Interview, Fireside Chat
Nebius Co-Founder on AI Infrastructure Bubbles | How Price Elastic is Demand for Compute
- Nebius anticipates navigating a capital-intensive environment for the next 6 to 24 months where competitors hold eight times larger capital reserves, potentially dictating deployment speeds.
- The company expects a transition from current "bubble" discussions to a "real adoption" phase requiring tens to hundreds of times more AI infrastructure capacity, with enterprise adoption currently in the first percent of volume and use cases.
- Nebius predicts accelerated growth across four dimensions: capacity expansion, product development, customer portfolio diversification, and capital deployment, noting that each subsequent stage demands increased capital.
- The firm plans to evolve its product stack from selling "megawatts" of physical infrastructure to "GPU hours" of managed infrastructure, and finally to "tokens" for managed inference enabling end-to-end task execution.
- A shift toward open-source and locally hosted models is expected as customers scale to improve economics, though this is not predicted to significantly harm frontier model providers serving complex, unsolved tasks.
- Nebius forecasts that falling intelligence costs will drive increased consumption rather than decreased demand, allowing customers to solve more complex tasks within existing budgets.
- The strategy includes diversifying the customer base from a dozen mega-clients to hundreds for managed infrastructure, thousands for inference, and tens of thousands for agentic applications to mitigate concentration risk.
- Exponential growth in AI adoption is expected for companies building foundational development systems, with specific mention of Revolut, Shopify, Prosource, and Booking.com overcoming "cold start" integration hurdles.
- The market is projected to remain diverse enough to support a mix of frontier, fast/cheap, and specialized models, utilizing agentic workflows that select the optimal model for specific tasks.
- Growth in niche and specialized models for sectors like life science, robotics, and cyber defense is anticipated, driven by the need for optimized inference and infrastructure.
- Nebius expects Europe to develop "sovereign AI" capabilities and "good enough foundational models" to avoid being left behind, relying on local builders to secure megawatts and power.
- Permitting and regulatory delays are viewed as preventing a market glut, with the main bottleneck expected to be unlocked over an 18 to 24-month period.
- The company intends to manage regulatory hurdles and public sentiment by treating data center deployment as a portfolio of projects where delays in one location do not impede overall delivery.
- Nebius is optimistic about "compute in space" becoming viable within three years, predicting more capacity will exist in space than on earth.
- Democratizing developer access is expected to convert tens of millions of people into creators of digital assets, generating new businesses and works that do not currently exist.
- Education systems are predicted to require drastic changes to focus on "how to think" and continuous adaptation, creating a major risk and opportunity for the future workforce.
- The workforce of the next decade will need "empathic communications" and "creativity" rather than just hard technical skills like math and engineering.
- Consolidation into "super models" or "super empires" is identified as the main threat to the business model, potentially reducing demand for diversified infrastructure providers.
- The significant investment from Leo Aschenbrenner is treated as justification and credit to execute, with a commitment to deliver for customers and investors without relaxing growth targets.
- Nebius plans to operate aggressively like a "shark," moving at the pace of the market to remain alive despite capital-intensive challenges.