In Today’s Issue:

💸 What actually happens when an expensive GPU sits waiting for data

🗄️ Why the cold archive became the "crown jewels" of enterprise AI

🧠 What an "AI operating system" means once you strip out the marketing

🔒 How security quietly decides where training and inference meet

📊 The one metric Breuner trusts more than GPU utilisation

🏭 What breaks first when you scale from thousands to tens of thousands of GPUs

Dear Readers,

Everyone talks about GPUs. The AI build-out is measured in H100s and gigawatts, and the whole industry is priced on who can secure the most compute. But ask the people who actually run these systems where the money leaks, and they point somewhere less glamorous: the storage layer underneath, the part that decides whether those expensive GPUs are working or waiting.

That is Sven Breuner's world. He spent more than a decade at Fraunhofer as chief architect of BeeGFS, a parallel file system that ran on some of the fastest supercomputers of its era, before joining VAST Data, the storage company that just raised roughly $1 billion at a $30 billion valuation on the argument that the data layer is now the center of AI infrastructure. As Field CTO International, he spends his days on exactly this problem with VAST's largest customers.

We sat down with him for a Superintelligence exclusive. He was refreshingly blunt about what happens when a GPU waits ("you are wasting a lot of money"), why the cold archive suddenly became the crown jewels, and what an "AI operating system" really means once you cut the marketing.

All the best,

Kim Isenberg

Kim Isenberg in conversation with Sven Breuner, Field CTO International at VAST Data.

VAST Data started in 2016 as an all-flash storage company. Over the last three years it has repositioned itself as what it calls an AI operating system, a single platform combining storage, database, vector search and an execution engine, and in April 2026 it raised roughly $1 billion at a $30 billion valuation. We wanted the argument underneath the marketing, so we put it to the person who makes it to VAST's largest customers. What follows is that conversation, lightly edited.

Tired of overpaying GoDaddy or Namecheap? There's a smarter option with Porkbun!

Porkbun is the domain registrar trusted by creators, developers, entrepreneurs, and anyone who wants low prices without the nonsense or the forced upsells.

Here's why so many people are making the move to the Bun:
• Most domains sold at cost
• Low, transparent registration and renewal pricing
• Free features like WHOIS privacy and SSL certificates
• Powerful web and email hosting options
• Real human support 24/7, 365 days a year
• Named the #1 domain registrar by Forbes Advisor and USA Today

Whether you're launching a business, building a personal brand, starting a side project, or creating your very first website, Porkbun makes the whole thing easy.

Get $1 off your next domain registration with Porkbun now.

Get Your Domain At Porkbun

How would you describe your own path, from BeeGFS and HPC file systems to today's AI infrastructure problems?

I was fortunate, and maybe unfortunate at the same time, to start my career at an HPC research center. I was surrounded by researchers, physicists, chemists, people developing algorithms that genuinely needed to run at scale to solve problems you cannot solve without that compute power. Then Fraunhofer tasked me with building a new parallel file system that was better and easier to use than what existed. I learned an enormous amount, and some of the problems we had back then are the same ones people are still struggling with today: lots of small files, algorithms that do not scale because of storage access.

The unfortunate part is that I was living in a bubble, in hindsight. In HPC you have a very particular view of the world, focused on simulations, and we were not really looking outside the box. That is why I was so happy when I found VAST, which came at this from the enterprise angle. It gives you a completely different sense of what actually matters for the world, not only simulations but how you keep economies and businesses running.

When you talk to enterprise customers, what is the first indication that their AI bottleneck is the data infrastructure rather than the model?

We all know how easy it is to ask an LLM a question now. But what does that mean for an enterprise? The enterprise does not primarily want to ask about general public knowledge on the internet. What matters to them is their own data, gathered over years, and turning that into something intelligent.

Data access can absolutely be the bottleneck. But there is a second question underneath it, which is how you actually learn something from the data. That is where the complicated AI stack sits, and where you normally need a lot of experts. Our ambition is to integrate as much of that stack as possible into the product so the barrier to starting is low.

For readers who understand GPUs at a high level but not storage architecture, what actually happens when a GPU is waiting for data?

Essentially you are wasting a lot of money. The GPUs are very expensive, so you had better put them to good use.

There are two ways to do that in an AI context. Either you train on the data you have, coming up with a foundational model that understands general things like how to interpret a picture or a video, which is a big undertaking and why you have specialised companies doing it. Or you take such a model and fine-tune it on company data. That is where enterprises need the GPUs. And it usually has to go together with security boundaries. Even if we are all part of the same company, that does not mean everyone has access to all of the company's information.

What does GPU utilisation mean in a real AI factory, and what level do customers think they have versus what they actually get?

The hurdle is getting the GPUs and getting the budget allocated. Very often people start small, they build experiments, everything runs nicely at that scale, and then comes the big step where they try to scale it out. Suddenly they find that the things that worked well at small scale do not work the same way at large scale. That is the whole point of the HPC community, understanding how you scale something, and I think that is what companies really need help with.

How are AI workloads different from traditional HPC workloads, from a storage perspective?

