Nebius: “For every GPU, there are three to four customers ready to buy it”

In 2026, AI model startups looking for training or inference will have a plethora of options available on the market to choose from. 

For example, CoreWeave (US AI cloud / GPU hyperscaler),  IREN (the former Bitcoin miner pivoting into AI cloud), and Nebius (European AI-native cloud) are all fighting in the same market today. 

But Nebius, headquartered in the Netherlands, wants to be seen as something bigger than a GPU-rental company.

Attempting to stave off the sweltering Paris heat, I sat down with Marc Boroditsky, Chief Revenue Officer at Nebius, during the packed RAISE Summit, to find out why the company is positioning itself as “the only AI-native hyperscaler.” It’s laying claim to the moniker of a European-rooted, full-stack infrastructure company built for the AI era, rather than the old cloud era. And he hinted that many of the newest AI startups that have raised in Europe recently may well be using Nebius for training, though he demurred on the details. 

Boroditsky’s core argument is that demand for AI compute is still far outstripping supply, and even massive frontier AI model companies building their own data centres are still coming to Nebius to buy compute. Inference is becoming the fastest-growing part of the business. He also thinks Europe has a real opportunity in ‘scientific AI’, provided it can solve the capital gap.

“We’re not just a GPU-as-a-service company,” Boroditsky said. “We’re actually on a path to be the next hyperscaler,” referring to how Nebius is putting AI into every facet of its operations, even DevOps. 

The company’s current customer base still skews heavily towards AI-native companies, including “many of the most well-respected AI-native companies,” he said. But, he said, the next phase is enterprise adoption, where the conversation is shifting away from raw usage and towards measurable value.

“There’s been reasonable discussion around rationalisation,” Boroditsky said. “‘Token-maxing’ in and of itself is a flawed strategy. But token-maxing where you know value is being created, where you can actually point to bottom-line contribution or top-line increase, is a great strategy.”

The mindset, he argued, has to move from “token-maxing” to “value-maxing.”

For Nebius, the pitch is that legacy hyperscalers were built for a different computing era.

“We’re building a full-stack, vertically integrated hardware, software and services AI platform that is built for the AI-native requirements through the enterprise requirements,” he said. “Unlike AWS, which was built for a different era, this is for cloud CPU compute.”

Boroditsky said the difference shows up in the developer experience. On traditional cloud platforms, AI engineers are often forced into infrastructure work they do not want to do. On Nebius, he claimed, much of that operational layer is automated.

“If you ask that same person, ‘What’s the experience on Nebius?’ they’ll tell you, ‘I don’t do any DevOps. It’s actually done for me,’” he said. “We’ve automated all the DevOps. We’re taking care of managing spin-ups, spin-downs, the health of the environment.”

But a key question for any AI infrastructure provider is whether the largest model companies will simply build their own compute and bypass external platforms. Boroditsky said Nebius is not seeing that happen.

“The amount of demand… it’s crazy,” he said. “We’re not seeing it affect us at all.”

In fact, he said, even AI model companies that are building their own, vertically-integrated infrastructure are also coming to Nebius.

“They are also recognising that building data centres is a lot harder than they may have imagined,” he said. “There’s a lot to standing up infrastructure and operating infrastructure. They have to ask themselves: is that their primary purpose?”

He pointed to Nebius’ deals with Microsoft and Meta as evidence of how constrained the market remains.

“The hyperscalers and foundational model companies are at our doorstep all the time,” he said.

By the end of the year, Boroditsky said, Nebius expects its compute footprint to be roughly evenly distributed between the US and Europe. But customer demand still tilts heavily to the US.

“We actually have a lot of our American customers on European compute,” he said. “Because they’re looking for it anywhere they can get it.”

And Sovereign AI offerings are also in its sights.

That European footprint is increasingly central to the company’s positioning. In a geopolitical environment beginning to be shaped by export controls, frontier-model regulation, and the sovereign AI debate, Boroditsky said Nebius can offer customers a European alternative to US-owned hyperscalers.

“We’re uniquely positioned to help them,” he said. “You can look at all the IP, with the exception of the hardware that is manufactured in the US, it is European. You look at our software, you look at our services. That’s all European.”

He described Nebius as “a more European-reliable alternative to the hyperscalers, which are all US-owned and controlled.”

