can elevate their role in “providing access to the data” and partner more closely with data engineering and AI teams because at the scale of unstructured data, you cannot have data engineering teams who lack broad access to the data nor the right tools to organize the data for each use case. Storage teams who do have an organization-wide view of data should systematically provide approaches to rapidly classify and find the right data.
Also, as AI goes mainstream, you need a systematic, automated way for any user to profile, classify and pick the right data with proper data governance, which underscores the need for storage teams to elevate their roles. Data storage teams are evolving to become data services providers and this will require a closer relationship with the teams focused on how to derive greater value from the data that they manage and protect.
B&F: What are the advantages and disadvantages of having two separate teams? Why can’t one team do both jobs?
Krishna Subramanian: Historically, storage infrastructure teams have focused on delivering the infrastructure while data engineering teams have focused on data quality, data cleansing and serving the needs of data analysts and data scientists. So, storage teams focused on the technology and the use of data while data engineering focused on what’s in the data and how to get insights and value from it. This makes sense as manipulating structured data requires deep SQL expertise, which is the purview of data engineering teams and a different skill set than data storage management.
AI necessitates some tuning of this approach for two reasons: a) because AI relies on unstructured data which lacks a unifying schema and is not stored in SQL databases and spreadsheets, and b) AI is going to be used by everyone in the enterprise so storage teams who have access to and management responsibilities for all the data should play a bigger role in providing tools and processes to classify the data, find the right data sets, tag sensitive data, etc. Essentially, the two separate teams are still needed but greater collaboration between the two is necessary due to unstructured data and AI. Ultimately a new set of SLAs and KPIs will emerge and CIOs may create new org structures to address these emerging use cases and requirements.
B&F: Is the data team’s area of responsibility restricted to AI? If not, where does it stop? Does it represent the needs of all data access users and processes to the storage team?
Krishna Subramanian: Data engineering teams were traditionally focused on data analytics processing and are now expanding their role to delivering AI services. Many organizations also have data scientists, departmental IT, and other roles that help users get access to data. The dramatic shift that AI brings is the democratization of data re-use. While historically, analytics teams were the ones reusing data, now every user may need to feed the data they can access to AI. So, getting the right data to the right AI use case should be standardized for every user with automated data governance.
B&F: Will storage teams need to evolve from infrastructure maintenance into data service providers to help data analytics teams as they gather the requirements and design processes for AI. What skills will they need for this?
Krishna Subramanian: Yes absolutely. Storage teams should think about how they work with departments to profile and enrich metadata to make it easier to curate data for projects. Learning how to provision and manage GPU-ready infrastructure and the data lifecycle to balance cost, performance and security is another skillset. At the same time, they will need to enforce governance to safeguard sensitive information, ensuring compliance with regulations and preventing leaks into commercial AI models. So the storage admin/engineer role is changing, and becoming much broader vis a vis AI, data preparation and data governance.
B&F: Will data platform teams need to build common metadata definitions with storage teams and collaborate on an inclusive data governance strategy incorporating unstructured data?
Krishna Subramanian: Yes. Data teams, storage teams, security and compliance teams and data owners should collaborate on common metadata definitions, sensitive data labels and a data governance strategy.
B&F: Won’t each team have its own metadata? Can you provide an example of common metadata definitions?
Krishna Subramanian: Yes, there will be both. Common metadata definitions include things like sensitive data labels aka PII, IP and employee IDs, author information, project codes or grant numbers. SIDs (security-identifier) can be very useful in M&A scenarios and to enforce compliance. Storage teams can also help with extraction of latent metadata that is outside of the filesystem such as header metadata and application-generated metadata.
B&F: Are you seeing a role for Komprise as an interface layer or function between the storage and data teams, helping represent each to the other? How would that work?
Krishna Subramanian: Absolutely. Komprise provides a single place to search across all enterprise unstructured data and it provides ways to continually enrich that data. Komprise has role-based access so data owners and data engineers can view the data they have access to and do things like tag and search data with Komprise Deep Analytics. Storage administrators can move data, govern data usage and execute data workflows systematically through Komprise Smart Data Workflows. Komprise provides a common interface through which data teams and storage teams can interact.
Komprise diagram.
B&F: Do you have any customer examples of customers having two separate teams? How do they work?
Krishna Subramanian: One of the world’s largest cancer research hospitals uses Komprise to achieve collaboration across their data teams and storage teams. The data teams use Komprise to tag data associated with each research project. When a project is complete, files are automatically tiered based on policies the storage administrator set in Komprise. This customer has saved millions of dollars using this approach of collaborating effectively between departments and storage IT via Komprise.
Another example is an oil and gas customer that recently went through a divestiture. The company’s compliance team used Komprise to segregate the data for each entity based on the compliance teams selecting the SIDs they wanted mapped to each entity and storage IT executing the data migrations via Komprise.
Research house GigaOm’s 2025 Object Storage Radar report identifies 22 suppliers, up from last year’s 18, with some dramatic changes in their status as the market shifts towards products with generic market appeal.
Whit Walters.
Analyst Whit Walters says; “The key characteristics of enterprise object stores have changed, with more attention paid to performance, ease of deployment, security, federation capabilities, and multi tenancy.” This is against a background of the volume and variety of data continuing to grow exponentially.
He identifies 10 leaders in this 6th edition of the report, which compares to 8 last year; DDN and Hitachi Vantara having been promoted into the leaders’ category from last year’s Challengers’ status. Walters sees a market where the suppliers are increasingly mature and focussed on building general platform-focus products rather than niche market area offerings.
Here is the 2025 chart with the 2024 chart included for reference;
We can see several suppliers have crossed the chart from 2024 to 2025; IBM moves from the Mature-Platform quadrant to the Mature-Feature Play quadrant. WEKA moves from the Innovation-Feature Play to the Mature-Feature Play quadrant. Hitachi Vantara changes from last year’s Mature-Patform Play challenger to this year’s Innovative-Platform Play leader. All of last year’s entrants become challengers this year: Quantum, Quobyte, SoftIron and Zadara. There are also four new suppliers: Cohesity, Seagate, SoftIron and Spectra Logic.
As there is no distance-from-the-center number we have ranked the leaders in distance-from the center terms visually: Cloudian, Pure Storage, Dell, Scality, NetApp, MinIO, Hitachi Vantara, VAST Data, WEKA, and DDN in that order.
To produce the chart, suppliers are rated, with a zero to five score, on three dimensions: key features, emerging features, and business criteria. They each have sub-categories. These ratings contribute to a supplier’s position in the Radar diagram which “plots vendor solutions across a series of concentric rings with those set closer to the center judged to be of higher overall value. The chart characterizes each vendor on two axes—balancing Maturity versus Innovation and Feature Play versus Platform Play—while providing an arrowhead that projects each solution’s evolution over the coming 12 to 18 months.” The arrowheads indicate forward (normal), fast (quicker) and out-performers.
Walters says about the radar chart positioning: “Key features, emerging features, and business criteria are scored and weighted. Key features and business criteria receive the highest weighting and have the most impact on vendor positioning on the Radar graphic. Emerging features receive a lower weighting and have a lower impact on vendor positioning on the Radar graphic. The resulting chart is a forward-looking perspective on all the vendors in this report, based on their products’ technical capabilities and roadmaps.”
We charted each supplier’s rating for the three dimensions, ranked by key features score;
The lack of a green bar indicates a zero emerging features score. Get a full-size image here.
We should note that: “The Radar is technology-focused, and business considerations such as vendor market share, customer share, spend, recency or longevity in the market, and so on are not considered in our evaluations. As such, these factors do not impact scoring and positioning on the Radar graphic.”
Walters points out: “The majority of evaluated vendors are positioned within the Platform Play hemisphere (with an emphasis on the Maturity quadrant), while a few exceptions appear on the Feature Play side. This distribution indicates that most vendors prioritize stability and reliable user experience while offering comprehensive platform solutions.”
He adds: “Vendors are expanding their solutions to address a wider range of requirements, integrating functionalities beyond core object storage. This expansion reflects the demand for unified storage platforms that simplify data management across diverse workloads.”
Three vendors did not respond to Walters’ inquiries and their ratings are based on desk research, meaning their published documentation and webpages. These vendors were Cohesity (SmartFiles), IBM (Cloud Object Storage), and Spectra Logic (BlackPearl S3 Hybrid Object Storage).
Hitachi Vantara has kindly made the whole Radar report available via its website.
Quantum, which has delayed its fourth quarter and full fiscal 2025 results because of accounting problems, now says that the previous quarter’s revenues were wrongly stated as a 1 percent Y/Y rise to $72.6 million. It should have been a 4.5 percent fall to $68.7 million.
