Overview of 20VC with Nebius Co-Founder Roman Chernin
In this episode of 20VC, Harry Stebbings speaks with Roman Chernin, co-founder of Nebius, about the state of the AI infrastructure market, whether the sector is in a bubble, and how open-source models are reshaping demand for compute. Chernin argues that AI adoption is still in its early stages, that cheaper intelligence tends to drive more consumption rather than less, and that Nebius is positioning itself as a full-stack infrastructure provider across compute, managed cloud, managed inference, and eventually agentic workflows.
Key Takeaways
-
AI infrastructure is still early, not bubbled out
- Chernin does not believe the market is in a true bubble.
- His core argument: most enterprises are only using AI in a tiny fraction of workflows today.
- He sees current adoption as the “first steps” of a much larger transition.
-
Cheaper models increase demand
- He repeatedly points to a Jevons paradox effect: when intelligence gets cheaper, companies use more of it.
- Lower inference costs unlock new use cases, broader production deployment, and larger workloads.
- The market doesn’t shrink when models get cheaper; it expands.
-
Nebius is building a multi-layer infrastructure stack
- The company is moving from raw capacity into higher-value software and managed services.
- Their strategy is to serve customers at multiple levels of abstraction, not just as a GPU supplier.
-
Consolidation is the biggest long-term risk
- Chernin says Nebius’s biggest threat is not competition, but a world dominated by a few giant AI platforms.
- A more diversified ecosystem creates more opportunities for infrastructure providers like Nebius.
Nebius’ Four-Layer Product Strategy
Chernin breaks Nebius into four evolving layers:
1) Capacity
- Core physical infrastructure: data centers, power, racks, GPUs.
- This is where the company competes with hyperscalers and other large-capital players.
- Demand is strong, but the main constraint is execution and build speed.
2) Multi-Tenant Cloud
- Traditional cloud-style infrastructure for research-heavy teams and builders.
- Customers get compute, storage, networking, security, and observability without managing the underlying physical stack.
- Revenue is typically measured more in GPU hours than raw megawatts.
3) Managed Inference
- Nebius’s TokenFactory product helps customers run open-source or customized models in production.
- This layer is aimed at enterprises and vertical AI companies that want:
- lower costs,
- model flexibility,
- better reliability,
- and less infrastructure complexity.
- Chernin says this is where customers increasingly move once they’ve proven a use case.
4) Agentic Layer
- The next step is full workflow optimization for agents.
- Instead of customers thinking in tokens or specific models, the platform would decide:
- which model to call,
- how much context to use,
- whether to use one model or several,
- and how to optimize for reliability and cost.
- This would make Nebius closer to an orchestration and optimization engine than a simple compute provider.
Open Source vs Frontier Models
Chernin makes a strong case that open-source models are not destroying the market for OpenAI, Anthropic, or Nebius:
- Frontier models still offer the best general capabilities.
- Open-source models are increasingly useful when:
- the use case is known,
- economics matter,
- customers want to fine-tune,
- or they need specialized behavior.
- The shift to open source is often about optimization, not replacement.
- He argues frontier model providers keep moving the frontier forward, so the total market keeps expanding.
Important distinction
- The debate is not “closed vs open” in a binary sense.
- Customers increasingly want a portfolio of models:
- smartest model for hard tasks,
- cheaper/specialized models for routine tasks,
- and agentic routing to choose among them.
Demand, Pricing, and Elasticity
A major topic in the conversation is how elastic demand for compute really is.
- Nebius recently raised prices and still saw strong pipeline pressure.
- Chernin says pricing is only part of the equation:
- the real issue is TCO (total cost of ownership),
- reliability,
- uptime,
- orchestration,
- and how many effective tokens or workloads can actually be delivered.
- In inference, better optimization can dramatically reduce real customer cost even if nominal GPU pricing rises.
- He believes there is meaningful elasticity, but not infinite elasticity.
How Nebius Differentiates Itself
Chernin says Nebius is differentiated by full-stack integration:
Downstream integration
- Building data centers, racks, servers, and infrastructure directly.
- This gives Nebius more control over cost and speed.
Upstream integration
- Building higher-level products that match customer needs as they evolve.
- This lets Nebius serve more than just the small set of customers that only want raw infrastructure.
Business model advantage
- By moving up the stack, Nebius can serve:
- a handful of mega-scale customers at the physical layer,
- hundreds at the managed cloud layer,
- thousands at the inference layer,
- and potentially tens of thousands in the agentic future.
Customer Behavior: What Changes in Practice
Chernin highlights a shift that he thinks is underappreciated:
- The key change is not simply “training to inference.”
- It’s that companies are building real AI products with real economics.
- Once an enterprise starts shipping AI features, it generates:
- data,
- feedback loops,
- evaluation needs,
- and continuous model improvement.
- This creates a flywheel:
- deploy inference,
- generate usage data,
- evaluate quality,
- tune or swap models,
- improve the product,
- scale further.
He says this is already happening at companies like:
- Revolut,
- Shopify,
- Booking.com,
- Prosus,
- and others with strong product/engineering cultures.
Competition, NVIDIA, and Power Dynamics
On NVIDIA, Chernin is pragmatic:
- He does not frame the relationship as adversarial.
- His view is that respect is earned engineer-to-engineer.
- Nebius focuses on:
- building good products,
- having strong engineers,
- and executing reliably.
- In his view, that’s how to build a healthy relationship with NVIDIA and maintain leverage.
Capital, Buildout, and Regulation
Chernin emphasizes how capital-intensive the business is:
- AI infrastructure requires massive upfront investment.
- More capital would let Nebius build faster and serve more customers.
- But timing matters:
- in the next 6 months, money can’t solve everything,
- over 18–24 months, it can meaningfully accelerate growth.
- He also notes:
- permitting,
- regulation,
- and community pushback can slow data center expansion.
Public backlash
- He acknowledges rising public scrutiny around AI infrastructure and data centers.
- His response: the company must work closely with communities, local authorities, and regulators.
- He compares the dynamic to early Uber expansion: new technology often creates friction before it becomes accepted.
Future Outlook
AI and the workforce
- Chernin thinks many future jobs will be created by the democratization of building software and AI products.
- He expects more people to be able to turn ideas into working digital products.
- That will create new industries and new roles we can’t fully predict yet.
Education and skills
He believes future success will rely less on hard skills alone and more on:
- empathy
- communication
- creativity
His advice to his teenage daughters:
- learn how to understand and communicate with people,
- be creative,
- and adapt continuously.
Space data centers
- He’s open-minded about the idea of compute in space.
- His view is basically: with enough smart people working on it, it could happen.
- He doesn’t claim it will happen soon, but he does not dismiss it.
Notable Quotes / Ideas
- “We are just at the beginning of this amazing moment.”
- “When intelligence gets cheaper, we consume more of it.”
- “The biggest threat for Nebius is the world becoming too consolidated.”
- “You need to build what you build. You need to tell your story.”
- “It’s like a shark. You’re alive when you move.”
Bottom Line
This conversation presents Nebius as more than a GPU supplier: it’s a company trying to become a full-stack AI infrastructure platform as the market moves from raw training compute to inference, specialized models, and agentic systems. Chernin’s central thesis is that AI demand is still early, cheaper models will expand usage rather than compress it, and the real winners will be those who can reduce friction for customers while staying fast, reliable, and economically efficient.