Nebius (NASDAQ: NBIS) is the AI cloud company building a full-stack platform for everything from data and model training to production deployment. Nebius is headquartered in Amsterdam and publicly traded on Nasdaq. The cloud it runs was built from the ground up for AI workloads, and the team’s way of describing it: what you'd get if a hyperscaler and a supercomputer had a baby.

Nebius is wagering that the AI cloud race will be won at the software layer - that capacity, runtime optimization, model adaptation, and developer experience all have to come together in one platform, and that GPU count alone won't decide who comes out ahead.

Over the past year, the center of gravity in AI has moved from training to inference, and Nebius has climbed the stack to meet it. Its managed inference platform, Token Factory, is the centerpiece of that strategy. In six weeks this spring, the company made three moves to deepen it: the acquisition of Eigen AI for model and runtime optimization, the addition of the Clarifai team for the system layer, and the acquisition of agentic-search company Tavily.

This is Nebius's second Deep Dive with Cerebral Valley. The first ran in December 2025 with co-founder and CBO Roman Chernin, and a lot has changed since. Few people are better placed to tell that story than Danila, who has run the software stack since the very beginning.

In this conversation, Danila gets into why execution speed is the real moat, what it actually takes to build across all three layers of the inference stack, the thinking behind the Eigen, Clarifai, and Tavily deals, and why he wants Nebius to be the independent infrastructure for independent builders.

Let's dive in ⚡️

Read time: 9 mins

Dan, welcome to Cerebral Valley! To kick us off, give us the short version of who you are and your role at Nebius, plus a quick intro to the company for anyone meeting Nebius for the first time.

My name is Dan, full name Danila, and I'm CTO here at Nebius. I've been with the company from the very beginning. Call it the founding team, the initial team, the first employee, whatever; this was way before the whole thing took shape, back when it was still a prototype of a prototype of an idea. I've been here ever since, always responsible for the software stack. After what was actually our third pivot, we landed on building a hyperscaler-like cloud product aimed mainly at AI workloads: Jupyter-oriented, high performance, all of that. Our internal slogan is that the Nebius AI cloud is what you'd get if a hyperscaler and an HPC cluster had a baby. That's the mental model.

We think we got the product right, and the timing too. It lets us address the whole market, not only the top tier that procures huge clusters. We also have a healthy fraction of clients at the retail end, some consuming with an MRR under $100. Since then we've closed some noticeable deals, started new business lines, and stayed busy on the acquisition front.

In a recent post you wrote that the AI cloud will be won at the software layer, and that execution speed is the moat. From your seat as CTO, what does that mean in practice, and how do you read the last 18 months as an expression of it?

It's less about being a CTO, or about technology specifically. It's about the decision-making process. You can't execute if you stall on decisions, and that's something we've worked on hard internally. Sometimes we get it right, sometimes we don't. But if you can make quick decisions, you're already in the leading pack. After that you make smaller decisions, but you keep the pace. That's your execution speed.

You take the risks. You stay less focused on potential problems and more on the upside. That's how we try to operate, and it's how we've shaped the culture. We don't talk about potential problems, we address them. We talk about the upside instead, about future products, about what else we can do and how to get there.

The last 18 months have been a lot. Someone will join an internal planning discussion for the next quarter and say, "but in three years we might have this problem." Three years ago we lived in a completely different world; a company like Nebius couldn't have existed at all. Less demand, less hype, the revolution wasn't really here yet. ChatGPT 3.5 and 4 were only about three years ago, so this whole industry has grown up inside that window. Eighteen months is half of it, and a huge amount has happened. Across build-outs, financing, corporate structure, technology, and product, I'd say we've been moderately successful. From the inside you always know what you could have done better. But I'm happy with where we are.

You've described the center of gravity in AI shifting from training to inference, from where models get created to where AI becomes the product. When did that shift become legible from inside Nebius, and how did it change what you were building?

Over the past year it stopped being an opinion and became a fact. You can see it inside the cloud, in the workloads themselves. There was a point where, if you looked at the internal dashboards, almost half of our capacity was running inference, which is unheard of at our scale. It fluctuates as clients come and go, but compare that to when we started, when 98 or 99% of capacity was training, and you can see the industry changing shape.

