Valida Pau’s analysis in The Information puts numbers on how far Anthropic’s infrastructure commitments have run since October: agreements covering at least 14.8 GW of compute capacity, with total spending over the next decade potentially reaching $517 billion.
Start with the power figure, because it is the one that converts into physical reality. 14.8 GW is grid-scale. It is the kind of load that shows up in a utility’s integrated resource plan and forces decisions about generation, transmission, and interconnection queues that take years to clear. Contracting for it is not the same as having it energized, and the gap between those two states is where most of the execution risk sits.
The obligation is the point
Deals of this shape are usually take-or-pay in substance. The buyer commits to a capacity ramp; the developer finances the build against that commitment. Which means the spending figure is less a forecast than a schedule of obligations, and it holds whether or not revenue arrives on the assumed curve.
That is the structural bet every frontier lab is now making. Revenue growth has to outrun a contracted cost base that was locked in years earlier. For Anthropic specifically, enterprise and coding workloads have been the strongest part of the demand picture, and those are the loads that would need to keep compounding to make the arithmetic work.
Who else is exposed
Numbers like these are why the memory, networking, and power equipment names have been trading on AI capex headlines rather than on their own end markets. A commitment schedule at this scale reaches back through HBM supply, advanced packaging, transformers, switchgear, and gas turbine backlogs, most of which are already sold out well into the future.
The figure to watch is not the headline dollar amount. It is how much of the 14.8 GW actually comes online on schedule, because that is where a contracted plan meets a substation.