This is revealed in an August 8, 2025, SEC filing by CFO Lewis Moorehead who stated that the existing Q3 FY 2025 financial statements “should no longer be relied upon.” Quantum’s audit committee and management team found that there were “service and subscription revenue inconsistencies” in the way revenue was recognized as well as necessary standalone selling price adjustments that needed to be made, “resulting in an adjustment to revenue.”
Quantum has long suffered standalone selling price problems, it now appearing a seemingly endemic issue.
The company expects that the resulting restatement: “will result in a decrease of approximately $3.9 million in revenue and a similar decrease in net loss from operations,” which could increase the GAAP net loss for the period from the existing $71.4 million; possibly, we think, to $75.3 million if the $3.9 million revenue decrease is reflected 1 for 1 in the net loss number.
The company has, it says, deficiencies or material weaknesses in its internal controls and: “the Company’s full assessment of the effectiveness of its internal control over financial reporting will be described in more detail in the Annual Report on Form 10-K for the year ended March 31, 2025.” There is no date set for this report to be filed with the SEC.
The SEC filing states that there were also adjustments to the fiscal quarters ended June 30, 2024 and September 30, 2024. However, these adjustments were not material and are not required to be restated.
Grant Thornton LLP is the Company’s independent registered accounting firm and was appointed on August 18, 2023, for the fiscal year ending March 31, 2024 (FY 2024).
Comment
Quantum has had a board makeover and CEO/CFO and CRO change in the last three months, and a round of layoffs, as it tries to recover from years of losses and declining revenues. The new board and exec team think the company can grow its revenues and return to health.
Investors sold shares when they found out about the latest snafu, with the peak August 8 stock price of $8.02 falling to the current $7.42 after recovering from a $7.02 low point on August 12. The investors may be thinking the board and exec team’s turnaround plan is not credible until, and unless, it can demonstrate believable and transparent financial reporting.
This latest financial reporting blunder makes the new board and exec team’s job even more difficult. The reputation of the company will fall lower and its debtors may have to reassess Quantum’s value. Perhaps, they may think, the firm has more total value if component parts are sold off and it exits from unprofitable business sectors. The independent accounting firm may now be facing an exit as well.
Louhi, Mistress of the North, attacking Väinämöinen in the form of a giant eagle with her troops on her back when she was trying to steal Sampo; in the Finnish epic poetry Kalevala by Elias Lönnrot. (The Defense of the Sampo, Akseli Gallen-Kallela, 1896)
Veeam’s Coveware business unit has released its Q2 ransomware report showing average ransomware payments are double last quarter’s amount at $1.13 million.
Data protector Veeam bought the Coveware business, with its cyber-extortion incident response facilities, in April last year. The quarterly report is a snapshot of what it’s seeing, based on “firsthand data, expert insights, and analysis from the ransomware and cyber extortion cases that they manage each quarter.” And it is seeing business digital data theft increasing with socially-engineered entry points the main attack vector.
Coveware CEO Bil Siegel stated: “The second quarter of 2025 marks a turning point in ransomware, as targeted social engineering and data exfiltration have become the dominant playbook. Attackers aren’t just after your backups – they’re after your people, your processes, and your data’s reputation. Organizations must prioritize employee awareness, harden identity controls, and treat data exfiltration as an urgent risk, not an afterthought.”
Bill Siegel at Veeam event.
The report says social engineered attacks are now the biggest threat with three ransomware groups prominent in the quarter: Scattered Spider, Silent Ransom, and Shiny Hunters. They are “using novel impersonation tactics against help desks, employees, and third-party service providers. … Credential compromise, phishing, and exploitation of remote services continue to dominate initial access, with attackers increasingly bypassing technical controls via social engineering.”
The top three top ransomware variants were Akira (19 percent), Qilin (13 percent), and Lone Wolf (9 percent), with Silent Ransom (5 percent) and Shiny Hunters (5 percent) entering the top five for the first time.
This quarter’s report notes: “The Average Ransom Payment: $1,130,070 (+104 percent from Q1 2025) and Median Ransom Payment: $400,000 (+100 percent from Q1 2025), jumped substantially in Q2 2025 versus the prior quarter. We attribute this increase to an increase in payments by larger organizations impacted by data-exfiltration-only incidents. While the quarterly increase is dramatic, we note that similar jumps quarter to quarter have occurred in the past and do not yet believe these metrics to be an inception of a trend.”
We are told that data theft has overtaken encryption as the primary extortion method. “Exfiltration played a role in 74 percent of all cases, with many campaigns now focusing on data theft rather than traditional system encryption.”
Access defences can be weak: “Credential-based intrusions dominate, with groups like Akira regularly exploiting exposed VPNs and remote services using stolen or weak credentials, often sourced from infostealers or successful phishing campaigns. Social engineering also continues to mature, with actors leveraging trusted communication channels like Microsoft Teams for vishing, SEO poisoning to deliver malware, and deceptive scripts masked behind fake security prompts or CAPTCHAs.”
And: “These tactics bypass technical controls by targeting human behavior, a trend exemplified by groups like Scattered Spider, whose tailored impersonation techniques make help desks the front line of compromise.”
The report says: “The percentage of organizations that opted to pay a ransom regardless of impact remained relatively low at 26 percent. We are encouraged that the overall rate of payment has not shown regression over the prior quarters. As compared to years past, companies are generally better prepared to defend themselves against extortion attacks, and are getting better prepared at navigating the nuances of cyber incidents via IR preparedness.”
However payment rates for data exfiltration are higher, with the report noting: “The payment rate on data exfiltration only matters increased in Q2 and remains in a stubbornly high bracket. Some threat actors are increasingly focusing on data exfiltration only as they feel the effort-impact / payout economics are more favorable to the encryption attacks. Encryption attacks do still cause the most impact and urgency though.”
The top three victim organization types were:
Professional services (19.7 percent)
Healthcare (13.7 percent)
Consumer services (13.7 percent)
Victim organization size is a factor: “Ransomware attacks most commonly affect small to mid-sized organizations, with companies ranging from 11 to 1,000 employees making up a combined 64 percent of incidents. This suggests that attackers often target firms that are large enough to offer a potential payout but may lack the robust cybersecurity infrastructure of larger enterprises. Mid-sized organizations (1,001 to 10,000 employees) account for 17 percent of attacks, showing that as companies grow, they remain attractive targets.
“Interestingly, very large enterprises—with over 25,000 employees—make up just 8 percent of incidents, indicating that scale may offer better protection through more mature security programs. At the smallest end, companies with fewer than 10 employees represent only 4%, likely due to limited assets or lower visibility. Overall, the data highlights a ransomware “sweet spot” in the small to mid-sized range, where vulnerabilities are more common and defenses often underfunded.”
There is much more information in the report which is a sobering read. Cut and paste a copy from the website into a file and go read it in a quiet room. If you are in a small-to-mid-sized organization then consider yourself at risk, and realize that your online employees and partner staff represent the main attack surface. They need constant social engineering attack awareness.
You can read a Bill Siegel blog to get more background info.
Interview.CyberSense is an Index Engines’ malware hunting and detection technology that it sells through an OEM channel with Dell (Cyber Vault), Hitachi Vantara, IBM and Infinidat. But resellers and SIs could sell a CyberSense service that protects general primary and secondary storage, including backup vaults.
We talked with Jim McGann, the company’s VP of strategic partnerships to find out more.
B&F: I’m impressed by Index Engines because you’ve managed to knock off these really impressive OEMs and they’re all basically talking about your message.
Jim McGann.
JimMcGann: The thing with the OEMs is that what we’re doing is very difficult and challenging and it’s very difficult for them to do it. So they recognise the technology. It’s something that would be challenging for them to get to the level of resiliency that we provide. So I mean, all the vendors are, and if you look at the magic quadrant for Gartner and Primary Storage, the ones that are kind of moving over to the left, outside of the upper right are the ones that we’ve partnered with. So they need more technology to stay in that upper right quadrant. It’s all about cyber resiliency and that’s what we provide.
And you can see, if you look at that magic quadrant, the ones that we haven’t partnered with yet are the ones that are still in the upper right. So they’re sitting in a happy place as they shift. And I think if you talk to the folks that run the magic quadrant for primary storage, this is one of the five pivotal SLAs that vendors need to provide. And the folks at Gartner are smart enough to know that some of the things that the vendors have done are just not good enough. Because you read the newspapers and you see Marks & Spencer and you see people like that that are still having attacks.
B&F: You’ve gone for storage system vendors like Dell, Hitachi Vantara, IBM, and Infinidat. Do you have any data protection vendors as customers for your CyberSense technology?
