Dylan Patel — The single biggest bottleneck to scaling AI compute

Dwarkesh Patel asked Dylan Patel of SemiAnalysis, in a podcast episode, how roughly 600 billion dollars of forecast big-tech capital spending, plus fundraising by OpenAI and Anthropic, fits the yearly cost of renting AI data center capacity. Patel's first answer was timing: much of that spending is not for capacity arriving that year. He said Google's 180 billion dollars includes turbine deposits for 2028 and 2029, construction for 2027 and power purchase agreements.

On demand, he extrapolated Anthropic's revenue: about 6 billion dollars added per month would mean 60 billion more over ten months, which at media-reported gross margins implies roughly 40 billion of compute spending, or about four more gigawatts of inference capacity at roughly 10 billion dollars per gigawatt-year. All of this is his estimate and a straight-line extrapolation, not reported contracts.