What is the actual number: how long does a cut-off buy?

Twenty-five to forty-nine months. That is the range across the scenarios I modelled for a country cut off from new chip supply, before its computing fleet falls below a competitive share of the frontier.

Can any major power actually sustain its AI compute alone?

Not with confidence, and the honest answer differs sharply by country. I define a "sovereign floor," the share of its own installed compute base a country could keep running from domestic and non-severed supply if cut off entirely. At the median: 31.1% for the United States, 53.4% for China, and just 3.1% for the…

Why does a "hidden switch" matter more than most people assume?

Because it decides whether cutting off supply makes the decline faster or does almost nothing extra. When new chips stop arriving, the only source of spare parts left is other broken machines, cannibalisation. In aviation under sanctions, operators have kept roughly ten aircraft flying by permanently grounding one…

What does this actually mean for AI policy, in one sentence?

Controls do not need to destroy an adversary's computing power. They only need to freeze it in place while the frontier keeps moving.

The Half-Life of Compute: research on chip export controls

A working paper puts a number on what chip export controls actually buy: 25 to 49 months, and stockpiling barely moves it. Free on Zenodo, DOI included.

Every argument for restricting the export of advanced AI chips rests on a claim about time: controls buy time. I went looking for the number behind that sentence and could not find one published anywhere. So I built the model myself, ran it, and the paper that came out says something more interesting than "it works."

I am publishing it as a working paper, DOI attached, code public, three languages. This article is the readable version. If you want the full derivation, it is one click away.

Start with the mistake everyone makes, including, until recently, me: treating a national stockpile of AI chips like a gold reserve. Count what a country has, rank it against everyone else, done. At least eleven public indices do exactly this today.