How Neoclouds Power the AI Boom
The compute substrate behind GPT-5, Claude, Gemini, Llama, Mistral, and the rest.
Two stories from the last twelve months tell you everything. OpenAI signed an $11.9B five-year compute deal with CoreWeave. Mistral signed a multi-year capacity deal with Nebius before its Series E. Neither lab is a small CSP customer anymore — they're anchor tenants in a category that didn't exist in 2022.
Frontier training is mostly on neoclouds now
We tracked publicly disclosed compute deals across the top-20 AI labs in 2025. The split: ~60% of training-cluster-hours ran on neocloud capacity, ~30% on hyperscalers, ~10% on lab-owned infrastructure (OpenAI's own builds, Meta's RSC, xAI's Memphis cluster).
That's a massive inversion from 2022, when ~80% of training ran on hyperscalers. The flip happened almost entirely in 2023–2024 and is now structural.
The two-sided dependency
Labs need neoclouds because neoclouds can ship 10,000-GPU clusters on a 6-month timeline that hyperscalers quote at 18 months. Neoclouds need labs because labs are the only customer base large enough to underwrite the multi-billion-dollar capex of a single new region.
The result is contracts that look more like power-purchase agreements than cloud bills — multi-year, take-or-pay, with the lab effectively financing the buildout by prepaying.
The five biggest 2025 deals
- 1OpenAI ↔ CoreWeave — $11.9B / 5 years, primarily H200 / B200.
- 2Microsoft ↔ CoreWeave — extension to $14B total commitment, multi-region.
- 3Meta ↔ Crusoe — $3.4B, three sites, flared-gas-powered.
- 4Mistral ↔ Nebius — €1.1B / 4 years, EU data residency.
- 5Anthropic ↔ Lambda — undisclosed but reported as "in the billions," multi-year reserved.
What it means for the rest of the market
Anchor-tenant economics squeeze out smaller workloads. If you're a startup wanting 64 H100s for fine-tuning, you'll increasingly buy from a marketplace neocloud (Vast, RunPod) rather than from a CoreWeave or a Lambda — because the big neoclouds prioritise their anchor contracts.
That dynamic is creating a tiered market: tier-1 neoclouds serve frontier labs and enterprise, tier-2 neoclouds serve mid-market AI companies, tier-3 marketplaces serve the long tail. Each tier has different unit economics, different margins, and arguably different multiples. Treat them as distinct businesses, not one category.
Why this is the most under-priced structural shift in AI
Everyone talks about model labs. The compute layer underneath gets a fraction of the coverage but is, on a margin basis, the most leveraged way to own AI's growth. CoreWeave's equity has compounded faster than NVIDIA's since IPO, and that's with all the concentration-risk discount built in. The category as a whole is going to look like the cellular tower REITs of the 1990s — boring infrastructure, enormous compounding.