US and China are failing the same AI race, only differently

In late September, construction crews in Ulanqab, on the edge of the Gobi Desert, filed through face-scanning gates to raise another block of AI data centers above fields of sunflowers. Six weeks later, American voters did the opposite. Restrictions on data centers passed in 44 of 45 local referendums where counting was complete.

The contrast writes its own headline: China builds, America argues … Beijing wins. That storyline is tidy, and it is wrong.

Every data center needs three things, namely advanced chips, dependable electricity and permission to operate. Then it needs paying customers. No country has all three in abundance. China is short on chips. America is short on permission. And the Asian countries hosting both are discovering they lack sufficient power and water.

Start with China, which deserves credit for treating power and computing as one project. Its “Eastern data, Western compute” strategy steers investment toward regions with land and energy to spare.

Nearly 90 data centers are planned for Ulanqab alone, some designed to draw most of their electricity straight from wind farms. In the first half of this year, China created enough surplus renewable power to run Mexico for a year. Putting the data center where the electricity already is counts as a real advantage.

Last year, China’s own state planner began surveying a glut of locally backed computing centers, many running at 20% to 30% of capacity, after local governments built them in anticipation of state purchases rather than actual demand.

Beijing’s answer was a national cloud service to sell the idle capacity. But chips remain the harder constraint to China’s AI ambitions. Cheap electricity can offset the cost of running less efficient hardware, but it cannot conjure the hardware.

America’s shortage is the mirror image. Its companies have the chips and the money; by one estimate, four American firms control more AI computing power than all of China combined. What they often lack is a socket.

Last October, the US Energy Secretary directed federal regulators to overhaul how large loads connect to the grid, because interconnection queues and inconsistent rules were holding projects back. But America – and its AI-fueled stock markets – are finding that capital committed is not the same as computing delivered.

Then there are the neighbors. When American heartland residents ask about data centers’ impact on electricity bills, water or noise, they are also asking who pays for an investment whose benefits seemingly accrue mostly in Silicon Valley.

It does not follow that every data center objection is right or every delay useful. Blanket moratoriums will stop well-designed projects along with speculative ones.

The sensible goal is a clear route to approval for projects that can meet credible resource and community tests. Treating all opposition as disloyalty in a contest with China makes that route harder to build.

Malaysia shows why this matters beyond the two AI superpower rivals. In Johor, just across the causeway from Singapore, which froze new data center builds for three years over energy constraints, dozens of projects serving giants from Microsoft to ByteDance are rising almost side by side.

Some of those projects give Chinese firms access to advanced computing they could not legally import, because the chips never enter China. Where the chips sit, who owns them and who rents their output are three different questions, and US export controls drawn around geography struggle with the second two.

The local costs are easier to locate. Johor residents have protested dust so thick they can no longer dry laundry on their balconies, and the data center boom is straining power and water supplies.

Malaysia has started setting new terms. In July, Deputy Minister of Investment, Trade and Industry Sim Tze Tzin told parliament that new data centers would be approved only when power and water capacity exceed residents’ needs. “The water supply must be prioritized for residents,” he said.

That is the right test, and a simple one: does the investment serve its host? For Asian governments courting AI data center investors from both the US and China, this is a negotiating opportunity. Accommodating superpower rivals widens a country’s options.

It does not settle, however, how much local capacity to commit to them, or what residents get back. Those terms should have been fixed before the first data center foundation was poured.

Three conditions would do most of the work. Developers should pay the grid and water costs their projects create and post guarantees against abandonment, so households do not inherit the bill when demand forecasts prove overly rosy.

Community agreements should carry enforceable, publicly reported commitments on water, noise and local benefits. And projects should be built in stages, expanding as paying AI customers and power actually arrive. In return, governments owe data center developers a predictable yes or no.

China and America are not running opposite AI races. They are failing at the same one in different ways: one on the supply of silicon and real demand, the other on the supply of sockets and consent. The next data center deal, wherever it is signed, should open with three questions: who pays, who benefits and what happens when the promises fail.

Y. Tony Yang is an endowed professor at George Washington University in Washington, DC.