All the books

Book · 6 contracts · 12 expiries

The price of compute

What an hour of a GPU and a million tokens cost, and how much of the market open models take: the prices that decide who earns the margin.

The theses it settles

The contracts

The next expiry of each; a strike opens its question, the title its whole ladder.

Who sits on each side

The commercial parties a print hurts, on each side: the ones who would hedge with a contract on it.

  • GPU rental price index (Ornn H100, H200, B200)
    Loses if it rises
    • AI labs and startups renting GPU capacity by the hour
    • inference platforms selling tokens on fixed prices over rented GPUs
    Loses if it falls
    • neoclouds leasing H100, H200 and B200 fleets (CoreWeave, Lambda, Nebius, Crusoe)
    • lenders holding loans secured by GPUs and their rental income
  • Flagship AI model token price
    Loses if it rises
    • application companies and enterprises buying tokens by the million
    • coding-agent and chat apps reselling model output at flat prices
    Loses if it falls
    • AI labs selling API access (OpenAI, Anthropic, Google, xAI)
    • inference clouds reselling open models on per-token prices
  • Chinese and open-weight share of model tokens
    Loses if it rises
    • closed US labs selling API access
    • investors in closed-model valuations
    Loses if it falls
    • Chinese labs and open-weight makers (Alibaba's Qwen, Moonshot, Z.ai and others)
    • inference clouds built on open models