The Book

• GPU rental — What a chip costs by the hourWhat an hour on a chip cost through the month.p.5
• Models — Open-weight models released each dayHow many open-weight models were put out each day.p.5
• Models — Which model people prefer in a blind voteWhich model people picked in a blind vote.p.6
• Models — Where an AI app sits in the free chartWhere the AI apps sat in the free chart.p.6
• Power — How much of the fleet is runningHow much of the American reactor fleet was running.p.7
• Power — How hard California's grid is workingWhat the California grid was carrying.p.7
• Power — How much power is waiting to connectHow much power is still waiting to be connected.p.8
• Accelerators — How fast a system serves Llama 2 70BHow fast the fastest measured system serves a 70B model.p.8

The only words in the book a machine did not write.

• What the rental price did, and what it means

  Still to be written.

• Power is the constraint, not chips

  Still to be written.

• What I would watch next month

  Still to be written.

• GPU rental — What a chip costs by the hourNothing landed on this line this month.
• Models — Open-weight models released each day250 open-weight models. New in this issue — nothing to compare it with yet.
• Models — Which model people prefer in a blind vote1,508 Elo. New in this issue — nothing to compare it with yet.
• Models — Where an AI app sits in the free chartnumber 2. New in this issue — nothing to compare it with yet.
• Power — How much of the fleet is running91.8%. New in this issue — nothing to compare it with yet.
• Power — How hard California's grid is working33,348 MW. New in this issue — nothing to compare it with yet.
• Power — How much power is waiting to connect2,061 GW. New in this issue — nothing to compare it with yet.
• Accelerators — How fast a system serves Llama 2 70B101,527 tokens a second. New in this issue — nothing to compare it with yet.