Compute Governance: Regulating AI Through Chips and Data Centers
This page is a structured working draft — real analysis, not yet expanded with the full expert sourcing given to the flagship pages. Safe to build on; treat specifics as provisional until sourced.
Why compute is the governance lever of choice
Algorithms and trained model weights are easy to copy and hard to track. The physical hardware needed to train a frontier model — advanced AI chips and the data centers housing them — is comparatively rare, expensive, and traceable. Because capability has scaled predictably with compute (see Compute and Scaling Laws), monitoring and restricting compute functions as an imperfect but real proxy for monitoring capability itself.
Current mechanisms
- Export controls restricting the sale of the most advanced AI training chips to specific countries, intended to slow the rate at which a competing state can compound frontier capability.
- Reporting thresholds, such as those in the U.S. executive actions and the EU AI Act, requiring labs to disclose training runs above a certain compute threshold to regulators before or during training.
- Know-your-customer requirements on cloud compute providers, aimed at preventing restricted actors from accessing frontier-scale compute indirectly through a cloud intermediary.
The known weaknesses
Compute governance is a blunt, evadable instrument: smuggling, third-country transshipment, and algorithmic efficiency gains (doing more with the same or less compute) all erode its effectiveness over time. It buys time rather than solving the underlying problem, and most policy researchers in this space describe it that way rather than as a permanent fix.
Why it’s still worth doing
An imperfect brake is still a brake. Compute governance is one of the only currently available policy tools that acts on physical reality rather than promises, which is precisely why it has become a centerpiece of the international governance conversation covered in International Treaties and Summits.