NextFin News - China is closing the artificial-intelligence race with the United States not by outspending it on chips, but by outbuilding it on concrete and copper: a wave of data centers rising across its western and northern hinterland, where electricity is cheap, land is plentiful, and the state decides where the wires run.
The scale of the buildout is the story. In Ulanqab, a city of about 1.5 million people on the Inner Mongolian steppe, nearly 100 data centers have opened or begun construction since 2016, with Chinese companies pledging a combined 12.5 gigawatts of capacity — and more than 70% of those commitments announced in just the past year, according to a recent research note. For comparison, OpenAI's $500 billion Stargate Project is set to reach only 10 gigawatts when complete.
This is not a spontaneous land rush. It is the latest phase of "Eastern Data, Western Computing," a nationwide program launched in 2021 that designated eight national computing hubs and 10 data-center clusters — five of them in western China — to shift compute load away from the power-constrained, land-scarce coast. The western hubs, from the karst mountains of Guizhou to the wind-swept plateaus of Inner Mongolia and the solar basins of Gansu, offer electricity costs roughly 40% below the national average, near-zero land costs versus eastern tier-one cities, and climates that allow months of free natural cooling.
The numbers behind the ambition are stark, and they just got a fresh policy anchor. On September 7, 2026, China's Ministry of Industry and Information Technology published its 15th Five-Year Plan for the information and communications sector, setting a national target of 9,800 exaFLOPS (EFLOPS) of intelligent computing capacity by 2030 — more than four times the roughly 2,185 EFLOPS the country had reached by the end of June, which was itself up 177% from a year earlier. The plan calls for 3.8 trillion yuan ($532 billion) in cumulative information-infrastructure investment over the 2026–2030 period.
China's intelligent computing power reached 725.3 EFLOPS in 2024, up 74.1% year on year, and was projected to hit 1,460.3 EFLOPS in 2026 under an earlier industry forecast. Using a different definition, the National Data Administration put China's intelligent computing capacity at 1.59 million PFLOPS — 1,590 EFLOPS — by the end of 2025, the world's second-largest pool after the United States.
And the state is putting money behind the map. Beijing is preparing to spend around 2 trillion yuan ($295 billion) over five years on a network of interconnected computing hubs, with state firms China Mobile and China Telecom operating the bulk of the centers, and a requirement that at least 80% of the technology — including AI chips — come from domestic suppliers such as Huawei, effectively squeezing out Nvidia and AMD.
Yet the paradox at the heart of this race is that China is doing all of this while locked out of the world's best AI chips. The question this piece answers: can a country that cannot buy frontier semiconductors still win the AI infrastructure war by throwing cheap power, cheap land, and central planning at the problem?
The Mechanism: Why the Hinterland Is China's Unfair Advantage
The transmission channel is simple economics. Training a frontier AI model does not need to happen next to the end user; it needs electricity, cooling, and time. A training run can take months and requires little real-time interaction, which is why latency — the original drawback of remote data centers — no longer decides location.
With the rise of AI in 2022, there was the realization that actually, those remote data centers could be well-utilized for model training.
That is Andrew Stokols, a professor at Singapore Management University who studies China's compute infrastructure. Inner Mongolia sits at high elevation with long, cold winters, so servers need far less energy for cooling. It is close enough to Beijing that two dedicated fiber-optic cables, built in 2017 and 2019, have cut average latency to under five milliseconds — fast enough even for inference workloads. And electricity there is cheaper than almost anywhere else in China, driven by a surge in wind and solar plus abundant coal.
The result is one of the fastest-growing compute clusters in Asia forming in a place that, until recently, was known for sheep and coal. Huawei opened its first Ulanqab data center in 2016; Apple followed in 2019. Now AI-native companies are building their own infrastructure rather than renting from cloud providers: DeepSeek, ByteDance, Alibaba, and Xiaohongshu are all reportedly building large centers in Ulanqab.
But the advantage is not only geographic. China's cloud utilization rate sits at roughly 28%, against about 60% in the United States, according to data from China's state-backed Academy of Information and Communications Technologies — a gap created by years of state-owned enterprises and government agencies defaulting to on-premises deployments. The 15th Five-Year Plan's push for a National Unified Computing Power Network is designed to pool that scattered capacity and route workloads to underutilized centers, squeezing more output from every chip the country already has. When you cannot buy the best chips, you make the chips you have work harder.
The Constraint That Became the Strategy
Here is the counter-intuitive part. The US export controls that were meant to slow China's AI progress have, in effect, forced Beijing to optimize for a different variable. Washington assumed the race would be decided by chip quality — the number of transistors, the memory bandwidth, the performance per watt. China is competing on total compute deployed, powered by the cheapest electrons it can find.
The domestic substitution is already underway at scale. Roughly 812,000 Huawei Ascend chips shipped in 2025, with AI processor revenue of about $7.5 billion, projected near $12 billion in 2026. The Ascend 950PR entered production in early 2026, and Huawei plans around 750,000 units this year with mass production beginning in April. Chinese firms have little choice but to adapt: in early 2026, parts suppliers paused production of Nvidia's China-targeted H200 after Chinese customs blocked shipments of the newly approved processors, and the models that define China's AI moment are now written for domestic silicon. GLM-5, a 744-billion-parameter model, was trained entirely on Huawei Ascend chips; DeepSeek V4, at 1 trillion parameters, is optimized for Ascend.