On models, however, Boroditsky argued against a narrow winner-takes-all view. Nebius wants to support a broad range of AI models rather than bet everything on one frontier layer.

“This is too early in the technology lifecycle to say that we’re done,” he said. “We need to be inspiring, supporting, investing in the innovation that has yet to take place.”

He said Europe may have a particular opportunity in scientific models, given its academic and research base.

“Europe is uniquely positioned to actually build a lot of the scientific models because of the heritage here with the basic sciences,” he said. “It has a unique opportunity to potentially be leaders in science-based models.”

One trend he is seeing is that companies outside the obvious AI lab category are now considering building their own models. He cited a conversation with the CEO of a gaming company in Europe which was planning to build a model of its own.

“When a gaming company is saying, ‘I’m going to go down the path of building my own instead of utilising one of the off-the-shelf models, that’s saying quite a bit,” he said.

Part of that is about controlling IP, but it is also about strategic positioning.

“It’s getting a horse in the race,” Boroditsky said. “It’s transforming their business to become AI-oriented, but also transforming their business to become a service provider to other gaming companies.”

On the split between training and inference, Boroditsky said inference is currently Nebius’s highest-growth area, even if it is not yet the majority of the platform.

“Inferencing is not the majority of our platform today,” but, he added, “inferencing is the highest growth area in our business.”

Nebius sees inference demand coming from three directions: customers running their own inference, inference platforms built on Nebius, and Nebius’s own managed inference service.

For AI-native companies, the managed service is a way to expose a model to customers through a private or public endpoint. For software vendors and enterprises, Nebius is helping implement open-source or specialist models, often alongside or instead of foundational models.

“They start with us using inferencing on open source,” Boroditsky said. “Then they oftentimes end up using us for retraining, and then sometimes they end up using us for model creation.”

His broader view is that inference alone is not the end-state business.

“Inferencing by itself is not a business,” he said. “Inferencing is a feature of a bigger set of requirements” in other words, part of a bigger business offering.

Energy remains one of the biggest constraints for AI infrastructure. Boroditsky said Nebius approaches data-centre projects locally, looking for jurisdictions where local authorities have both the will and the authority to support a project, especially around power.

“We want to do projects where we’re a net benefit to the community,” he said. “That we’re a good partner, neighbour, participant in the community, because we’re going to be there for a long time.”

Admittedly, that approach still can’t insulate it from the concerns of some local communities, most recently the new Nebius AI factory data centre off Lakeshore Parkway in Birmingham’s Oxmoor Valley neighbourhood, where residents have been mounting small protests about the project.

But, Boroditsky said Nebius tries to avoid projects where local ratepayers or consumers might suffer negative consequences.

“In a lot of situations, we’re adding power to the facility; we’re adding power to the grid,” he said. “In some situations, we’re actually putting in place capabilities that didn’t exist before, like redundancy or surplus capacity.”

Boroditsky claimed Nebius is not facing the same power issues as some competitors because it is disciplined before signing projects.

“When we say we’ve got committed power of the four gigawatts, that’s committed to be able to be delivered,” he said. “That’s not hypothetical.”

Demand, however, a sort of bottleneck. “We have three to four times the demand to the capacity to buy,” he said. “For every GPU, there’s three to four customers that are ready to buy it.”

He said the demand is no longer just venture-backed startups burning through investor capital on training models. 

“It’s expanding pretty dramatically to more and more inferencing and revenue-producing activity,” he said.

On Europe’s AI prospects, Boroditsky was optimistic about the ability of European talent to innovate, but raised the concern that it still does not have the capital it needs to compete.

“Many of the AI engineers in the world came from Europe,” he said. “They were trained in your universities. They’ve been doing the research for a long time.”

Nebius already counts European model-creating companies among its customers, he said. But model creation itself remains hard to package into a repeatable product because the patterns differ so much by team and use case.

“Model creation is very unique to the model team,” he said. “Today, the patterns aren’t repeatable.”

He thinks Europe can still produce state-of-the-art AI model companies, but only if it can finance them properly.

“My suspicion, and this is just an opinion, is that state-of-the-art model creations happen in Europe,” Boroditsky said. “The real question is, are European companies getting the same opportunities for capital formation?”

“The amount of capital that’s getting raised in the United States against the same requirements is a dramatic multiple of the amount of capital being raised,” he said. “Capital formation is the potential throttle that the European market faces.”