Roman Chernin on AI Compute Demand, Inference and Infrastructure

Nebius co-founder and chief business officer Roman Chernin tells Harry Stebbings that he does not believe AI infrastructure is a bubble. He expects much more compute and enterprise adoption, while acknowledging that his position as an infrastructure provider gives him a reason to believe in that growth.

He sees coding as one early use case that has begun to work at scale and says many companies are still at an early stage of AI adoption. His assessment of current adoption and expectation of many future applications are observations and forecasts, not independently measured industry totals.

Stebbings asks whether cheaper, locally hosted open models could hurt frontier labs and infrastructure providers. Chernin says some customers already look for alternatives as their products scale, because they want to improve cost, quality or both for a specific use case.

He distinguishes the initial breadth of frontier models from open-weight models that customers can adapt or fine-tune. In his view, a specialized model may perform better for a particular workflow, while frontier providers continue to tackle new and more difficult tasks. That coexistence is his market thesis, not a guaranteed outcome.

Stebbings questions whether continually moving to new problems is enough to justify very high model-company valuations. Chernin argues that many problems remain unsolved and expects room for both frontier capabilities and specialized systems; the interview does not establish future revenue or returns.

Chernin recalls that Nebius shares fell around an earlier DeepSeek release while the company experienced what he describes as an exceptionally strong sales week. He says cheaper inference encouraged customers to run workloads that had previously been uneconomical. This is an anecdote from the company's experience, not proof that lower prices always increase total spending.

He organizes Nebius's work around four dimensions: physical capacity, product, customers and capital. Building data centers requires power, equipment, permissions and supply chains; he says the team wants to move faster on all of them.

He describes a physical infrastructure layer for large labs and hyperscalers that want substantial compute capacity and bring much of their own software. Delivering at that scale is difficult, he argues, even when a customer primarily wants raw infrastructure.

A second layer is a managed, multi-tenant cloud with compute, storage, networking, security and observability. Customers can provision clusters and manage their own workloads without building the underlying infrastructure; the commercial unit is often GPU hours.

A third layer is managed inference, including Nebius Token Factory, where customers consume model outputs rather than manage GPU clusters themselves. Chernin says it is aimed at product builders and enterprises that want to use open-weight or specialized models.

He sketches a possible fourth layer for agents: a system could choose models, manage context and optimize an entire task for quality and cost. He explicitly describes this as speculative product thinking, not a capability Nebius has already delivered.

Asked whether Nebius could sell ten times its current capacity overnight, Chernin says not overnight but believes there is demand for much more. He wants a portfolio spanning large infrastructure customers and smaller users of managed services; this is a company assessment, not a contracted demand figure.

Customer concentration is a central strategic question, he says. Very large customers may require only physical capacity, whereas a software platform can serve more enterprises and product builders. He wants a diversified portfolio but does not give a specific target or audited customer breakdown.

Stebbings asks whether doubling prices would affect demand. Chernin says Nebius had raised prices and still saw pipeline pressure, but he does not know the market's precise balance. Inference economics eventually constrain customers: if the cost of serving users is too high, their products cannot grow sustainably.

He says the nominal GPU-hour price alone does not determine a customer's total cost. Utilization, uninterrupted runtime, model configuration and the number of useful tokens produced can change the effective economics. His comparisons are general operating claims, not guaranteed savings for every workload.

For the managed-cloud layer, his main question is what customers need as work shifts from model training to inference and agents. Nebius cannot build everything, he says, so it must choose where its products can provide distinctive value.

Chernin says production inference creates new requirements: collecting and observing data, evaluating outputs, and using feedback to improve a model or application. He describes a potential improvement loop rather than treating an inference workload as simply training on a different GPU.

He believes the barrier to building AI products is falling as platforms hide some infrastructure and optimization complexity. His goal is to let developers focus on users and use cases, while the platform handles more of the underlying service.

Asked to compare Nebius with another neocloud, he describes the company's approach as vertical integration: data centers, servers and platforms below, with managed products above. He believes this can support cost control and a broader customer base; the competitive advantage is his assessment as a company executive.

He explains a common managed-inference problem: a team that has built on a closed API may want to try an open-weight model for cost or customization, but deploying it reliably requires orchestration, caching, monitoring and optimization. Token Factory is Nebius's proposed service for handling that work.