Jim McGann: So in the Cyber Recovery Vault with Dell, the data that’s in that vault is backup data. We support the Dell backup formats. So PowerProtect data manager, Avamar Networker. We also support Commvault, NetBackup and the IBM backup software. So the vault is a good thing and backup is a 24 hour view of the world, which is good, but we’re moving into production, which is every X minute view of the world, which is better. But not to say that the vault isn’t important because, when they wipe out an array, which some of these variants do, you need to have a gold copy in the vault that’s clean for recovery.
B&F: I’ve not noticed that Rubrik or Cohesity or Commvault are actually using your technology.
Jim McGann: No, they’re not. I mean the modern backup format vendors feel they have it covered. I mean Rubrik is now a security vendor versus a data protection vendor. So they feel they have cyber resiliency covered. That’s questionable from our point of view, but for them, this is the capability that they want to provide as a security player. We’ve had conversations with them and there are conversations going with some of them today. They’re not as interesting as the ones with people more in the production storage side.
There’s still a lot of vendors there. There’s a lot of other vendors that we’re talking to in production storage. I think what we want CyberSense to be is really a data service. That’s how Gartner looks at us; as a data service that overlays on top of both primary and secondary storage.
I think what customers want is to have a unified data integrity platform from primary to secondary storage that says, “Hey, the data’s good.” So if I’m a Marks & Spencer or a Co-op that gets attacked, it is not like a panic mode. It’s like I know where I have clean data and I know how to recover and I can bring the business back to an operational state versus people running around like crazy saying we don’t know how to recover.
I think everybody talks about these cyber resiliency strategies, but they’re clearly not robust enough. And you would think someone like Marks & Spencer would’ve a better look, right? I think taking orders is important for a company like them.
B&F: Infinidat is changing its status.
Jim McGann: Yes. Infinidat will be Lenovo over the next couple months and I think Lenovo wants to get in the enterprise. I don’t know their strategy, but I’m assuming they want to get into the enterprise space. Cyber resiliency is a key aspect to that. So I think they will embrace us very well.
B&F: What about other OEMs?
Jim McGann: If you look at the magic quadrant for primary storage, the folks that are hanging in the upper right are the ones that we haven’t cracked yet because they’re living in a world where they think they have good enough. Whereas, if Gartner starts saying, “Hey, not quite good enough,” and you’re starting to slide down and left, then you’re going to see the doors will open up.
But I think as CyberSense becomes a standard across the enterprise for data integrity and a key component for cyber resiliency, they will come to us because customers don’t just have Dell or don’t just have IBM or don’t have Hitachi, they have mixed environments where they say: “We are using CyberSense here. Why can’t I get it over on my NetApp? And I think you’ll see what we’re doing as a company is; we’ve been a very OEM-focused company where we partner aggressively and that’s worked out very well. But we’re heading into a direction of more providing an independent offering that vendors can integrate to.
So opening up and going to these vendors and providing toolkits that allow them to connect to other environments. So that opens us up to even the security vendors, … the managed service providers, a lot of different folks, the cloud vendors opens us up to those different spaces as well.
B&F: Let’s say CyberSense is a service and let’s say as you’ve indicated, I’m sure it’s absolutely true, a customer would say to themselves, well, I’ve got CyberSense on my Dell kit, but there’s my NetApp kit, there’s my cloud storage. Why can’t I have it layered across everything?
Jim McGann: Indeed.
B&F: So let’s suppose you wanted to layer it across everything. You’ve got this CyberSense capability and it has to integrate with whatever storage vault it’s going to look into. Now this is where I’m getting tricky. How do you do that integration? The OEM sale strategy suggests you have to do an integration per vendor per storage array. But that may not be true.
Jim McGann: It won’t be true in the second half of this year.
The OEM partnerships are strategic and we can create a very elegant orchestration with them. There are toolkits that are available that are being more formalised now that will allow for integration that will work, and work in production. And I think, if you see vendors like some that you have mentioned that we’ve not strategically partnered with, they may want to take it to the next level. They could say: “That’s good, but we can even be more sophisticated about that.”
I think there’s multi-phase approach that you’ll be seeing here. And the same on the security side too. So I think as we’re running and doing scans in production on a regular basis, if CyberSense sees unusual activity, data manipulation activity, that could be integrated to security systems to say: “Hey, there’s unusual activity. Shut this down.”
And I think if you study Marks & Spencer versus the Co-op; the Co-op saw unusual activity and they disconnected the Internet and they were able to recover quickly. Marks & Spencer didn’t do that. So it spread. I mean if you see unusual activity on a single server, being able to automatically sever that from the internet and isolate it is a very god thing. CyberSense can integrate into those security strategies very well and provide that level of telemetry data and knowledge.
But then also I know where the last good version of data is. So that’s the data that you can recover. And our goal as a company is to turn a ransomware attack into a normal disaster that can be easily recovered from. And so it just minimises the impact and the effect of it. And again, if customers are down for months or handicapped for months, like Marks & Spencer, that’s just devastating for them and they don’t need to be at the mercy of these bad actors.
B&F: If Marks & Spencer could six months ago have gone to Index Engines and said: “We want to buy a subscription to the CyberSense service to cover our storage,” you’d have said no at the time, whereas in a few months time, would that be possible?
Jim McGann: Yes.
B&F: Wow! Okay.
Jim McGann: A lot of what we’re seeing in the channel is that the channel sellers are really selling cyber resiliency. So they’re selling a strategy that more integrates with our message and how we sell this. And I think selling through the channel, with the channel sellers having relationships with Marks & Spencers and other, that they can help provide this. So if you think about cyber software, there could be connectors on there to all sorts of different platforms. So the idea is, if you are maybe even a managed service provider, that you can just say: “Hey, I want to connect to A, B, C, and D that that’s possible.” It is going to take us time, but we’re laying the foundation to do that. And that’s really the future.
I think the OEM relationships have allowed us to get into customers that we wouldn’t be able to get into on our own. But now that CybersSense has a brand reputation out there, customers are asking for it across other environments and I think that’s the need that we’re going to satisfy in the direction that we’re going to be taking.
B&F: The core product, the ransomware scanning, the malware scanning; you’ll keep that updated with every flesh flavour of malware you can find?
JimMcGann: That’s so. Even better than that, we just got a patent issued for an automated process. So in our colo offsite, we have a secure lab. Any new variants that are detected on the market through different sources are downloaded into that lab and detonated. We study that and then we make sure that our machine learning is up to date to support that.
It’s an automated process that we’ve just patented that will say basically, if there’s new variants, cyber central will find it. And variants are basically classified into about 30 or so different general categories. So they’re not reinventing the wheel every time. They’re changing the name, they’re changing the encryption algorithms. So we’re seeing them, we’re classifying and say, okay, we’re good. And if not, then the machine learning would need to be updated because we have a 99.99 percent SLA to find that corruption.
B&F: This is a core constant ongoing engineering development effort for CyberSense?
Jim McGann: There’s 800 to a thousand or more new variants every day. I mean there are just tweaks and modifications of existing variants. So the idea is if you’re going to play in a soccer match, you study the study the opponent and understand what they’re going to do. That’s what we’re doing. We’re studying the variance on a continual basis to know exactly what they’re doing. Most of the times they don’t change much, but if they do that, we’ll see that and customers can know that. They’ll have that confidence that any corruption with these new variants will be detected with 99.99 percent confidence
B&F: And then you’ll be adding connectors?
Jim McGann: it’s all about just being able to feed as much data to CyberSense as possible. CyberSense looks out as data changes over time. The idea to be agnostic of platforms is the vision that we have and I think that’s what customers want. They want a data integrity service that can look at data across the portfolio, across the data centre, and give them the assurance that their data is good and, if not, where’s the data I need to use for recovery.
B&F: With a storage array, a filer, an object storage or a block storage array with a database, there are APIs you can use to get into that storage array. If you can talk NFS, then you can talk NFS to any storage array that uses NFS. But, with a backup vendor’s vault, that’s not the case, because the only doorway into that vault is through the backup vendor’s own proprietary API. So you have to do that, I think, on a case by case basis.
Jim McGann: We’ve engineered access to these backup formats. Part of our intellectual property here at Index Engines is understanding complex formats so that we can look at something like NetBackup and understand it. If those partners or backup vendors cooperate with us and give us the schema, that helps. But we have engineers that can really understand and look at the bits and bytes of these proprietary formats and engineer access into them. So that’s what we’re doing; cracking them open without rehydrating them, and looking inside of them.
That is key to the Dell Cyber recovery vault because it’s all backup data that’s in there. Also key to snapshots is to be able to understand a snapshot and how to mount it to block storage. So it sounds easy saying that we scan data, but, as you understand, the devil’s in the details here. How the formats work. And a lot of these vendors like IBM or Infinidat; they participate with us to create, to orchestrate into those APIs.
