Jun 8, 2026, 10:52 AM ETInterview, Fireside Chat

Nebius Co-Founder on AI Infrastructure Bubbles | How Price Elastic is Demand for Compute

Nebius operates in a capital-intensive AI infrastructure market with a 2025 CapEx program of $20–25 billion.
Nebius's capital expenditure is roughly eight times smaller than that of its hyperscaler competitors.
The company faces a critical 6–12 month execution bottleneck where capital cannot accelerate infrastructure build-out due to regulatory, permitting, and supply chain delays.
The primary long-term threat identified by co-founder Roman is excessive market consolidation, which could reduce the addressable customer base to a few super-entities.

Roman dismisses the notion of an AI infrastructure bubble, arguing the industry is only at the beginning of widespread "useful AI" adoption.
Current enterprise AI adoption remains below 1% in volume and use cases, indicating significant room for growth.
The first widely successful AI use case at scale is coding, which began maturing only a few months prior to the interview.
Cheaper intelligence (Jevons Paradox) drives increased consumption rather than reduced demand, as companies can economically solve previously infeasible tasks.
Nebius experienced its highest sales volume in company history during a week when its stock dropped 40% due to the "DeepSeek" open-source model release.

Nebius defines its growth through a four-layer product evolution moving up the stack:

Layer 1 (Bare Metal): Raw infrastructure sales measured in megawatts, serving a small group of hyperscalers (e.g., Meta, Microsoft) who bring their own software stacks.
Layer 2 (Multi-tenant Cloud): Managed infrastructure for research teams, measured in GPU hours, offering virtualized compute, storage, and networking.
Layer 3 (Managed Inference - Token Factory): A service for vertical AI companies and enterprises, measured in tokens, allowing users to deploy and tune open-source models without managing underlying optimization (e.g., caching, distillation).
- The Token Factory platform supports 60+ open-source models and claims to reduce inference costs by up to 70% via system optimization.
Layer 4 (Agentic Orchestration): Future focus on end-to-end task execution where the platform autonomously selects models and manages workflows, measured in task outcomes rather than tokens.

Nebius aims to diversify its customer portfolio to avoid over-reliance on a "dozen" hyperscale clients, targeting thousands of enterprise and product-focused customers.
The company argues that while raw GPU pricing varies ($3–$5), the Total Cost of Ownership (TCO) for optimized platforms can be an order of magnitude better for customers.
Enterprises require robust "foundation engines" (CI/CD for AI, evaluation pipelines) to transition from closed models (OpenAI, Anthropic) to specialized open-source models.
Non-AI native enterprises like Revolut are currently in a "cold start" phase of foundational investment before achieving exponential AI budget growth.
The company maintains that the market is large enough to support both frontier closed models and specialized open-source models without one erasing the other.

Nebius positions itself as a key enabler for European "sovereign AI" by providing infrastructure and platforms to local builders (e.g., Mistral, Black Forest Labs) rather than just competing on power capacity.
The company believes the US and China dominate model development, creating a demand for a European ecosystem of builders to generate local demand for infrastructure.
The relationship with NVIDIA is described as an engineering-driven partnership based on mutual respect for technical execution rather than transactional dynamics.
Roman notes that permitting and regulatory hurdles, while difficult, have prevented an immediate glut of data center capacity, helping stabilize the market.

Roman predicts that AI infrastructure in space will become a reality within three years, driven by the convergence of smart engineering teams.
The job of "developer" is expected to become democratized, allowing tens of millions more people to convert ideas into digital assets, creating new workforce demands.
Educational focus for the next decade must shift from hard skills (facts, math) to soft skills: empathy, communication, and creativity.
Leo Ashenbrenner recently disclosed a 5.3% stake in Nebius, which the company views as external validation rather than a license to relax operations.
The company operates on a "shark" mentality: the business is only alive as long as it continues to move and execute.