Stebbings cites a claim of inference-cost reductions of up to 70%. Chernin describes techniques such as distillation, speculative decoding and caching. The quoted saving is a maximum product claim, not a universal result; actual cost and quality depend on the model, baseline and workload.

He says frequent model releases make it valuable for customers to test alternatives without rebuilding infrastructure each time. A managed platform can help with that work, in his view, though the interview does not independently verify how smoothly every migration works.

Asked whether the pace of model releases will continue for five years, Chernin says he does not know. He expects more specialized systems across modalities and gives a cyber-defense model as an example of tailoring an existing foundation model to a particular task.

He describes Revolut as a customer that initially relied heavily on closed models and invested in evaluation before adopting alternatives for some use cases. The budget mix and customer experience are Chernin's account, not an independently confirmed disclosure from Revolut.

The lesson he draws is that teams need metrics and an evaluation process before changing models in production. Without that foundation, a cheaper alternative could damage output quality; with it, he believes teams can experiment and deploy faster.

Asked how much more Revolut might spend in three years, Chernin declines to give a number. He expects advanced enterprises' AI budgets and use to grow quickly, but future customer spending remains uncertain.

He rejects a simple closed-versus-open choice: frontier, specialized and lower-cost models can each serve different tasks. He imagines future orchestration selecting among them, while noting that reliable deployment and economics matter alongside model capability.

On sovereign AI, Chernin says regions need access to capable models and a strong community of builders, not just megawatts of capacity. He argues that infrastructure follows demand from researchers and product companies; that is his view of how the ecosystem should develop.

Asked where he would invest across infrastructure, models and applications, he acknowledges Nebius's own infrastructure focus but praises founders building end-user products. They take the risk that customers may not want what they make, he says, and help create demand for the rest of the stack.

Stebbings asks about negotiating with Nvidia. Chernin says his approach is to build a strong product and engineering team, and he believes mutual technical respect supports the relationship. He notes that Nvidia may have a different view; he does not claim equal bargaining power or disclose contractual terms.

He returns to the four dimensions of scale, product, customers and capital. Cloud is a post-sale business in his view: signing a contract creates a promise, and customer-facing engineers must then help the buyer make the service work.

With unlimited capital, he says he would build more data centers and deploy GPUs faster. He cites Nebius's capital program and larger hyperscaler budgets, but the figures are his company-level discussion and should not be treated as verified future expenditures.

Stebbings asks whether permitting delays may prevent an oversupply. Chernin says bottlenecks differ by time horizon: more capital cannot instantly create capacity in six months, but could enable parallel projects over a longer period. He does not quantify a safe level of industry supply.

Nebius, he says, tries to secure power and land before later stages of construction and GPU deployment. Each stage needs additional capital, and preparing a portfolio of projects can help when individual sites are delayed. This is the company's planning approach, not a promise that every site will proceed.

Asked about community opposition to data centers, Chernin says developers must engage local authorities and address concerns. He describes managing a portfolio of possible sites and acknowledges that Nebius needs to improve its outreach; the interview does not independently assess the impacts of its projects.

On space-based data centers, he says many capable people are trying to solve the problem but acknowledges substantial challenges and uncertainty. He does not predict that space will replace terrestrial capacity on a particular timeline.

In the quick-fire round, Chernin hopes easier software creation will let more people turn ideas into useful products. He also sees a challenge for education: helping people learn, think and adapt as tools and occupations change. These are expectations, not measured employment outcomes.

For his daughters, he emphasizes empathy in communication and creativity, while admitting he cannot know precisely which skills will be needed. His advice reflects his personal priorities rather than a prescription for every student.

He says the biggest threat to Nebius is excessive industry consolidation: if a few companies control the market, independent providers might be confined to supplying physical capacity. He hopes a diverse set of builders and customers will preserve more choice, which also aligns with Nebius's business interest.

Asked about investor Leo Aschenbrenner's disclosed interest, Chernin says investment and customer contracts create an opportunity and an obligation to deliver, not a reason to assume success. He does not offer a forecast of share returns.

Chernin says the company should recognize the team's work more often while continuing to execute in a fast-moving market. He closes by comparing the business to a shark that must keep moving, and Stebbings thanks him for the conversation.