A data center does not sleep. It does not observe weather, or seasons, or the particular slant of light at four in the afternoon in November. It draws power at the same punishing rate at midnight as it does at noon. The newest ones, the ones built to train and run artificial intelligence, draw more of it than anything humans have built before.
Somewhere in that mismatch, between a machine that never stops and a sun that only shows up for half the day, sits the most consequential energy question of the decade.
Two numbers
In May 2025, China installed approximately 93 gigawatts of solar capacity. In a single month. That is more than the United States installed across the entirety of 2025, which came to about 43 gigawatts. By year’s end, China’s cumulative solar capacity had crossed 1.1 terawatts, the first country ever to cross that line, while America finished the year near 279 gigawatts. Do the division and China now holds roughly 4.7 times as much installed solar as the United States.
If capacity were the whole story, the piece would already be written, and the headline would read: China wins. But capacity is not the whole story. It is closer to a rumor of the story.
Here is the number that complicates the first one. Despite that fleet, roughly seventy percent of the electricity actually running Chinese AI data centers still comes from coal. China has installed more solar in the last twelve months than most countries have ever built, and its own machines for building intelligence still run mostly on the fuel they always have.
The real question, the one that should organize any serious comparison, is whether either country can convert an intermittent, weather-dependent, daylight-only resource into something a data center can actually run on, hour after hour, regardless of cloud cover or time zone. Solar panels generate electricity. AI infrastructure needs reliability. Those are not the same commodity, and the gap between them is where this story lives.
The paradox at the center of China’s buildout
The coal dependence is not simply inertia. It is geography, and it is grid physics. China’s best solar sits in the northwest: Xinjiang, Gansu, Qinghai, Inner Mongolia. Its computing demand has historically clustered in the east, around Beijing, Shanghai, and the Pearl River Delta. Even after electrons travel a thousand miles on China’s newest transmission lines, the western hub zones can only absorb so much of what their own solar farms produce. In parts of Gansu and Inner Mongolia, curtailment, solar generated but never used, runs as high as twelve to twenty percent. The country is building panels faster than its own grid can drink from them.
China has answered this with genuine engineering ambition. By the end of 2025 it had 46 ultra-high-voltage transmission projects in operation, more than 62,000 kilometers of line, moving upward of 340 gigawatts of capacity from west to east. The Ningxia-Hunan line alone carries 8 gigawatts across five provinces with less than three percent transmission loss per thousand kilometers. State Grid Corporation spent roughly seventy billion dollars on grid construction in 2024 alone.
None of it has been fast enough to outrun the buildout. “Even China can’t develop that fast,” the analyst David Fishman has observed, and the data backs him up: solar curtailment actually rose in the first half of 2025, not fell, as new capacity kept outpacing the grid’s ability to absorb it.
So the solar sits there, real and enormous, while the servers run largely on the same fuel they always have. A nation can lead the world in a technology and still not be using it for the thing everyone assumes it is used for.
There is one genuine exception worth naming, because it is more instructive than the headline numbers. Tencent’s Huailai Dongyuan facility in Hebei, a modest microgrid combining wind, solar, and battery storage, runs at roughly seventy-one percent renewable power at peak. It has received a fraction of the attention paid to America’s five-hundred-billion-dollar Stargate announcements. It is, by the evidence, closer to actually solving the problem than anything currently running in Texas.
America’s own sleight of hand
The American story has a version of this too, and it may be the more interesting one, because it is not primarily a story about a technology gap. It is a story about what happens when a country decides the grid-connection problem is too slow to solve, and reaches for the fastest available substitute instead.
The American grid is, at the moment, in something close to gridlock. By the end of 2024, roughly 2,300 gigawatts of generation and storage sat in interconnection queues nationally, nearly double the capacity of every power plant currently operating in the country. Historically only about fourteen percent of queued projects ever actually get built. In the PJM system, which serves thirteen states and Washington, the average wait for a connection now runs past three years, against a target of under a year.
Even when a project clears the queue, the physical grid may not be ready for it. The country completed only about 322 miles of new high-voltage transmission line in 2024, a fraction of what it built routinely in the early 2010s. Large power transformers, the least glamorous piece of the entire system, now carry lead times of two and a half to nearly three years, with prices up seventy-seven percent since 2019 and roughly eighty percent of supply still imported. The bottleneck holding back American AI power turns out to run through a piece of equipment nobody outside the utility industry could have named a decade ago.
Facing that wall, hyperscale developers have made a choice. In April 2026 alone, American companies committed more than seven gigawatts of behind-the-meter natural gas generation, turbines and reciprocating engines built directly beside the data center, bypassing the public grid entirely. xAI’s Colossus campus in Memphis runs largely on gas turbines while a hundred-megawatt solar project, backed by a $439 million federal loan, gets built alongside it, supplying perhaps ten percent of the site’s actual load once finished. Meta has done something similar in Ohio. Several of the announced Stargate sites in Texas and New Mexico describe themselves, in their own official materials, as powered by onsite gas microgrids, with solar listed as a future addition rather than a current source.