The access patterns are much more random. That is one of the reasons spinning disks no longer work and why it mattered so much to us to let people put all of their data on flash. It is also about accessing far more of the data than before, because the more the AI can be trained on, the better it understands the world.

In classic HPC there was always an active working set that you put on fast storage, plus a big archive that was low touch and rarely accessed. That has changed significantly. The archive is now basically the crown jewels, the material the AI can learn everything from. So suddenly you need very fast access to all of the data instead of only small working sets.

Semianalysis recently argued that inference speed is becoming as important as model quality. Does that change the picture?

It does. As we move to agentic AI, it is one thing if a human asks the computer a question and waits a few seconds. But if different LLMs are talking to each other, it takes far too long if each one has to wait for the answer from the other. That is why speed matters so much now.

Let's be frank about the term. Beyond the marketing phrase, what does an AI operating system actually mean?

Some people hear operating system and think of macOS or Red Hat. That is not what we mean. We mean something that stores the data, can do computation on the data, and has an execution engine to actually run things.

We started as a storage company that made data AI-ready by getting it onto flash. Classic storage was built on spinning disks, which are mechanically bad at the fast random access AI needs. All-flash was cost prohibitive before us, so we introduced new data reduction algorithms that made it affordable. That was the first step.

Then we asked what comes after storing files and objects. We found that many customers were putting their data into databases that were mapped to files on a file system, which is unnatural. You have nicely structured data and you put it into an unstructured format. So we added database functionality, and that decision happened just before the ChatGPT wave. We quickly realised a classic database would not cut it and we needed vector functionality, so that chunking and vectorisation could happen on the same platform where the data already lives.

The next step was an execution environment. Whenever a new image or a video from a meeting gets stored on the system, that should instantly trigger an action that teaches your AI model about the new information. AI is already a workflow in itself, and companies have any number of more complicated workflows that can be triggered the same way.

If the AI OS unifies storage, database and compute, what becomes simpler for the customer and what stays complex?

Having a vector database does not automatically tell you what to do with a new video file, or how it gets indexed. There are tools and standard mechanisms for that, and we provide the automation to run all of it directly on the platform.

We also learned something from enterprise customers. We would say the storage system has an execution engine and can submit jobs to your GPU cluster to vectorise new videos. And they would say yes, but the GPU team is completely separate from the storage team, and the networking team sits between them. Just getting those three teams to work together is complicated. That is why we concluded the only way to make this simple is to integrate everything, which is where our CNode-X server model comes in. It includes a GPU, so you do not have to worry about the network and a separate GPU cluster. Then with a graphical interface you can say: if a new video lands in this directory, do all of the processing so it becomes instantly queryable.

What stays hard is knowing what you are doing. We define predefined templates, but it still helps to understand the differences between models, which one is appropriate for analysing meeting recordings versus security camera footage, how you do the chunking. Any company serious about this should work with a consultancy. We do not build the models for our customers. We provide the framework to integrate everything easily.

How does the storage layer affect inference, not just training?

Thank you for that question, because it lets me come back to security. One fundamental problem with LLMs is that they happily share everything they have learned with whoever asks. That is not what a company wants. If the boss asks, give them all the company information. If someone from a specific department asks, answer only on the basis of what that person has access to.

This is where training and inference converge, through retrieval augmented generation. The LLM no longer has all the knowledge baked in. It has a full index, like an internet search engine, which is the vector database. It knows which files or objects are relevant to a question, and it then tries to access that data on behalf of the user asking. If the user does not have access to a document, the model cannot use it in the answer. That is how you enforce security.

It also means the model does not need to be fully trained up front, which is a much more lightweight process, and it is far closer to real time. A new video comes in, you do a lightweight analysis and indexing, and it is immediately available for questions. You had a session with a customer and you want it available now, not after next weekend's training run.

What should enterprises actually measure to know whether the data layer is helping or hurting?

GPU utilisation is a good starting point. But I would say enterprises and research organisations should measure themselves by productivity. Do you feel you are at the forefront of technology, or do you feel you are falling behind? There is no point in having your GPUs fully utilised if you do not think it is making a difference to your research or your business. Productivity should be the measure, and that is why we try to keep the barrier to starting as low as possible.

When a vendor says it can solve I/O bottlenecks, what should a technical buyer ask before accepting that claim?

I/O patterns for these workloads are much tougher than they used to be, so customers going through modernisation find out quickly that the old infrastructure does not cut it. The next question is how much you need. Our answer is that you should not build a small isolated system for one purpose, because then you are bound by that little slice of capacity and you need so much performance out of it that it becomes unreasonable. This should be treated as a strategic topic. If you have a good all-flash system for all of your data, there is easily enough performance in it that this stops being a worry.

The important thing is that it is no longer only humans using the system. You have to account for agents accessing it. I/O patterns can be very fast and furious.

Are there workloads where a classical HPC file system is still the better choice?

Hopefully not. I would see that as a mistake nowadays. It would be a clear sign that you are not enabling your researchers and data scientists to do what they should be doing. You are holding them back by sticking with what you have done for the past twenty years, and that will not carry you into the future.

