← @bgSponge

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

Summary

Dylan Patel tells Dwarkesh that Anthropic and OpenAI already take about a third of new AI compute and, if they keep outbidding everyone on revenue per megawatt, will control most usable FLOPs by late 2028.

  1. Two labs take most incremental compute by 2027–28.

    OpenAI and Anthropic go from about 30% of new watts this year to 40–50% next year; newer chips make that half of incremental supply most of the world's usable FLOPs by end-2028.

    The labs are going to allocate less and less compute to inference over time.

    every force is screeching towards centralization.

  2. Outbidding, not ownership, is the mechanism.

    Anyone can make money at $10–15M/MW today; labs already generate far more, so they can pay $25–50M/MW and pull rented capacity from SpaceX, Meta, and clouds.

  3. Inference share falls as models get more valuable.

    Non-consensus: labs already steer more of each new megawatt to R&D because internal progress beats selling the same tokens, even as public-company investors would prefer inference profit.

  4. China stays compute-poor relative to the frontier.

    Sub-10% of incremental watts, weaker domestic chips, and much smaller clusters; Chinese labs currently need less because pretrain sites max out around 200 MW while the rest is research.

    China domestically still continues to have sub-10% of incremental new compute.

  5. Capex crowds out the rest of the economy.

    SemiAnalysis models about $11T of AI capex through 2029 with about $5T of new credit; higher rates, sovereign defaults, and collapsing non-AI multiples are the political brake on takeoff.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →