Meta just committed $27 billion over five years to rent AI infrastructure from Nebius, and that tells you everything about where the compute wars are heading.
The Signal
This isn't a partnership announcement. It's a declaration of dependency. Meta, with all its cash and all its data centers, is paying Nebius $5.4 billion a year for GPU access because they can't build fast enough to keep up with their own AI ambitions. That's the same company that spent $28 billion on capex last year alone, now writing checks to a third-party cloud provider.
Nebius isn't Amazon or Google. It's the rebranded cloud business of Yandex, spun out after sanctions forced Russian tech into creative restructuring. They've been quietly stockpiling H100s and building out European data centers while everyone else fought over limited NVIDIA allocation. Meta's bet here is that Nebius has better access to cutting-edge chips than Meta can secure directly, at least for the next five years.
The math is stark. $27 billion buys you roughly 500,000 H100-equivalent GPUs at current cloud rates, maybe more if Nebius has favorable NVIDIA pricing. That's enough to train multiple frontier models simultaneously and run inference at the scale Meta needs to compete with OpenAI and Anthropic. But it also means Meta is ceding control of a critical piece of infrastructure to an outside vendor, something they've historically avoided.
This confirms what insiders have been whispering: the constraint isn't capital anymore. It's access. The companies that can guarantee GPU supply over multi-year timelines now have pricing power that would make OPEC jealous.
The Implication
Watch who else signs these megadeals. If Google or Microsoft start renting compute at this scale, it means even hyperscalers can't keep pace with AI demand. And if you're building AI products, your real competitive moat isn't the model. It's your relationship with whoever controls the chips.
Source: Bloomberg Tech