Definition
Compute governance uses AI’s underlying hardware (chips, data centers, cloud capacity) as a policy lever, because that hardware is physical, measurable, and made by only a few companies.
At a glance
- Cutting-edge AI needs enormous specialized computing power that is far easier to track than software, data, or models.
- Compute is governable because it is detectable, excludable, quantifiable, and concentrated in a few suppliers like Nvidia and TSMC.
- Tools already in use: US export controls on advanced chips to China, plus reporting above a compute threshold (10^26 operations in the US, 10^25 in the EU).
- The rules are volatile and shifting fast.
Why hardware is the lever
You cannot easily regulate an idea or a model file, both copied instantly. But frontier AI runs on physical machinery: thousands of chips in power-hungry data centers, made by only a few firms like Nvidia and TSMC[5]. That makes it hard to hide, easy to count, and easy to gate[1].
What it lets governments do
Three things[2]: visibility (require labs and cloud providers to report large training runs), allocation (steer compute toward beneficial research or slow the pace), and enforcement (block sales or limit how chips connect). The main mechanisms today are export controls and FLOP reporting thresholds[4].
Why a business owner should care
If you buy cloud compute, deploy AI tools, or touch advanced chips, these rules shape your costs, suppliers, and markets. In May 2025 the US rescinded the Biden-era AI Diffusion Rule, shifting toward chip access as a negotiating tool[3]. Expect ongoing change.
Bottom line
Frontier AI can’t be built without scarce, visible hardware from a few suppliers, giving governments a rare handle on it, but treat the rules as a moving target.