← @bgSponge

We’ve Built 1% of the AI Compute the World Needs

Summary

Patrick Toulme argues the world has built only 1% of the AI compute required to support a future where users command agent swarms rather than managing individual models.

  1. Agent swarms drive massive demand.

    Users will interact with a single AI that orchestrates hundreds of subagents in parallel, shifting the compute requirement from one agent per human to billions of always-on agents.

    The human stays in a single conversation; the machine fans out into hundreds.

  2. Silicon demand is effectively infinite.

    Because demand sits hundreds of times above combined manufacturer output, vendors are not competing for a fixed market share of customers.

    When demand sits anywhere from a couple times to many hundreds of times above the combined supply

  3. Physics limits chip production.

    The primary bottleneck for vendors is not logic dies or market competition, but rather the physical constraints of CoWoS packaging and HBM memory.

    The real competition isn’t vendor versus vendor for customers. It’s all vendors versus physics

  4. Efficiency gains trigger Jevons' paradox.

    Making compute cheaper through model distillation does not reduce demand; instead, it encourages users to run sixty agents instead of one.

    make something cheaper to consume and total consumption goes up, not down.

Highlights

  • The real competition isn’t vendor versus vendor for customers. It’s all vendors versus physics for CoWoS and HBM, the shared, sold-out chokepoint upstream of all

  • Make an agent 60× cheaper and nobody runs one agent and banks the savings. They run sixty. Then six hundred.

  • The chips are useless without the watts, and the watts are harder to get than the chips.

  • The binding constraint on AI chip production is not logic dies; those are abundant, and AI silicon vendors just outbid everyone else for leading node

  • The world needs a hundred times more compute than exists.

Read the original ↗Save to your Sponge

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

More from @bg

All readings →