The US does most AI inference spending — but does that mean its labs lead, or just that its customers pay the bills?

This explores whether US 'inference dominance' (the large share of AI model usage traced to the US) is about where the labs sit or where AI gets put to work. The collection has only one direct data point, and it measures something narrower than the question suggests. Chainalysis traced payments made over x402, a protocol that lets software agents pay each other on-chain, back to the countries of the paying wallets. It found US wallets made 88.1% of payments for AI model inference, while Chinese wallets clustered around security services Which countries dominate agent-to-agent AI payments on chain?. Those are buyer locations, not seller locations. So the number describes demand: who is paying for model calls. It says nothing about where the models were trained or which servers ran them. In this dataset, at least, US dominance reflects deployment.

That split between buying and building matters more than it first appears. Inference is where AI money goes. Epoch AI's look at Anthropic, Minimax and Z.ai found that R&D and inference compute together make up 54–62% of operating expenses, while staff costs stay under 25% Do compute costs really dominate AI company budgets?. The physical bill is mostly hardware. In a 1-gigawatt data center, servers cost about $5B a year and energy only about $0.6B What actually drives AI data center costs?. Server spending lands wherever the data centers are built, and that can differ from both where a lab has its offices and where its customers are. 'Where inference happens' can therefore mean three different places: the lab's home country, the data center's location, and the paying user's location. The Chainalysis data captures only the third.

The gap between these measures could grow. One model of the AI compute stack argues that inference will far outstrip training because every autonomous agent keeps using inference for as long as it runs, a source of demand training never had Will inference compute soon exceed training compute demand?. If that holds, inference geography will increasingly follow where agents are deployed and who pays for them. The x402 payments, which are agents paying agents, are an early example of that pattern.

The collection has a clear gap here. It has no notes that break down inference by data-center location, by where labs are headquartered, or by user geography outside crypto. On-chain agent payments are a small and unusual slice of overall AI usage, so the 88.1% figure shouldn't be read as the US share of global inference. A full answer would need data the library doesn't yet have, such as cloud-region traffic or API usage broken down by country.