With something like NetApp, they have things like SnapDif and they have APIs to be able to help you interpret their data. So we’re leveraging anything we can just to understand the data and how CyberSense looks at it to see how it changes over time. So for example, with a Oracle database or a health records database, we need to see either a snapshot of that or a backup of it so we can see a static copy of it and compare it to the previous version.
We’re not going to be scanning production databases because customers don’t want that. So we’re looking at, for example, health records databases like Epic, and snapshot it every 15 minutes. CyberSense can scan it every 15 minutes. And when there is an attack on that database, then CybersSense will say: “Hey, the one from the previous 15 minutes is clean. You can go back to that with confidence.” That turns a recovery into minutes versus trying to say which one should we recover? Is that clean or not? And they have to go to a clean room and test it. CyberSense is giving you that confidence on a continual basis that you don’t need to study it or go into a clean room. We were telling you that it’s good and you can recover with confidence.
B&F: Is there anything else you care for me to know about what CyberSense may be doing over the next three to six months?
Jim McGann: I think the future is definitely maintaining these strategic relationships because they’re important to us, but also opening up so that cybers sense can easily integrate into other environments and really that’s what customers are asking for and we’re providing that solution out to the market.
Bootnote
We think Index Engines has just signed up Panzura as an OEM.
Cyber-security and backup company Rubrik has an Agent Rewind offering that enables customers to undo mistaken agentic AI changes to applications and data by providing visibility into agents’ actions and an undo capability.
The technology behind it comes from Rubrik’s acquisition of AI agent development startup Predibase in June. Agentic AI systems will be able to independently execute tasks, modify data, or alter IT configurations to achieve their goals, potentially introducing risks to system stability, data integrity, and compliance. Johnny Yu, Research Manager at IDC, is quoted by Rubrik: “Agentic AI introduces the concept of ‘non-human error,’ and as with its human counterpart, organizations should explore solutions that allow them to correct potentially catastrophic mistakes made by agentic AI.”
Anneka Gupta.
Anneka Gupta, Chief Product Officer at Rubrik, stated: “As AI agents gain autonomy and optimize for outcomes, unintended errors can lead to business downtime. Agent Rewind integrates Predibase’s advanced AI infrastructure with Rubrik’s recovery capabilities to enable enterprises to embrace agentic AI confidently. Today’s organizations will now have a clear process to trace, audit, and safely rewind undesired AI actions.”
Last month a rogue Replit AI agent inexplicably deleted a company’s database, despite a code and action freeze being in place.
Rubrik says “Agent Rewind makes previously opaque AI actions visible, auditable, and reversible, creating an audit trail and immutable snapshots that facilitate safe rollback. Current observability tools only show what happened, but not why or how to reverse high-risk actions.”
As we understand it, an agentic AI setup, a collection of interacting agents, can make changes to the state of IT infrastructure components. For Agent Rewind functionality to work reliably, every change has to be detected along with the change agent’s identity, time-stamped, captured and stored in an immutable form. We are talking about continuous agent action backup. And we are also implying that agents are made to have their actions be observable.
Rubrik Agent Rewind agent map.
Then, an unwarranted or undesirable change has to be detected and its scope and cause understood. The next step is to identify the clean starting state before the change and roll back the changes, rewind, to that clean state. Preferably with minimal downtime, and without the rogue agents trying to reinstate the changes they made.
A rewind function can only roll back changes to IT infrastructure components that have been made by co-operating and observable agents, that it has recorded and stored.
Rubrik says Agent Rewind features:
Context-Enriched Visibility: View an inventory of your agents and identify high-risk ones. Surface agent behavior, tool use, and impact while contextualizing each action, mapping it back to its root cause, from prompts to plans to tools, to enable recovery when something goes wrong.
Agent action audit – Trace agent actions from the agent to the data or application they have accessed from a store of agent action logs.
Safe Rollback: Uses Rubrik Security Cloud to rewind what changed, whether that’s files, databases, configurations, or repositories.
Broad Compatibility: Will integrate seamlessly with a wide range of platforms, APIs, and agent builders, including Agentforce, Microsoft Copilot Studio, and Amazon Bedrock Agents, and will be compatible with any custom AI agent.
Read more in a Rubrik blog and watch a demo here. General Agent Rewind availability will be in Autumn (the Fall) this year.
Comment
Rubrik is the first of the cyber-resilient, data protectors we follow to have introduced this rogue agent action recovery technology. It told us: “Today, there are no direct competitors as Agent Rewind is the only solution offering unique ‘rewind’ capabilities. Rubrik is pioneering this functionality, providing the ability for true data reversibility, which is not offered by existing solutions.”
We expect others to follow close behind. We understand that suppliers like Trend Micro, Accenture, and Broadcom emphasize monitoring, simulation, and controlled environments (e.g., digital twins, zero trust models) that could support reversibility if paired with robust logging and backup systems. That sounds like a partnership opportunity for our set of data protecting, cyber-resilience suppliers.
Arcitecta has evolved its Mediaflux data access and management fabric to an AI-ready platform with structured and unstructured data support, a vector database, and AI model support.
MediaFlux is distributed data management software supporting file and object data storage with a single namespace and tiering capability. It works across the on-prem, public cloud, and hybrid environments, using storage tiers formed from SSDs, disk, and tape. It has a Livewire data mover and metadata database. Mediaflux Multi-Site, Edge, and Burst offerings help geo-distributed workers get fast access to shared data – text, images, video, time-series, etc. – with Mediaflux Real-Time offering virtually instant access to content data. Now Mediaflux supports AI workflows and users increase AI applicability in areas such as cancer research, genomic analysis, defense, M&E, finance, government, and scientific discovery.
Jason Lohrey.
CEO Jason Lohrey stated: “As organizations increasingly rely on AI and machine learning, the challenge of making vast, diverse datasets accessible and usable for AI training has become paramount. With an enhanced version of Mediaflux that powers AI, we are delivering a revolutionary data fabric that integrates any data asset into an AI-ready resource pool, allowing our customers to achieve better models faster and with unparalleled operational efficiency.”
He reckons: “This integrated approach bypasses the need for fragmented software development tools and separate vector stores, setting a new standard for AI data management. The result will be outcomes such as transformative advancements in cancer research, accelerated drug discovery and preservation of the world’s most important cultural archives.”
There are many potential data stores that could contribute data to AI models, such as transactional databases, analytic databases, SaaS transactional systems such as CRM, data warehouses, lakes and lakehouses, unstructured file and object stores, backup stores and archival data vaults. We just enumerated ten types of data silo. Getting data in these silos ready for AI can involve silo-specific or partchworked tools for ingest, tagging, ETL plus a vector database and vector search capability.
Arcitecta defines this as a mess and claims Mediaflux is a unified data management platform designed from the ground up to turn all your data—text, images, time series, genomics, and more—into AI-ready assets.
Lohry told us in a briefing: “Basically you need data to train your models. Indeed. That’s been the whole mission along the way; to get as much data from disparate sources into the one system, a single pane of glass, and enable your pathways to AI. And for me, actually in some ways, there’s nothing particularly special about AI. It’s a compute framework. It’s computation that’s just using a particular method; vectors basically.”
He said: “We see people moving towards unlocking the value of AI or the perceived value of that. You actually need really good infrastructure and software to drive that. It’s pretty clear though that if you win that race in many ways, then you are competitively advantaged compared to others.”
And: “You need unified data platform. So that’s our role. We are not in the business of doing the models themselves. Our job is to house the output of the models, the vectors themselves, and just treat it as general data. So we will orchestrate the external systems to go and do their analysis.”
For Arcitecta that means: “I’m heading us into a position where we are the data company basically; that’s all. Not the producer of it. Everything we do is centered around data. Doesn’t matter whether it’s in your traditional database or your non-database, your unstructured data, it’s all the one thing. So you end up with this platform that brings it all together and spans on-premises to the cloud.”
From the B&F point of view Mediaflux AI – our name – doesn’t just make make data passively available to AI. It actively helps AI with a set of features providing a linkage layer between data and AI models, that helps make data ready for use by AI, without involving separate, external infrastructure components. Its features include:
Metadata – a schema-less metadata catalog and fabric covering diverse data sources with full metadata and vector indexing in a single system, combining metadata, vector, file and object data across multiple locations
Vector embeddings stored in Arcitecta’s in-house developed XODB database
Similarity search with native vector search engine and fast semantic queries across trillions of records in milliseconds
Built-in pipelines to automate ingest, tagging and transformation
Single-pane orchestration
The XODB database is a foundational pillar for AI-ready Mediaflux and has built-in capabilities for vector embeddings and plugin support for new models managed within Mediaflux. It supports object, time-series, geospatial and vector data, and manages metadata in real time, instantly directing users toward their data, regardless of scale or location.