Bootstrapping a frontier lab and doing your own pre-training is borderline impossible now. Some friends and I were doing napkin math recently, and my estimate is you'd need a seed round of roughly $50 to $100 billion today to set up the infrastructure, the talent, the ambition. I'm not saying it's impossible, but it's a complex endeavor, so we don't see new leaders being shaped on pre-training.

At the same time, the adoption is everywhere. AI went from something you read about in the news to something in your pocket every day. There's no category of "AI-enabled software" anymore; every piece of software ships AI features as a minor version bump, and suddenly your SSH terminal is completing your commands or summarizing their output. Inference workloads are a function of that adoption, not of investment dollars or the talent in frontier labs. That's training. Inference is where the use cases live, and there are more of them every week. A couple of months back, when Anthropic was struggling with Claude availability, that was a pure inference-capacity problem. If Anthropic hits that, you can infer what it means for the rest of the market.

Part of our answer was to change how we work. We rebuilt our acceptance-testing frameworks, for one. Most of the workloads we used to test against were modeled on training cycles; now they're shifting toward inference, because the latency demands are different. Our deployed engineers retrained their focus toward inference, and we stepped into managed inference: not just handing someone capacity to run a model, but running it for them. Endpoints, traffic, KV caching, optimization. This is not a solved problem, not by a long shot. Take a freshly released open model, follow the manual to run it on some hardware, and your cost will be insane, on the level of what you'd pay a closed API, which is not how open models are supposed to work. So we've poured effort into the expertise of running them efficiently and tuning them to a specific traffic pattern. The AI cloud is our proven product, and our big custom build-outs for large clients are doing exceptionally well, but managed inference is the strongest bet we have for the next iteration.

You break the inference stack into three layers: infrastructure, model and runtime, and the system layer where everything comes together. Most providers focus on one or two. Walk us through what it really takes to build across all three.

The key thing is that no single layer wins it for you. Being exceptional at one of the three doesn't put you ahead. Infrastructure matters, and capacity availability is a big deal these days, but it isn't everything. As I said, take the out-of-the-box recipe and your performance numbers are atrocious.

So you invest in runtime. You learn how to tune the inference server, how to write and deploy custom kernels, how to manage KV cache, which is huge. For modern long-running agent workloads, your KV cache hit rate can sit at 98%, which means that for the heavy traffic of a successful product, your GPUs are only computing a fraction of what they'd otherwise have to. That's what lets you push throughput and hold your SLAs.

Then there's the model layer. When we say model, we don't mean pre-training your own. You don't have to be the team behind Kimi to own this. It's fine-tuning, reinforcement learning, post-training: shifting a model's behavior toward your scenario, which lifts accuracy and throughput because you spend fewer tokens. There are also ways to optimize the model itself, like speculative decoding, where a smaller model predicts the next tokens without going through the big one, or model pruning, where you know the use case well enough to modify the weights so the knowledge you need stays and the rest falls away. A smaller model runs faster.

A lot of this is public but not collected anywhere. There's no online course for running inference at scale. It's frontier research across all three fields, and you have to push on it daily. You can't win without all three, and to be the best you have to be the best at all three. They're fairly independent, the overlap isn't large, but gains in each one stack up in the final result.

You've also said inference isn't one workload that serves all. Some teams optimize for latency, some for cost, some for compliance, some need on-prem or air-gapped. How do you architect for that, and where does the flexibility turn into a real differentiator?

I'm not sure I have one clean answer today, because the market isn't mature enough. We see the shift, but we don't yet know exactly what will matter most, so we place bets everywhere, watch what wins, then double down. What's already clear is that privacy is a big issue. It's part of why I think there will always be a market for open-source models; they give you control and flexibility that closed providers probably won't. The bigger OpenAI or Anthropic get, the more niches they can serve, but there will always be people unwilling to hand over the very sensitive data that inference and agentic workloads now touch.

So air-gapped will be a real case, and that's where the three layers earn their keep. In an air-gapped install that the customer owns and runs, capacity might be a constraint, but the other two layers still matter; you can bring the orchestration and model layers on top of it. We built Token Factory for exactly this. From day one it was a standalone product, essentially a client of the cloud rather than something wired into our internal infrastructure, which makes it portable. It runs best alongside our cloud and our data centers, but in principle it can move outside Nebius's perimeter and still work. That was a deliberate decision from the start, because we could see the resource scarcity coming, and we could see that some agentic workloads would have to live inside a tightly controlled environment.

