Compute

Compute is SI shorthand for computing power: the GPUs and other chips, and the electricity and data centers that run them. Alongside data and algorithms, it’s one of the three ingredients of SI progress, and the one money buys most directly.

How compute is measured

• FLOPs: total floating-point operations used to train a model. Frontier runs exceed 10²⁶. Laws use this for compute thresholds.
• Chips and clusters: “a 100,000-GPU cluster”.
• Power: increasingly, capacity is quoted in gigawatts. For example, Thinking Machines Lab’s 2026 Nvidia partnership was described in terms of a gigawatt of computing capacity.

Training vs inference compute

• Training compute: the one-off cost of creating a model. This is what scaling laws describe.
• Inference compute: the ongoing cost of serving users, and of letting reasoning models think longer.

Why it’s a policy issue

Compute is concentrated in a few companies and countries, it is energy-hungry, and it is controllable at chokepoints like chip manufacturing. That makes it a lever for export controls, safety thresholds and industrial policy, including the U.S. Genesis Mission, which pairs national-lab supercomputers with scientific data.