The gap on raw chip performance remains real. A Council on Foreign Relations model published in December 2025 put Huawei's 2026 aggregate AI compute at around 4% of Nvidia's, even on aggressive production assumptions. But that arithmetic may understate the strategic point: 4% of Nvidia's performance, deployed on power that costs 40% less, in facilities the state is paying to build, can still amount to a formidable national compute stock. Efficiency compounds the advantage. If a domestic cluster is 40% slower per chip but runs on electricity that costs 40% less and sits in a facility subsidized by the state, the cost per training run narrows faster than the benchmark gap suggests.
The American Problem: The Grid, Not the Chip
While China builds outward into the hinterland, the United States is running into a wall at home. US data center power demand is expected to climb from 31 gigawatts in 2025 to 41 GW in 2026 and 66 GW in 2027, according to the forecast from a major Wall Street research team — more than doubling capacity by the end of 2027, with year-over-year additions accelerating to 36.3 GW in 2027 from 8.5 GW in 2025. Data centers' share of total US peak summer power demand is projected to rise from 4.1% in 2025 to 8.5% in 2027.
Anthropic has estimated that by 2027, training a single frontier model will require five gigawatts of power, and that the US AI sector alone will need 50 GW of new electric capacity by 2028 to maintain global leadership — roughly twice the peak electricity demand of New York City.
The second-order implication is uncomfortable for the conventional read of this race. The United States leads on chip design, on model quality, and on private capital. But its buildout is fragmented across private developers, local permitting authorities, and a grid that was not built for this load. China's advantage is not that it spends more — by most measures it spends less — but that it can site, wire, and subsidize faster because the state decides where the electrons go. In a race where the binding constraint is no longer the chip but the substation, central planning is not a bug; it is the product.
The Counter-Thesis: Coal, Water, and Empty Halls
The strongest case against the hinterland thesis attacks it at its foundation: China's buildout may be fast, but it may also be wasteful, dirty, and ultimately underutilized.
First, the power is not as green as the brochures claim. About 37% of Ulanqab's electricity still comes from coal, and because data centers must run around the clock, operators have historically favored fossil fuels for reliability.
Inner Mongolia has long been the West Virginia of China. It's a coal country.
That is Damien Ma, director of Carnegie China. He adds that the region is racing to replace coal with wind and solar, but "in three years, maybe it will be completely powered by renewables." That "maybe" is doing a lot of work.
Second, water. Ulanqab receives only about 14 inches of rain a year — about as dry as Denver. The local water company was forced last month to shut several waterworks for seven hours each night to manage peak demand, before many of the planned centers are even running. AI racks generate heat that must be removed, and water-cooling is the cheapest way to do it.
Third, utilization. In recent years, local governments raced to build data centers to hit regional digitalization targets and national KPIs, often without regard for grid capacity, renewable availability, or actual demand. The 15th Five-Year Plan's directive to "coordinate the layout and orderly construction of computing infrastructure" is itself an admission that too much was built in the wrong places. Empty server halls do not train models.
These are real risks. But they are risks of execution, not of direction. The state's response to overbuilding is not to stop building — it is to centralize the siting decisions. The response to coal dependence is to build dedicated renewable stations next to the parks; Envision, one of China's largest wind-turbine makers, announced this month it will build a 2 GW AI data center in Ulanqab connected directly to its own clean power supply. And the response to chip scarcity is to force the domestic ecosystem to improve faster than it would have under open procurement.
Conclusion: The Race Moves to the Steppe
The mechanism, cashed out: China cannot win the AI race on chip quality in the near term, so it is changing the unit of competition from performance-per-chip to total compute deployed per dollar of electricity. The beneficiaries are the western provinces with cheap power and cool air — Inner Mongolia, Guizhou, Gansu, Ningxia — and the domestic chip and equipment suppliers that the 80% local-content rule protects. The exposed are the foreign chip vendors locked out of the world's second-largest AI market, and the US grid, which now faces a buildout race it did not plan for.
Split by time horizon, the picture is mixed. In the short term, sentiment will swing on individual model releases and chip benchmarks, where the United States still holds a wide lead. Over the medium term, the binding constraint shifts to capacity: who can bring the most compute online, fastest, at the lowest cost — and there China's state-coordinated siting is a genuine edge. Over the long term, the structural question is whether domestic silicon can close the quality gap before the next generation of models makes today's clusters obsolete.
Three scenarios frame the path ahead:
- Base case: China's intelligent computing capacity grows at 30–40% annually through 2027, closing the gap in total deployed compute while remaining behind on frontier training quality. The US hits most of its 2026 capacity additions but slips on 2027 as permitting and interconnection queues bite.
- Upside for China: domestic chips deliver a step-change in software efficiency, and the western hubs absorb surplus renewable power faster than expected, letting Beijing approach its 9,800 EFLOPS 2030 target ahead of schedule.
- Downside for China: water constraints and grid bottlenecks strand capacity, coal dependence triggers carbon or trade friction, and the 28% cloud utilization rate stays stuck — meaning the centers get built but do not get filled.
The falsifying signal is specific: if China's in-service AI computing capacity growth slows below roughly 30% year on year in 2026–2027 — well under the nearly 40% pace recorded in 2025 and the trajectory implied by the 15th Five-Year Plan — the hinterland buildout is not closing the gap, it is just adding empty halls. Conversely, if US data-center capacity additions miss the 13.6 GW scheduled for 2026, the American power constraint is tighter than the models assume.
The race for AI supremacy was supposed to be won in the clean room, one transistor at a time. It may instead be decided on the steppe, one gigawatt at a time — and China has just made the steppe its home field.
Explore more exclusive insights at nextfin.ai.