You've brought in Eigen AI for the model and runtime layer and Clarifai for the system layer. Make it concrete: what does each team change about what Nebius can deliver, and what was the ceiling before them?

Part of me wants to say the sky was the limit, and it always is. But execution speed is the moat, and building that expertise organically, getting people experienced enough over time, would have taken a while. The acquisitions are a shortcut. With the Eigen team we get a group that's deeply versed in model and runtime optimization: publications, real practical experience, proven against industry-recognized benchmarks, strong ties to academia. That gets us there today, not in half a year.

Clarifai is a different angle; serving, orchestration, hardware support, reliability, and the practical engineering needed to make customer workloads run in production. But they are not just one layer, they've been very successful in model and runtime optimization too. The two teams consistently rank the highest on industry inference performance benchmarks. It's a natural fit. We get talent with proven execution, and both teams fold straight into the Token Factory team and the wider engineering org, involved in everything we ship for customers, how we shape the product, the benchmarks we publish, and this is a case where one plus one equals three.

The Tavily acquisition takes Nebius beyond inference into agentic search, the idea being that Token Factory helps agents reason while Tavily helps them ‘know.’ For developers building agent systems, what was broken about the status quo, and what changes when search is a native layer in the stack?

This one's simple. It's night and day. Before, you took the agentic approach and stitched together retrieval, context management, RAG, internet search, whatever you want to call it. The whole industry quickly figured out that baking common knowledge into the model isn't enough anymore. Remember the term "cutoff date"? It used to be central to every model launch: this many experts, this size, this cutoff. It's not a thing anymore, because it doesn't matter that much. General knowledge is now about the ability to reason and draw conclusions, while every meaningful workload is grounded in current data, whether that's internet search, intranet search, or some memory the agent manages. A model's intelligence is really its ability to interpret those grounded facts, not to pull baked-in world knowledge out of its weights.

Getting at the real world knowledge, which is the internet, is close to universal for agentic workloads. The catch is that the internet being what it is, raw access is a problem. You need filtered access, controlled access, and the ability to audit it, and that's what Tavily is. We loved the product, the team, and the vision. As a company with strong infrastructure roots and strong foundational software-engineering roots, we saw the room to collaborate. The internal goal for the team now is to recreate a large-scale web index, to make search genuinely better. The product surface isn't changing much, the great Tavily product and execution stays, but we back it with our infrastructure and engineering muscle. I have people on my team who came from Yandex, Google, and Meta and built large-scale indexes. Tavily touches something like 90% of agentic workloads one way or another. It's another case where one plus one equals three.

Matthew Zeiler is joining as SVP of Research. You've described the research profile you want as "visible in the field, useful in production." What does that mean at Nebius, and how does the research agenda connect to the inference stack?

We've had an internal research team from the very beginning, even before we launched the cloud at any real scale. They were the cloud's first users, and the feedback they gave shaped the product and the requirements, which mattered, because most of the people building the product are infrastructure engineers building infrastructure products. We needed that outside research point of view. The team has always worked on novel AI, training, and post-training, and over the past year a lot of that went into agentic coding and reinforcement-learning environments. One of their best-known releases is a large dataset of RL-ready environments that plenty of top-tier labs have cited in their own model launches. They've also done real work on novel speculative decoding and optimizations, the kind of thing researchers publish at NeurIPS and ICML, which we then push into production pipelines. 

Matt is the person who lives at the intersection of research, business, and product. He's done his own research, he's excellent at reading and interpreting the whole research ecosystem, and he ran a research-heavy machine-learning company for ten years, well before the LLM wave. That blend, grounded in product and immediate business cases but with a researcher's helicopter view, was a big part of the appeal, both with Matt and with the Clarifai team. The deal has closed, he's already with us, and he'll lead the research team and scale it into something bigger than it is today.

Let's get concrete. Walk us through a customer story or two, something founders and engineers reading this can picture themselves in.

I'm a little biased, because I care more about usage patterns than customer stories. But at our scale it's hard to name an industry we haven't touched: fintech, biotech, foundational labs in both text and non-text models, world models, physical AI. I'm a bit skeptical of where physical AI is today, but it's heading the right way; I'm bullish that in five years we'll see humanoids handling our laundry.

