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

Enterprise adoption sits at 1%

Large companies are only beginning AI integration, with coding being the sole use case proven at scale so far, indicating massive expansion potential remains.

Infrastructure is not overbuilt

Chernin believes tens to hundreds of times more investment is required to meet future demand as useful AI adoption moves beyond early startups into the enterprise.

Consolidation is the primary threat

The main risk to Nebius is excessive market consolidation among hyperscalers, not oversupply or a collapse in compute demand.

DeepSeek panic drove record sales

When DeepSeek emerged, Nebius stock dropped 40% in a week yet the company recorded its best sales week ever as customers realized cheaper inference enabled viable production workloads.

Demand follows Jevons paradox

Every time intelligence becomes cheaper, consumption increases rather than decreases because companies can finally economically solve tasks that were previously budget-prohibitive.

Frontier models move upmarket continuously

OpenAI and Anthropic are not threatened by open source because they continuously advance to unsolved complex tasks while open-source models optimize previously solved use cases.

Layer 1: Bare metal megawatts

Physical infrastructure sold to large labs like Meta and Microsoft, measured in megawatts and serving only dozens of customers who bring their own software stack.

Layer 2: Managed cloud GPU hours

Infrastructure-as-a-service with virtualized compute, storage, and networking for hundreds of research teams who pay for GPU hours rather than raw hardware.

Layer 3: Token-based inference

Managed inference platform serving thousands of vertical AI companies who consume tokens via API without managing clusters, GPUs, or model optimizations.

Layer 4: Agentic task execution

Future optimization layer where developers pay for end-to-end task completion rather than specific models or tokens, automatically routing to optimal models for cost and quality.

$2B program vs 8x larger competitors

Nebius's 2025 capex budget is $2 billion compared to hyperscalers spending roughly $16 billion, requiring disciplined capital deployment in a capital-intensive arms race.

10x capacity would find buyers

Current demand meaningfully exceeds supply, with the primary constraint being physical build-out speed including regulatory approvals and supply chain logistics rather than customer acquisition.

Customer diversification is existential

The company actively limits revenue concentration from any single hyperscaler, balancing low-margin bare metal deals with higher-margin managed services to reduce dependency risk.