What is compute governance?

Compute governance is a subfield of AI governance that focuses on controlling access to the computing hardware needed to develop and deploy AI. Some have argued that compute governance is a particularly promising method of AI governance.

Lennart Heim has proposed three main strategies for compute governance:

  • Monitoring compute usage: Compute can be monitored for the purposes of verification, transparency, and identifying which systems are high-risk. If compute is well-monitored, it can be used as an index for policy-making, such as by imposing regulations based on how much compute a model uses.

  • Restricting access and usage: For example, limits on compute can be used as “emergency brakes”, to avoid diffusion, and to regulate specific training runs.

  • Promoting beneficial usage: For example, more safety-conscious actors could be given more access to compute, and safety research could be promoted through sponsoring compute.

As of October 2023, there are few policies in place for governing compute aside from US export controls for advanced microchips to China, and much of the research done in compute governance is exploratory. Examples of current work in compute governance include:

Further reading: