Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028

Open original

Dylan Patel and Dwarkesh Patel discuss how OpenAI and Anthropic are on track to command most of the world's usable computing capacity within a few years by outbidding everyone else, thanks to their ability to monetize inference at much higher margins than the raw cost of compute. They also explore whether the $10+ trillion in AI infrastructure spending by decade's end could trigger a sovereign debt crisis.

  • OpenAI and Anthropic are capturing 40-50% of new compute capacity next year (up from 30% this year), and at current growth rates will control most of the world's usable computing power by end of 2028, since they're deploying the most efficient latest-generation chips while competitors use older hardware.
  • These labs have flipped from venture-funded losses to profitability by achieving 50x revenue per megawatt of compute (Anthropic), allowing them to reinvest all profits into training and continuously outbid other companies for scarce compute capacity.
  • The concentration of compute in two companies raises questions about whether massive hyperscaler debt could drive up interest rates globally, push non-AI countries into bankruptcy, and whether any force can counteract the economics pushing toward centralization.

Dwarkesh Patel interviews Dylan Patel about how AI lab economics will reshape global compute distribution over the next few years. The core story: OpenAI and Anthropic are capturing an outsized share of new compute capacity and converting it into revenue at unprecedented rates. Anthropic generates up to $50 million in revenue per megawatt of inference capacity, compared to a base compute cost of $10-15 million per megawatt. This margin advantage lets them reinvest profits into training compute rather than relying on venture funding. Both labs started 2024 around 2 gigawatts each and have already scaled to above 5 gigawatts—roughly tripling capacity in a single year. They're taking about 30% of all new compute added globally right now, and that's expected to jump to 40-50% next year. By the end of 2025, more than half of the world's incremental new compute will likely belong to these two companies. This centralization accelerates because the labs can simply outbid everyone else. They have the revenue to pay higher prices for compute, which means suppliers prefer leasing to them over smaller players. New entrants like SpaceX are building capacity specifically to lease to OpenAI and Anthropic. Meanwhile, the labs themselves are starting to build proprietary chips—OpenAI developing their own, Anthropic using Google's TPUs. The shift from inference to training is also happening; as models become cheaper to run, the labs are redirecting freed-up inference capacity toward R&D and new model development. The conversation touches on bigger structural questions: whether $10 trillion in AI capex by 2030 could trigger a sovereign debt crisis if hyperscaler borrowing drives up global interest rates and bankrupts countries without AI exposure, and whether anything can actually counter the forces pushing toward centralization—economies of scale in training, compute scarcity, and eventually continuous learning. Dylan and Dwarkesh couldn't resolve whether any counterforce exists. China receives less than 10% of new global compute, though its labs apparently need less to remain competitive.

Ask about this article