Vector search is critical here. Lohry tells us: ”We do see, obviously, one of the things that underpins AI vector database is in vector maths. … The beauty of vectors is we don’t have to know exactly what they mean. We can perform similarity matches on them just using very standard maths.”
We shouldn’t think of XODB as being localized: ”An emphasis for us is to say the database is not necessarily in one place. It can be in multiple places and actually joined together. You might do your analysis on data that’s 2000 kilometers or 20,000 kilometers away and produce outcomes that come back into a central system or remain in place where those are for distributed querying. The end user can get at that global network.”
XODB has a uniqueness, with Lohry saying: “I don’t know of anybody else that’s put a database in the middle of their file system and done it successfully. … A file is just a container of data. So it’s just data, and that’s all in the database and that’s how we get to the ability to do a wild card searches across file systems. …When we show people this, they are just gobsmacked and can’t believe it. But you can do an arbitrary wild card search for digital certificates in two or three building files and do it in tens of milliseconds and then go and do the same search across all points in time.”
“It’s very fast and … it handles objects, [in] which a file is an object, a directory is an object, a person is an object, an airplane is an object and has time series aspects.
This is a key point for him and Arcitecta: “We’re a database company. You can’t be in the space of managing data and have sophisticated data management systems really without a decent database because that is the limiting factor of any system. The more densely you can pack your information, the better you are. And the holy grail in databases is maximizing the information content per unit byte that you’ve got. And if you do that, then you can get at it much more quickly. And again, that’s why building these systems and incorporating vector support actually puts us in a fairly unique position with supporting AI. You’ve got everything in the one system without trying to integrate multiple systems.”
Arcitecta thinks its updated AI-ready Mediaflux offering is a good for for organizations working with massive data volumes needing a scalable, high-performance infrastructure in areas such as R&D, data science, genomics, medical imaging and machine learning Ops. Lohrey says of Arcitecta: “I really do emphasize the point that we are a data-based company as much as we are a data management company and all of this is just part of that whole persona of being in and around data.”
The new Mediaflux AI-ready capabilities are available as an integrated part of the existing Mediaflux platform. It is licensed by user count, with no capacity-based fees, and, Arcitecta says, offers a pricing edge compared to patchworked tools.
Object storage supplier Cloudian has added the Milvus vector database into to its Hyperstore offering. Scality supports external vector databases. Data orchestrator Hammerspace does not explicitly support a vector database. Ditto data manager Komprise. Both can help with the collection of data and its feeding to a vector embedding model with resulting vectors stored in an external vector database.
The hoped for AI uplift of SSD revenues can’t come soon enough for Kioxia which reported another consecutive down quarter, but there are encouraging signs in the revenue weeds.
Revenues in its first fiscal 2025 quarter, ended June 30, were ¥32.8 billion ($2.36 billion), substantially beating its ¥310 billion ($2 billion) guidance, although down 20 percent year-on-year, but just 1.2 percent down Q/Q, indicating the revenue drop rate is slowing significantly. There was a profit of ¥18.5 billion ($127.6 million), down 73.5 percent annually.
CFO Hideki Hanazawa stated: ”The first quarter results exceeded the upper end of guidance, mainly due to increased ASPS and some exchange rate effects. …We are seeing clear signs of recovery from the downward trend that we saw between the third and fourth quarters of fiscal year 2024.”
Financial summary:
Free cash flow: ¥27 billion ($186.2 million) vs ¥64.1 billion ($427 million) a year ago.
Kioxia said it’s been cash flow-positive for 6 consecutive quarters.
This latest two revenue hole is going to be temporary, Kioxia hopes.
It earns revenue from three business segments, and their revenues were;
SSD and storage: ¥217.4 billion ($1.5 billion) down 2.5 percent Y/Y, though up 1 percent Q/Q
Smart devices: ¥79 billion ($544.9 million) down 48 percent Y/Y, down 0.7 percent Q/Q
Others: ¥46.3 billion ($319.3 million) down 13.5 percent Y/Y, down 11.4 percent Q/Q
The SSD and storage revenues are split pretty much 50:50 between data center (DC) and enterprise sales on the one hand and PC sales on the other. DC/enterprise sales were flat sequentially. Although demand held up due to AI adoption, there were lower selling prices. PC sales increased sequentially as due to what Kioxia described as “partial pull-in demand.”
Segment revenue drops in SSDs and Storage, and Smart Devices, are evidently ending while Others is still trending down.
Smart device sales were flat Q/Q as, although customers started buying due to stopping inventory digestion, the transition to BiCS gen 8 218-layer flash started and it reduced demand for BiCS 5 112-layer flash.
The Others category covers sales to consumers of SD memory cards and revenue recorded through sales to Sandisk via the NAND fab joint-venture. The overall Others revenue level was down and bleak.
Trumpian tariffs have had a negligible effect thus far.
Kioxia’s guidance for the second quarter is ¥445 billion +/- ¥25 billion ($3 billion +/- $172 million). This is based on strong demand for data center and enterprise SSDs, driven by AI servers. It reckons that the PC and smartphone makers face growing demand for AI models and their inventories aren’t sufficient to meet that demand. Kioxia also expects the transition from its BiCS gen 5 to gen 8 NAND to result in a substantial increase in bit shipments.
Kioia said it is investing in equipment for its BiCS 10 332-layer flash, but that will be a 2026 shipment story.
Wedbush analyst Matt Bryson told subscribers that he is seeing a better outlook consistent with “recent reports/guidance from Hynix, Samsung, and Kioxia which have all highlighted better bit shipments (and bit shipment expectations) as well as improving pricing trends into Q3 (which we see as driven by improved cloud and AI demand).”
NAND, SSD and DRAM maker Micron raised its next quarter’s guidance today as well. It previously provided guidance for revenue of $10.7 billion ± $300 million, but now expects revenue of $11.2 billion ± $100 million. This is mostly down to better DRAM pricing.
Pure Storage looked at the VAST/Solidigm SSDs vs HDDs white paper and thought that its device-level TCO comparison was not all there was to say. The company claims that its Direct Flash Modules can do better than either alternative, specifically on power usage and embodied Carbon.
This is because its Direct Flash Modules (DFMs) having larger capacities than SSDs, and being managed at the system instead of the drive level, are needed in lower numbers to meet capacity or throughput targets and so need less electricity to operate. There is an embodied carbon advantage with its DFMs having less of this than either disk drives or commercially-available SSDs.
It also believes that just calculating the TCO does a disservice to potential buyers as “a fairly common performance requirement for cloud storage and archival tiers for throughput-optimized HDD applications (i.e., AWS st1, sc1, etc.) is an MB/sec level”, such as 10 MB/sec/TB. This, it says this is a “performance target for the storage system comparison.”
Pure sent us a trio of spreadsheet-style tables. To set the scene, it laid out a first table showing how HDD throughput and throughput-per-TB changes as disk drive capacity increases. We’ve added a Max Throughput per TB chart to the table:
There is an IO density problem here as, relative to its capacity, disk drive throughput decreases as capacity rises. Pure says that the problem is worse, as theoretical maximum throughput numbers are not achieved in real world conditions: “Layering in the real world characteristics for workload composition(aka not 100 percent sequential 100 percent of the time, disk is at least 50 percent full), expected throughput falls from 16 MBps/TB at 8TB to under 5 MBps/TB at 30TB
Having hammered this point home, it then specs out six theoretical systems which each have a 10MB/sec throughput requirement for 10 years and an effective near 10 PB capacity. There are two HDD systems, using 28 and 20 TB drives, two SSD ones with 15.36 and 30.72 TB drives, and two Pure systems with 75 and 150 TB DFMs. All the systems share the same 2:1 data reduction ratio.
It then calculates the total embodied carbon for each system and, separately, the electrical power needs. It doesn’t provide a TCO calculation as that would necessarily include its pricing details. Here are the embodied carbon table numbers:
Comparing the device count and embodied CO2e of different types of systems using updated values from a recent Harvard University research paper.
It assumes a 5-year lifespan for the disk drives, necessitating the disk drives and SSDs being replaced halfway through the 10-year period, but not the DFMs as they have a warranted 10-year life. It also adds in a few replacement drive purchases based on annual drive failure rates.
We charted the final row;
As expected, the SSDs have a higher total than the disk drives, but Pure’s DFMs have a lower total than either, as their per-drive capacity is so much higher and they have a system-level controller function, not drive-level controllers.