What are the trade-offs of the VAST architecture that customers should understand before adopting it?

We usually expect our customers to work closely with us. I hope this does not sound arrogant, but we think we are already very far ahead of the other companies in a similar domain. So we expect our customers to guide us on the next steps, because that is how we develop new features. We build what customers ask for. So customers had better ask us for something.

As AI factories scale from thousands to tens of thousands of GPUs, what breaks first if the data architecture is not redesigned?

Storage access is the first thing that breaks. Some people will say GPUs also break these days if you use them too heavily, but that is a different topic. Joking aside, if you did not plan for storage access, scaling out is not trivial.

The other failure mode is productivity. If you grow too organically from a small setup to a big one, you end up with a very complicated architecture. It becomes patchwork. This does not work so we replace it, then you suddenly need another thing, and you end up with something fragile. It makes sense to plan ahead. If your project today is successful, it will have to work at scale, so think about that up front rather than afterwards.

Do you expect enterprise AI infrastructure to converge with HPC infrastructure, or will AI require a separate stack?

That is an almost religious question. I would say HPC is the essence of AI, in the sense that HPC is the body of knowledge for making something work at scale, automating things to run in batched fashion on large systems, designing algorithms so multiple nodes can work together. AI would not work at scale without HPC knowledge and HPC technology.

In another sense you can see AI as completely separate. If I am an enterprise that wants to understand all the information stored on my systems, that is an AI topic, not an HPC topic. But there is so much research going on globally where it would make no sense to have an HPC centre separate from an AI centre. AI is helping with classic HPC workloads, so it needs to run on the same systems.

Tough question: what will matter more over the next few years, faster GPUs, more memory, better storage, or better orchestration across the whole system?

Can I say all of it? Faster GPUs, yes, and we can help keep them busy. I have no doubt the usual suspects already have the next generation in progress and probably the one after that. Energy is clearly becoming a problem, because not every organisation can have its own little power plant next to the data centre. And then there is understanding the data. AI can only be as good as we help it to understand the data. The good thing about AI is that it recognises patterns we did not see before. But the better we understand the data ourselves, the better we can teach the AI to find what is hidden deeper in it.

If an enterprise CTO is about to spend heavily on GPUs, what is the one data infrastructure question they should ask before signing?

Some of them should maybe not place that order at all. There are specialised neoclouds that spare you from buying the GPUs in the first place, especially if you know you will not need them around the clock. If part of your organisation is based in Germany, you probably do not need the GPUs overnight in Germany. So why not have them in a colocation where the operator can rent them out overnight to someone else?

The neoclouds increasingly understand that this is not about renting out affordable GPUs. It is about how easy they make it for a new enterprise customer to start taking advantage of them. So they are starting to offer infrastructure services and sometimes consulting. The question a CTO should always ask is: how easy is it for me to get started, and what do you bring to the table so that I can become productive quickly?

Anything you would like to add that we have not covered?

Yes. Some companies are reluctant to adopt AI because they had projects that did not turn out as well as they hoped. Everyone assumes computers do not make mistakes and always work perfectly, and then it turns out an LLM can occasionally give a wrong answer. But that is because LLMs are modelled on our brains. If you hire a new employee, that person will also occasionally give a wrong answer, and that is not a reason not to hire them. They can still be extremely useful. It is the same with AI. We have to accept that it will not be 100 percent accurate all of the time, and it is still enormously useful. Things are moving so quickly that if you do not start now, you are already falling behind, and it is very hard to catch up if you wait longer.

This interview has been edited for length and clarity.

Breuner's most useful point is the one about the archive. For twenty years, storage architecture rested on the assumption that a small working set is hot and everything else is cold. Training and retrieval have collapsed that distinction, because the cold archive is exactly where a company's differentiated knowledge sits. Once every byte becomes potentially hot, the economics of the whole stack change, and that is the argument behind VAST's $30 billion valuation as much as anything NVIDIA says about bottlenecks.

The second point worth carrying forward is his answer on measurement. GPU utilisation is a number that flatters procurement decisions after the fact. Productivity is harder to measure and easier to feel, and it is the metric that tells a CTO whether the last eight figures of capex bought capability or just capacity. If there is one idea to carry out of this conversation, it is that the expensive part of your AI stack is only ever as valuable as the boring layer feeding it. Our thanks to Sven Breuner and the VAST Data team for the exclusive.

VAST Data, founded in 2016, began as an all-flash storage company and now positions itself as an "AI operating system" spanning storage, database, vector search and an execution engine. In April 2026 it closed a roughly $1 billion Series F at a $30 billion valuation, led by Drive Capital with Access Industries as co-lead and NVIDIA, Fidelity and NEA participating. That is more than triple the $9.1 billion valuation from its 2023 Series E, and it puts the private company above Pure Storage's public market cap. VAST reports more than $4 billion in cumulative bookings and over $500 million in committed annual recurring revenue.

Sources:

🔗

🔗

🔗