A gas turbine sitting next to a data center is a rational, even elegant, engineering solution to the twenty-four-hour problem. It is also, quite plainly, not a solar solution.
The policy ground has shifted under this decision, too. The One Big Beautiful Bill Act, signed in July 2025, sharply accelerated the phase-out of federal solar and wind tax credits while preserving incentives for nuclear, geothermal, and battery storage through 2033. A companion executive order explicitly framed Chinese-linked solar supply chains as a national security concern. In the months since, gas applications in the Texas grid jumped roughly one hundred fifty percent while new renewable applications slowed and existing ones were withdrawn. The country did not run out of sunlight. It changed the economics of chasing it.
The mirage both countries share
Here is the through-line connecting both countries, and it may be the sharpest finding in the whole comparison. Nearly every “solar-powered AI data center” announcement, in Shenzhen and in Texas alike, rests on a distinction almost no one outside the energy industry understands. There is annual renewable energy matching. And there is hourly carbon-free operation. They are not the same thing, though the press releases rarely say so.
The first means a company bought enough certificates, renewable energy certificates in America, green electricity certificates in China, over the course of a year to offset, on paper, the electricity its data center consumed. The instruments are functionally identical across both countries. The second means the electrons flowing into the servers, at the actual moment they are needed, came from a carbon-free source.
Google’s celebrated Orion Solar Belt in Texas, its largest solar investment anywhere in the world, is a useful example precisely because it is a good-faith one. The nine-hundred-megawatt project, built by SB Energy and commissioned in October 2024, does not physically wire into Google’s Dallas-area data centers. It feeds the same public ERCOT grid the data centers already draw from, at whatever hour the sun happens to be out. The company purchased the output; it did not build a private wire.
A facility can be one hundred percent solar-matched on an annual balance sheet while running on coal, or gas, at eleven o’clock on a Tuesday night.
Most of what gets called solar-powered AI, on both sides of the Pacific, is the former. The claim is not false, exactly. It is answering a different question than the one the public thinks it is asking.
The twist neither country’s press office mentions
There is a final complication, and it cuts against the urgency of the whole race. By some credible estimates, roughly eighty percent of the AI computing capacity China announced and built in 2023 and 2024 was sitting idle by the middle of 2025, partly because more efficient models, DeepSeek’s among them, needed fewer chips than planners had assumed. China may have overbuilt its data centers at almost exactly the moment it was overbuilding its solar fleet, two enormous bets on a demand curve that has not yet arrived as steeply as forecast.
That does not make the underlying race less real. It does mean some of the panic embedded in both countries’ infrastructure announcements is running ahead of the actual computing load those announcements claim to serve.
Who is actually ahead
None of this means the question of who leads is unanswerable. It means the honest answer is not the one either country’s press office would prefer. China holds the more genuine long-term advantage in solar manufacturing, controlling somewhere above ninety percent of global polysilicon, wafer, and cell production, in the sheer speed of deployment, and in a centralized transmission-planning apparatus that the fragmented American grid cannot currently match. Its module costs, near nine cents a watt, run forty to sixty percent below what American developers pay even after reshoring efforts.
But that advantage is throttled by exactly the infrastructure China has not finished building: the absorption capacity that would let western sunlight actually reach eastern servers on demand, rather than sitting curtailed at the source.
The United States holds the advantage in capital and in the willingness of hyperscale companies to simply build their own power plant next to their own data center rather than wait for a public grid to catch up. Amazon alone has 13.6 gigawatts of solar under development domestically. Google’s $4.75 billion acquisition of the developer Intersect Power is the most serious American attempt yet at physical, rather than contractual, colocation of clean power and compute.
But that advantage increasingly means quietly walking away from solar altogether, in favor of the fuel that can run at three in the morning without asking anyone’s permission.
If solar-powered AI becomes as strategically important over the next decade as both countries currently claim to believe, the country better positioned to make good on that claim is China, on the strength of its transmission ambitions and its manufacturing base. The bottleneck that could overturn that advantage is not a shortage of panels. It is the unresolved fight, publicly unacknowledged on either side, between Beijing’s renewable mandates and its own grid operators, who have quietly resisted designing AI facilities around a power source they consider too variable to trust. The American equivalent of that fight is playing out in a less centralized but no less consequential form, inside interconnection queues and transformer factories that were never built for a customer that wants ten times the power a decade ahead of schedule.
Whoever solves the storage and transmission problem first, not whoever installs the most solar, will be the country that actually gets to say it runs its intelligence on sunlight.