Pure then supplied a power consumption comparison table for the same six systems, making allowances for the drives being housed in base system chassis and expansion chassis;
Naturally, we charted the final row again, to draw out the annual power consumption;
The 15.36 TB SSD system has the same power consumption, more or less as the disk drives. The 30.72 TB SSD system has about half that level. The HDD (and 15.36 TB SSD) systems need 2.4X more electricity than Pure’s 75 TB DFM system, but 4.8X more than Pure’s 150 TB DFM system. We might expect this advantage to increase once Pure starts shipping 300 TB DFMs towards the end of the year. And this would apply to the embodied carbon numbers as well.
Pure comments: “At the system level, flash outperforms HDDs not just in speed and density but also in sustainability, cost, and reliability. The performance-per-TB collapse in HDDs has rendered them obsolete for many enterprise needs. Flash storage options, particularly those built with Pure Storage DFM, result in far fewer devices, less energy consumption, and significantly lower embodied CO2e emissions compared to traditional HDD and SSD systems.”
Open-source storage hardware supplier 45Drives has a strategic partnership with LINBIT, creators of DRBD and LINSTOR, to deliver fully-integrated, enterprise-grade high-availability (HA) storage systems built on open-source technologies. It unites “Drives’ radically transparent hardware model with LINBIT’s production-grade software stack that powers HA deployments for companies like Apple, IBM, and Amazon.”
…
Cyber data protector Cohesity introduced its Aspire Partner Program. It encompasses all partner types including Resell, Manage, Integrate, Distribute, OEM Hardware, Embedded OEM, Technology Partners, and Cloud Marketplace. Aspire participants can qualify for Premier, Preferred, or Associate tier levels with corresponding benefits. The umbrella program also features an authorized distributor component geared towards bringing more value to Cohesity resellers. There are three dimensions:
Profitable growth – offers competitive margins, rewards for partner-sourced deals, and performance-based incentives to support partners at every stage of the sales cycle, including joint-go-to-market execution, co-selling, co-branding, pipeline sharing, and marketing and demand generation resources.
Technical strength – facilitates role-based partner learning opportunities and Cohesity Accreditations. Multiple certifications further allow technical professionals to demonstrate their Cohesity data management skills and mastery for real-world environments. A special Cohesity Aces program additionally recognizes top technical partner experts with privileged access, hands-on virtual labs, and more.
Differentiation – provides multiple paths for expanding partner services and tapping into new revenue streams through partner-delivered professional services, authorized training partner certification, or building and bringing new solutions to market with Cohesity as a Professional-level or Elite-level Cohesity Technology Partner.
…
Databricks and OpenAI have partnered to bring the gpt‑oss-20b and gpt‑oss-120b models directly to the Databricks data lake analytics platform. These are open‑weight LLMs with frontier‑level reasoning and tool use capabilities, so enterprises can build intelligent, domain‑specific AI agents next to their own governed data. They have been engineered for:
Advanced reasoning and tool use – enabling complex workflows like RAG, automated decisioning, and enterprise search.
High‑speed, cost‑efficient performance – powered by Mixture of Experts architecture, with 131k context length for long‑form documents and massive datasets.
Enterprise‑grade governance – deploy within HIPAA‑ and PCI‑compliant Databricks environments, with full observability and security controls.
Databricks and OpenAI say that the launch marks a significant step in closing the gap between open‑source flexibility and enterprise AI requirements, giving organizations more choice in how they combine open and proprietary models to suit speed, cost, and quality needs. Read more in a Databricks blog post here.
…
HPE will ship ProLiant Compute servers featuring Nvidia RTX PRO 6000 Blackwell Server Edition GPUs, with a 2RU form factor. The air-cooled ProLiant DL385 Gen11 server supports up to two RTX PRO 6000 Blackwell Server Edition GPUs in the new 2U RTX PRO Server form factor. The previously announced ProLiant Compute DL380a Gen12 server will support up to 8 Nvidia RTX PRO 6000 GPUs in a 4RU form factor and will ship in September.
The next generation of HPE Private Cloud AI will be available later this year and includes support for RTX PRO 6000 GPUs with ProLiant Compute Gen12 servers, seamless scalability across GPU generations, air-gapped management, and enterprise multi-tenancy. HPE Private Cloud AI, a turnkey AI factory solution for the enterprise co-developed with Nvidia, will support the latest versions of the Nvidia Nemotron models for agentic AI, Cosmos Reason vision language model (VLM) for physical AI and robotics, and the Nvidia Blueprint for Video Search and Summarization (VSS 2.4) to build video analytics AI agents.
…
Insurgo has released a white paper: Unveiling the Hidden Risks – Rethinking Tape Media Disposal in the Modern Age. The company says traditional tape disposal methods like shredding and degaussing pose significant cybersecurity risks for organizations, leaving a backdoor into terabytes of potentially sensitive data.
The company has a patented tape media destruction technology which employs a non-wearing neodymium magnet to erase all data on the tape film by wiping the entire length of the tape, both forwards and backwards, and it also destroys the tape’s cartridge memory data chip. UK MD Gavin Griffiths tells us: “We have, from what I can see, the only court-defendable documentation chain of custody on the market today. We also have the patent for the head in the major tape countries and patent pending on the film/chip erase service.”
…
Cyber-security/data-protecting supplier N-able, which posted its first loss for 15 quarters last quarter, has sharply rebounded with Q2 revenues up almost 10 percent Y/Y to $131.2 million, beating guidance, and GAAP loss declining to $4 million from Q1’s $7.2 million loss. ARR is up 14.5 percent to $513.7 million. William Blair analyst Jason Ader suggests that “the company is beginning to turn the corner operationally—supported by favorable market tailwinds in cyber resilience, a more integrated three-pillar platform, and improving execution across the business.”
Next quarter’s revenue outlook is $127.5 million ± $0.5 million; up 9.5 percent Y/Y at the mid-point. The full year outlook is raised to $501.5 million +/- $1.5 million; up 7.5 percent at the mid-point. It was $494.5 million ± $2.5 million.
…
Toronto-based MSP Storage Guardian reduced its physical storage footprint from 9 full cabinets to just 2 by consolidating its IT infrastructure on StorONE’s SW-defined and unified protocol storage. Storage Guardian provides managed backup, disaster recovery, and cyber resilience services. It also integrated Veeam and other data protection tools into its StorONE-based infrastructure, achieving faster backup and recovery times, as well as greater scalability for customer environments.
Fun fact. Storage Guardian founder and CEO Omry Farajun is the son of David Farajun, the founder of Asigra, and the brother of Eran Farajun, who serves as a board member and EVP at Asigra.
…
TeamGroup has released its NV5000 M.2 PCIe 4.0 x4 SSD offering up to 2TB of capacity and read speeds of up to 5,000 MB/s. It has a patented heat-dissipating graphene label, which lowers operating temperatures. Measuring under 1mm thick, the label eliminates installation interference and works seamlessly with motherboard M.2 heatsinks to further improve cooling efficiency.
Chinese supplier YanRongsays it’s set an AI benchmark record in the newly-released MLPerf Storage v2.0 results. Its all-flash F9000X appliance, running its self-developed distributed and parallel file system YRCloudFile, delivered a record 513 GB/s bandwidth in the 3D U-Net test, record that is for a 3-node cluster using commodity hardware.
Each F9000X storage node was equipped with Intel Xeon 5th Gen Scalable Processors, utilizing domestically manufactured PCIe 5.0 NVMe SSDs, along with 4 Nvidia ConnectX-7 400Gbps InfiniBand network cards. The system supports up to 22 NVMe SSDs for data (up to 30.72 TB/drive) plus 2 x 1.6 TB NVMe SSDs for metadata.
A JNIST Shanhe-Huawei OceanStor A800 system was faster, at 749.9 GBps but this is a a scale-out architecture array, that supports 100 million IOPS and petabytes-per-second bandwidth, run by China’s Jinan Institute of Supercomputing Technology. It is not a commodity hardware system.
YanRong was second fastest on the CosmoFlow benchmark in the Nvidia H100 category with 136.9 GBps, behind a UBIX UbiPower 18000 system delivering 293.4 GBps. [English translation here.] The UbiPower 18000 uses a distributed architecture with a cluster supporting up to 256 nodes. It combines SSDs and persistent memory for high-speed data access, and is also not a commodity hardware system.
The F9000 delivered 258.5 GBps on the ResNet-50 test with simulated H100 GPUs, beaten into second place by the UBIX UbiPower 18000 set up again, with its 380.7 GBps.
Here is YangRong’s summary of its results;
Note that the MLPerf Storage benchmark presents GiBps bandwidth numbers, not GBps values which are used in YanRong’s table.
The company compared its results to those of DDN, Nutanix, Hammerspace and HPE:
YanRong, with its commodity hardware, recorded the second fastest read checkpoint speed; 221 GBps, with 8 clients and 64 simulated GPUs in the newly added Llama3-70B Checkpoint workload. A faster IBM Blue Vela system delivered 232.5 GBps read speed. Blue Vela is an IBM supercomputer using the Storage Scale parallel file system.
