compute governance
Suppose you wanted to limit the spread of a dangerous drug, and noticed that one rare chemical was needed to make it, produced in only a handful of factories worldwide. You would have found a chokepoint: a single place where rules can actually bite. Compute governance is the idea of treating the specialized computer chips that train the most powerful AI as exactly that kind of chokepoint.
The reasoning rests on a few facts about 'compute' (raw computing power). Training a frontier model takes enormous amounts of it, concentrated in vast data centers full of cutting-edge AI chips. Those chips are physical, expensive, and made by very few companies using a supply chain (chip design, fabrication, and the lithography machines that print the circuits) that runs through a tiny number of firms. Unlike software, you cannot copy a chip over the internet. That makes hardware unusually governable: it can be counted, located, export-controlled, and in principle fitted with on-chip mechanisms that log or limit how it is used. For example, a government can restrict exports of the most advanced AI chips, or require large training runs to be reported once they cross a compute threshold.
Compute governance is attractive because it offers leverage when other levers are weak: you cannot easily inspect every line of a model's code, but you can ask who is buying tens of thousands of the world's most advanced chips. It is not a cure-all, though. Thresholds based on compute can become outdated as algorithms get more efficient (the same capability for less compute); export controls can spur rival supply chains; and on-chip monitoring raises real privacy, security, and concentration-of-power concerns. It is a promising but contested tool, not a settled solution.
A country bars the export of its most advanced AI accelerators to certain buyers, and separately requires any domestic training run above a set number of operations to be registered, because both the chips and the giant clusters are hard to hide.
Hardware is physical and concentrated, so it is one of the few parts of the AI pipeline you can actually count and control.
Compute thresholds are a proxy, not the thing itself. Because algorithms keep getting more efficient, a fixed operations count drifts over time, so a number that looks 'frontier' today may be ordinary in a few years.