What's interesting is what happens when infrastructure stops being the limit. One of our customers, doing large-scale post-training, was at one point hitting our object storage at around one terabyte per second, across almost 100,000 requests per second, to a relatively small training cluster. That number boggles my mind. I've done this for twenty years, and I had to make peace with the idea that a 30-megawatt facility, which a decade ago meant a huge data center serving tons of traffic, might now not be enough for a single client. Five years ago the most demanding storage workloads were measured in gigabytes per second; now we're at hundreds of gigabytes to terabytes per second per client. The primitives haven't changed; S3 was conceived around 2003. But co-locating that same old idea with data-hungry GPUs over the fastest network you can buy, backed by servers full of expensive NVMe, and making it so effortless that the customer just writes three lines of Python, still makes me a little nervous, in a good way.

On inference, we almost never see people running vanilla models. In about 95% of our use cases people are fine-tuning at least a bit. Nobody runs vanilla DeepSeek or vanilla Kimi; they run custom, post-trained open models. That never-ending flywheel of optimization is even more relevant for agents than it was for models, because there are more surfaces to tune: the model, the loop, the tools available. In a way it's easier than with models, because now you have a clear goal the agent has to hit, so the reward is obvious. The difference in how we set up those endpoints is staggering. Take the same model, say Kimi K2, serving large-scale agentic coding versus large-scale deep research. Those look similar on paper but behave nothing alike: input and output lengths, latency demands, how synchronous they are. Deep research is far more asynchronous, you can wait minutes for a result, while an interactive agent has to be fast enough for a human watching the console. Drill down to the hardware and, for token counts that look identical over a span of time, the GPUs needed can differ by an order of magnitude. That's why, for at least another year or two, this stays a field that demands real expertise, with no one-size-fits-all recipe.

For a Series A or B company deciding where to run their workloads, what's the single strongest case for choosing Nebius in 2026?

If it has to be one thing, it's that we're the best product and company in the market, so you go with the best. To be fair, though, Nebius is uniquely positioned, and you can name almost any competitor and I'll tell you why we're better. Against the traditional hyperscalers, those are great companies with far broader portfolios than we could offer, but their AI and GPU business is still relatively small for them. It draws less talent and less focus from decision-makers. 

While some others may be larger in scale in terms of megawatts, our platform is far more flexible. We've been software-first from the very beginning. That's where I believe we're the leading AI cloud, because we've invested in it for three years and we'll keep going. Our benchmark is the hyperscalers, AWS, Azure, Google Cloud, not the rest of the pack people call neoclouds. "Neocloud" has almost become a derogatory term; the umbrella is so wide that serious companies would rather not be lumped under it, and we don't. Against large bare-metal providers, sure, they're cheaper, but the flexibility is lower. We've put years into being flexible and developer-friendly at every level. 

We also love Series A and B companies in particular, because they're the most valuable over time. By the time they reach Series C we already have the connections and the partnership, and we don't see much churn. Start with us at A and chances are you're still with us at B and C.

Looking 18 to 24 months out, where do you want Nebius to be playing, and what does winning at the software layer mean to you?

I don't have one fixed picture, but I know we want to be the industry that supports AI builders. We probably won't be in the AI-builder category ourselves; we don't want to build end-user products, at least not now. We want to support the people who do build products and improve other people's lives with AI. We want to be the independent infrastructure for independent builders.

Right now the market is consolidating, large companies buying other large companies to form even larger ones, and you can already see the gravity wells forming. In 18 to 24 months I think Nebius will still be its own thing, an independent company supporting that never-ending flow of builders. You don't want OpenAI to be the only player, or Anthropic to be the only player. Anthropic is a good competitor for OpenAI, but they need a third company, and a fourth. I'd personally love to see something as good as ChatGPT built on other models by another company; I love competition. We're in a shift that isn't even generational, it's once in five generations, industrial-revolution scale. There's no world where two or three companies control most of it, so an independent layer of infrastructure building blocks, one that lets people with great ideas build something the world actually wants, matters enormously.

Last one. For the founders, engineers, and operators reading this, what's the single best way to engage with Nebius right now?

They can just go to nebius.com. We've put real effort into making it friendly and pointing people wherever they need to go.

Get started with The Nebius Builder Program

Stay up to date on the latest with Nebius, follow them on LinkedIn and X.

Read our past few Deep Dives below:

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.