YanRong recorded the fourth fastest checkpoint write speed, 79 GBps, with an Argonne National Lab/DAOS supercomputer (126.3 GBps) recording the highest speed. The JNIST/Huawei OceanStor system (123.1 GBps) was second with IBM’s Blue Vela third (118.1 GBps).
Broadcom’s new Jericho4 Ethernet fabric router scales to 36,000 HyperPorts, each operating at 3.2 Tb/s, with deep buffering, line-rate MACsec, and RoCE transport over 100KM+ distances. The 3.2T HyperPort technology consolidates four 800GE links into a single logical port — eliminating load balancing inefficiencies, boosting utilization by up to 70%, and streamlining traffic flow across large fabrics. Jericho4 ensures lossless RoCE across 100+ km, due to deep buffering and intelligent congestion control. Broadcom claims it enables distributed AI infrastructure unconstrained by power and space limitations at a single location. It is fully compliant with specifications developed by the Ultra Ethernet Consortium (UEC), ensuring interoperability across open, standards-based Ethernet AI fabrics. Visit the Jericho4 product page here.
…
Data management and analytics supplier Cloudera has acquired Taikun which makes software to manage Kubernetes and cloud infrastructure across hybrid and multi-cloud environments. Cloudera gains a fully integrated compute layer that unifies deployment and operations across the IT stack, delivering a consistent, cloud-like experience anywhere. With this move, Cloudera says, it accelerates the deployment and delivery of the complete Cloudera platform, including Data Services and AI anywhere – from the public clouds to on-prem data centers to sovereign and air-gapped environments – all through a unified control plane.
Taikun’s engineering team will join Cloudera’s Engineering, Product, and Support organization. Taikun, based in the Czech Republic, will become a new European development hub for Cloudera. This marks Cloudera’s third strategic acquisition in just 14 months, following the purchase of Verta’s operational AI platform in May 2024 and Octopai’s data lineage and catalog solution in November 2024.
…
DataBahn.ai has developed Smart Agent, a new AI agent integrated into its Smart Edge edge telemetry collector. Smart Agent reduces complexity, cost, and risk of managing multiple tool-specific agents. Security and observability teams now have a unified and cross-functional solution to collecting, processing, and routing telemetry data. The DataBahn Smart Agent for endpoint telemetry collection is available now in Smart Edge, DataBahn’s edge telemetry collector.
…
DataPelago has launched DataPelago Accelerator for Spark. It’s first in the world to combine native execution, CPU vectorization and GPU acceleration to speed up Apache Spark workloads. DataPelago’s Accelerator (powered by DataPelago Nucleus, the company’s universal data processing engine) delivers up to 10x speedup and 80% cost reduction for Apache Spark workloads without requiring any changes to existing applications, tools, data or infrastructure, making it feasible for companies to leverage all the data at their disposal for AI and analytics.
Customers are already using the Accelerator to achieve new price + performance advantages for their data processing workloads:
• Delivered 3-4x speed up and 60-70% cost reduction for petabyte-scale ETL workloads for a Fortune 100 customer
• RevSure, a major e-commerce company, deployed the Accelerator in just 48 hours, achieving measurable performance gains and cost savings processing hundreds of terabytes through its ETL pipeline
ShareChat, India’s premiere social media platform serving over 350 million users, increased job speeds by 2x with 50% cost reduction
…
DDN has announced a partnership with SK Telecom to integrate its EXAScaler and Infinia platforms into the Petasus AI Cloud, creating a next-generation architecture for GPU-as-a-Service. This collaboration delivers virtualized GPU performance that rivals bare-metal environments, while dramatically reducing workload spin-up times from weeks to just minutes. The Petasus platform features one of Korea’s largest GPU clusters, consisting of over 1,000 Nvidia Blackwell GPUs integrated into a single cluster. It delivers state-of-the-art performance in Korea and represents an advancement over the Nvidia H100 Tensor Core GPU-based GPU SaaS released in December 2024. It is expected to contribute significantly to nationwide AI infrastructure expansion and the growth of the Korean AI industry.
Oddly the DDN press release URL that this was based on now points to an FMS 2025 award release. And yes, we have asked DDN about this.
We found a May 15, 2025, DDN YouTube video which has Jian Li, Principal Cloud & AI at SK Telecom, explaining their AI vision and the critical role of storage virtualization and optimization. He talks about the GPUaaS system and a slide mentions DDN at the 5 min 41 sec point;
EXAScaler is mentioned in a subsequent slide;
An integration POC with Infinia is mentioned at the 10 min 29 sec point. It looks like DDN was supplying storage for the Petasus system back in May but, for some reason, the press release has been retracted.
…
ElephantSqlDB, with its subsidiary Dataark Systems, announced GA of ElephantSqlDB, the industry’s first software‑as‑a‑service (SaaS) database built on a quantum‑asymptotic architecture. By combining an enhanced implementation of Grover’s algorithm – Quantum‑Asymptotic Core which accelerates search by orders of magnitude without specialized quantum hardware. With Nvidia GPU simulation technology, the platform maintains single digit‑millisecond search speeds—even as data volumes grow from terabytes to petabytes—while keeping pricing predictable and affordable. ElephantSqlDB gives hyperscale performance, and transparent pricing that works for solo creators and Fortune 500 teams alike. Learn more at www.elephantsqldb.com.
…
GPU-powered RAID card provider Graid announced the official availability of SupremeRAID HE (HPC Edition). It’s designed specifically for high availability in clustered and distributed environments. It offloads RAID operations from the CPU to the GPU, unlocking extreme NVMe performance—up to 132GB/s read and 83GB/s write throughput after RAID processing—while freeing up CPU resources for upper-layer applications. It has support for cross-node array migration, failover without replication, and full compatibility with Ceph, Lustre, MinIO, IBM SpectrumScale, and more. Explore SupremeRAID HE: (HPC Edition) here.
…
NetApp’s Amazon FSx for NetApp ONTAP is now certified and supported as external storage for Amazon Elastic VMware Service (Amazon EVS), which is now generally available. Amazon EVS is AWS’s new native service that allows organizations to run VMware Cloud Foundation directly within Amazon VPC. With FSx for ONTAP support for EVS, NetApp becomes the only enterprise storage provider with a first-party, native AWS service certified for EVS, helping customers accelerate VMware workload migrations with no re-platforming or re-factoring, reduce costs by up to 50%, and strengthen cyber resilience. [But see Pure Storage item below.]
New capabilities include:
Migration Advisor for EVS (BlueXP workload factory): Automates VMware workload discovery and migration planning.
Expanded disaster recovery for EVS: Supports both NFS and VMFS/iSCSI for flexible recovery options.
Enhanced ransomware protection: ONTAP autonomous ransomware protection and BlueXP ransomware protection now natively support EVS workloads in AWS.
…
The NVM Express consortium announced important new features and engineering change notices (ECNs) across the set of 11 NVM Express (NVMe) specifications. The updated specs include:
NVMe 2.3 Base Specification
Command Set Specifications (NVM Command Set 1.2, ZNS Command Set 1.4, Key Value Command Set 1.3, Subsystem Local Memory Command Set 1.2, Computational Programs Command Set 1.2)
Transport Specifications (NVMe over PCIe Transport 1.3, NVMe over RDMA Transport 1.2, NVMe over TCP Transport 1.2)
Rapid Path Failure Recovery: Permits communication with the NVM subsystem through alternative channels, allowing for rapid recovery in the event of the loss of communication to a controller to avoid data corruption and duplication of outstanding commands.
Power Limit Config: Provides complete control over the maximum power of an NVMe device, particularly important for older systems with limited power capabilities.
Self-reported Drive Power: Allows the host to measure and monitor NVMe device power, as well as power consumption over the lifetime of the device, addressing ongoing maintenance and sustainability concerns.
Sanitize Per Namespace: Enables cryptographic erase sanitization on individual namespaces, as opposed to broadly impacting the entire NVM subsystem.
Configurable Device Personality: Supports mechanisms that allow a host to securely modify the NVM subsystem configuration, easing inventory management for device providers.
The NVM Express specifications and the new feature specifications are available for download on the NVM Express website.
…
Cloud file services supplier Panzura has announced AI capabilities, powered by ML models, for enterprise-grade threat detection and prevention in the Panzura CloudFS hybrid cloud file platform. CloudFS slashes the response time to seconds for ransomware, data exfiltration, and other suspicious behavior while eliminating false positive noise. It moves beyond traditional detection to predict, learn from, and eliminate threats before attackers can complete their mission. The new CloudFS capabilities, now part of the renamed Threat Detection feature set formerly called Detect and Rescue, harness AI to create unique behavioral fingerprints for every user. It immediately takes automated action when that behavior deviates substantially from the norm, ensuring operational continuity and rapid recovery.
