America Is Buying Up AI Compute. The Bills Will Settle the Argument

The US is buying up AI compute, and its appetite is growing. According to Epoch AI’s estimates, it received 72% of the world’s new AI computing capacity between 2022 and June 2026. China accounted for roughly 10%, the European Union about 5%. The US share of annual arrivals rose from around 60% in 2022–2024 to 72% in 2025, then 78% in the first half of 2026. American industry is buying as though tomorrow’s demand will make today’s spending look cheap.

Epoch tracks where the hardware goes using customs records, adjusting for re-exports and combining the results with estimates of accelerator sales. Almost every advanced AI chip is packaged in Taiwan, so the journey to a data center usually leaves a trail in trade statistics. Capacity is measured in Nvidia H100 equivalents. These figures track hardware arrivals; installation, power connections and paying workloads still have to follow. The calculation separately includes estimates of Chinese production by Huawei and Cambricon, while excluding chips both made and used within Taiwan.

Counting chips alone makes the gap look smaller. China has about 18% of the chips covered by the estimates but just 10% of the computing capacity. Much of its hardware consists of less powerful processors, including Nvidia’s export-compliant H20 and domestic accelerators. Export restrictions give Chinese labs a practical reason to squeeze more work out of the compute they can get. DeepSeek, Qwen, GLM and others are still producing competitive models. A vast accelerator fleet is an advantage, but rivals can compete through better engineering and lower prices.

There is a clear business case for grabbing as much of the market as possible, whatever the upfront cost. Buy the chips, secure sites and electricity, recruit developers and sign customers before demand explodes. If AI quickly becomes essential to most business workflows, the cautious players will have to catch up on terms set by the early builders. They will rent capacity from companies that took the risk first, handing over a share of their earnings. Spending that looks reckless today could turn out to have bought a commanding position in an enormous market.

Since 2022, 72% of AI compute has ended up in the US and about 10% in China, including chips China makes itself.

The more restrained approach has its own appeal: buy compute for workloads that already make sense, switch models as price and performance improve, and make money from customer relationships and business software. SAP, for instance, incorporates third-party models into its enterprise products, including Claude for Joule agents. Its cloud revenue reached €6.7 billion in the third quarter of 2026, up 25% at constant currencies. That figure covers the entire cloud business; AI helps sell an established platform. This business model can work perfectly well without owning the world’s largest collection of accelerators.

Waiting can be expensive, too. If capacity stays scarce, renters are exposed to someone else’s prices, waiting lists and allocation decisions. Even Europe’s Mistral is expanding its own GPU infrastructure, and Microsoft has agreed to use part of that capacity. Each company has to decide how much infrastructure to put on its balance sheet, how much to rent, and how much confidence in future demand is enough to justify building ahead of it.

For hardware owners, the cost of getting that judgment wrong is already tangible. In our breakdown of a $30 billion data center, chips and server equipment accounted for $17.1 billion—more than half the budget. A building can outlast several generations of accelerators. The computing hardware inside it has to earn its keep before a newer generation makes the same work cheaper. Electricity, maintenance and financing add to the purchase price. Every technological leap is good news for users and another reason for owners of last year’s equipment to revisit their return calculations.

The bills can also prove much more dependable than demand. As we examined in an earlier piece on AI infrastructure economics, Reuters reported from Anthropic’s confidential IPO prospectus that roughly 80% of its infrastructure commitments are noncancelable or payable regardless of usage. Sales can miss the forecast while payments carry on. If competition drives down inference prices, rising query volumes still have to leave enough cash after costs. AI adoption and infrastructure returns can move at very different speeds.

America’s 72% is an imposing result for the first stage of the race. The next stage will test whether customers bring in enough money to pay for the lead already purchased. Scarce compute and strong returns would vindicate aggressive expansion. Excess capacity and a price war would favor companies that kept cash available to buy more cheaply later. When the bills come due, we will find out whether the market grab was foresight—or an extraordinarily expensive way to build infrastructure for more patient competitors.