Conference Presentation, Fireside Chat
The Token-per-Watt Era: What's Next for AI Hyperscalers | Nebius x Accel | RAISE Summit 2026
- Nebius, co-founded and led by Roman Tchernin (Chief Business Officer) and backed by Accel, is celebrating its second anniversary as an independent AI-specialized cloud provider following an IPO.
- Nebius shares have appreciated 10x since the initial public offering.
- The company views the AI adoption curve as just beginning, citing a projected $11 trillion investment in AI data centers between 2024 and 2029.
- GPU capacity remains sold out, driven by demand that is currently concentrated among a select group of large players.
- Hyperscalers and frontier labs (e.g., Meta, Google, Microsoft) account for an estimated 50–70%+ of total GPU demand.
- The secondary demand segment consists of "AI-native" companies (e.g., Cursor, Cognition, Lovable) that build applications from the ground up on AI infrastructure.
- Enterprise adoption is currently limited to two distinct exceptions:
- Digitally native enterprises with strong technical in-house teams (e.g., Revolut, Shopify, ProZ.com, Booking).
- High-frequency trading firms with historical infrastructure expertise (e.g., XTX, Jane Street).
- Roman Tchernin notes that "real" enterprise AI adoption beyond these exceptions has not yet begun.
- Nebius structures its product vertically across three distinct layers:
- Physical Infrastructure (Bare Metal): Targeted at hyperscalers and super-labs with full-stack software capabilities.
- Managed Infrastructure (IaaS): Providing virtualized compute, storage, network, and security; currently the primary layer serving research labs and teams needing reliable managed environments.
- Managed Inference (Navio Stocking Factory): Offering token-based consumption for application builders (e.g., Ligora, LavaBull) and advanced enterprises requiring agentic workflows.
- The company plans to evolve beyond "token" or "GPU hour" pricing toward billing for "authentic outcomes" or solved tasks, decoupling cost from raw compute consumption.
- Nebius prioritizes a vertically integrated approach, owning data centers rather than renting them to enable co-design of hardware and software layers.
- Efficiency ("intelligence per watt") is identified as the critical industry metric, where increased software efficiency paradoxically drives higher overall compute demand.
- New model releases (e.g., deepseek, GLM) have triggered market panic but historically resulted in acceleration across application, adoption, and infrastructure layers.
- The gap between open-source and proprietary models is narrowing regarding task capture, though open-source models remain crucial for cost-effective, data-specific fine-tuning.
- Enterprise value in open-source models is driven not by vanilla baseline quality, but by the ability to fine-tune models on specific proprietary data to solve niche use cases safely and cheaper.
- Looking 2–3 years ahead, the infrastructure trend is focused on lowering barriers to allow non-experts to deploy production-grade AI solutions.
- The long-term vision (5-year horizon) aims to enable a broad population of developers to build and scale complex agentic workflows with built-in security and observability.