The system instantly logs incidents, notifies admins, and disables compromised accounts based on assessed risk severity, cutting response times from weeks to seconds. CloudFS combines data management with both passive and active defenses which are foundational to cyber resilience. It consolidates data from disparate sources, makes it immutable so that it cannot be permanently damaged, and provides restore points that are never more than 60 seconds old. This instant snapshot recovery eliminates backup dependency.
The CloudFS platform flags privilege escalation, after-hours access, and other anomalous patterns that indicate compromised accounts. Panzura’s data services platform, Panzura Symphony can then provide deep analysis of permissions to identify unauthorized modifications or suspicious access rights.
We understand that Panzura is using Index Engines’ CyberSense technology.
…
The PCI-SIG announced the PCI Express (PCIe) 8.0 specification will double the data of the PCIe 7.0 specification to 256.0 GT/s and is planned for release to members by 2028. The PCIe 8.0 specification is aimed at supporting emerging applications like Artificial Intelligence/Machine Learning, high-speed networking, Edge computing and Quantum computing; and data-intensive markets like automotive, hyperscale data centers, high-performance computing (HPC) and military/aerospace.
PCI-SIG technical workgroups will be developing the PCIe 8.0 specification with the following feature objectives:
Delivering 256.0 GT/s raw bit rate and up to 1 TB/s bi-directionally via x16 configuration
Reviewing new connector technology
Confirming latency and FEC targets will be achieved
Ensuring reliability targets are met
Maintaining backwards compatibility with previous generations of PCIe technology
Developing protocol enhancements to improve bandwidth
Continuing to emphasize techniques to reduce power
…
Customers using Supermicro’s Petascale Storage Family will be able to leverage Phison’s high-capacity 122.88 TB Pascari D205V SSD, in the E3.L form factor and with PCI gen5 NVMe performance. E3.L offers a longer form factor which unlocks double the capacity compared to E3.S as well as improved airflow and thermal management.
…
Pure Cloud Block Store (PCBS) supports Amazon Elastic VMware Service (Amazon EVS)which is now GA. PCBS on AWS is certified with VMware, ensuring compatibility and enterprise-grade reliability for virtualized workloads.It”delivers a seamless, storage-optimized path forward for Amazon EVS customers—whether you’re migrating from on-prem, extending hybrid operations, or evolving from another cloud deployment. It gives VMware environments the runway needed to take off into AWS—with the same tools, the same control, and the same enterprise-grade features, integration, and performance you rely on in the data center.” Read more in a Pure blog.
…
Quantum partner ASI Solutions, operating in Australia and NewZealand, has launched ASI Cloud InfiniStor, a cloud storage platform built on Quantum ActiveScale object storage. “InfiniStor reduces storage and environmental footprint by 40%-60% compared to traditional solutions, thanks to ActiveScale’s advanced erasure coding,” said Lloyd Vickery, country manager for New Zealand at ASI Solutions. “This allows us to offer New Zealand businesses lower costs and higher reliability, all hosted locally.” Customers pay per terabyte with no hidden fees and enjoy unlimited egress for zero-fee data retrieval. The University of Auckland, an early adopter, transitioned from legacy tape libraries to InfiniStor to manage growing research data. ASI plans to expand InfiniStor’s reach to Australia, explore white-label offerings for IT providers, and integrate AI-driven analytics.
…
Sophos has teamed up with Rubrik to launch Sophos M365 Backup and Recovery and provide a unified global platform to enhance cyber resilience against ransomware, account compromise, insider threats, and data loss in SharePoint, Exchange, OneDrive, and Teams. It integrates Rubrik’s SaaS-based protection directly into the Sophos Central platform and Sophos will offer this to its 75,000 MDR and XDR customers—enabling fast, secure recovery of critical Microsoft 365 data in the event of accidental deletion or malicious compromise.
…
Unified database supplier SingleStore announced its official launch in Japan. It says this expansion comes at a time of immense growth for SingleStore. The company grew from $10 million in ARR to over $100 million in just four years and achieved unicorn status with a valuation of over $1 billion following its most recent funding round in 2022. SingleStore currently has customers across 50+ countries, with its expansion to Japan marking its eighth office globally. This includes its presence across the APAC region, with customers in 12 APAC countries, a regional hub in Singapore and a strong presence in India. The company already serves over 100 customers in the region, including leading enterprises such as Nikkei, Samsung, Tata, DBS Bank, Posco, and 6sense.
…
Korea’s Chosun reportsSK hynix has decided to deliver the sixth-generation high-bandwidth memory (HBM4) 12-layer products to Nvidia at prices about 70% higher compared to the fifth-generation high-bandwidth memory (HBM3E) in the first half of this year. Chpsun says “It is understood that negotiations regarding prices and supply quantities for products to be delivered to Nvidia next year have been finalized ahead of Micron and Samsung Electronics. The HBM4 12-layer products are expected to be equipped in Nvidia’s next-generation artificial intelligence (AI) chip, the Rubin platform, set to be released in the market next year.”
…
VDURA has developed a scalable reference architecture with AMD that combines VDURA’s V5000 storage platform with AMD Instinct MI300 series accelerators. It’s a working blueprint that defines compute, storage, and networking for jobs running on 256 or more GPUs. Highlights include:
Compute Density – 32 MI300 servers × 8 GPUs = 256 GPUs per unit
Storage Throughput – Up to 1.4 TB/s & 45M IOPS (all-flash); 280 GB/s with flash + HDD expansion
Usable Capacity – ~5 PB in a 3 Director + 6 V5000 node layout
Data Durability – Up to 12 nines via Multi-Level Erasure Coding
Download the full 20 page architecture reference guide complete with BOMs, cabling maps, and tuning best practices, here.
…
VDURA has announced a strategic partnership with New Mexico State University (NMSU) to co-develop and commercialize post-quantum cryptographic (PQC) technology that safeguards the petabyte-scale data pipelines powering next-generation AI and HPC workloads. NMSU, recently elevated to Carnegie R1 status, brings world-class cryptography and cybersecurity research to the collaboration. Together, the teams will integrate NIST-selected PQC algorithms into VDURA’s parallel file-system stack, enabling line-rate encryption of GPU-accelerated workloads without sacrificing performance.
…
VergeIO says global sports entertainment business Topgolf has standardized on VergeOS to upgrade its IT infrastructure across all Topgolf offices, data centers, and venue edge data centers. It has over 100 high-tech venues across the United States and worldwide, including locations in Europe, Asia, and the Middle East. Each venue offers an interactive golf gaming experience that blends physical gameplay with digital analytics. Players use real golf clubs and balls embedded with RFID, aiming at a field of electronically monitored targets. Toptracer technology visually tracks every shot in real-time, while analytics systems power extended reality games and provide performance insights. The upgrade was triggered by escalating costs and complexity from VMware’s licensing model under Broadcom ownership.
Topgolf is replacing a VMware deployment on Dell VxRail with a leaner, more efficient three-node VergeOS architecture per site. It’s eliminating third-party backup software and hardware and using VergeOS’ built-in data protection capabilities, including ioClone, ioGuardian, and ioReplicate.
…
XConn announced successful interoperability and performance optimization between the XConn CXL 3.1 switch and the ScaleFlux MC500 CXL 3.1 Type 3 memory controller at the Future of Memory and Storage (FMS25) event. This collaboration marks a significant milestone in enabling memory disaggregation and pooling capabilities for AI and cloud infrastructure at scale. The ScaleFlux MC500 CXL 3.1 Type 3 controller, developed in close collaboration with hyperscalers, memory providers, and CPU vendors, introduces a groundbreaking list decoding ECC architecture to deliver unmatched reliability, availability, and serviceability (RAS) in DRAM-based memory systems.
…
SW RAID supplier Xinnor says Consorzio LaMMA, Tuscany’s leading meteorological organization, with a 32-node HPC cluster, has upgraded its mission-critical weather forecasting infrastructure with a 2RU Full Flash storage system from E4, BeeGFS 7.2.15 parallel file system SW, and Xinnor xiRAID. The upgrade delivered exceptional performance improvements (62GB/s read speeds, 125,000+ IOPS) while dramatically reducing power consumption and physical footprint—all achieved with zero downtime during migration.
The Full Flash storage node features:
Hardware: Supermicro 2125HS-TNR server with dual AMD EPYC 9124 16-Core processors
Storage: 12x 3.84TB Kioxia CM7-R RI NVMe drives in RAID6 plus 2x 1.92TB drives in RAID1
