In 2022, China launched a national infrastructure program with a name that sounded more like a bureaucratic slogan than a technological strategy: "East Data, West Computing." Four years later, the slogan has become a physical reality — and it is reshaping the geography of AI compute in ways that have no direct equivalent in the United States.
The premise is straightforward. China's eastern provinces — Beijing, Shanghai, Shenzhen, Hangzhou — generate the demand for AI compute. They host the users, the companies, and the capital. But they do not have the land, the power, or the climate to host the data centers that serve that demand. The western provinces — Inner Mongolia, Gansu, Ningxia, Guizhou — have all three.
China responded by moving the computation to where the resources are.
As of mid-2026, more than 80 percent of China's intelligent computing capacity is concentrated in eight national hub nodes, five of which are located in the west. The national intelligent compute scale reached 2,185 EFLOPS (FP16) by June 2026, a 177 percent year-over-year increase, according to the Ministry of Industry and Information Technology. The eight hubs account for over 85 percent of that total.
This is less a story about speed than about geography — and the reasons behind it.
The Numbers Behind the Migration
The scale of the western build-out is substantial.
Inner Mongolia has become the largest computing hub in the country. As of July 2026, the region's total computing power reached 345,000 PFlops, with intelligent computing accounting for 326,000 PFlops. More than 85 percent of the electricity used by regional data centers comes from green power. The region expects to exceed 400,000 PFlops by the end of 2026.
The city of Ulanqab alone has signed 89 data center projects with a combined agreed investment exceeding 500 billion yuan (about $70 billion). Its operational and planned capacity is approximately 12.5 gigawatts — more than ten times its operational scale in 2025. Huawei, Apple, Alibaba, ByteDance, Kuaishou, and DeepSeek have all established facilities there.
Guizhou hosts the China Telecom Cloud Computing Guizhou Information Park, a 7 billion yuan facility spanning 586 acres with a planned capacity of 800,000 servers. More than 90 percent of its computing power serves customers outside the province. Its best PUE has dropped below 1.2, and it has been recognized as a national green data center. Domestic computing capacity accounts for over 70 percent of the park's total.
Ningxia's Zhongwei cluster has reached 331,000 standard racks and 264,000 PFlops of computing capacity as of May 2026. The city has attracted 31 data center enterprises, including Amazon and China's three major telecom operators, with cumulative investment of 67.2 billion yuan. In the first half of 2026 alone, information transmission investment in Zhongwei grew 5.26 times year-over-year, rising from 8.8 percent to 46.7 percent of the city's total investment.
Gansu's Qingyang cluster reached 185,500 PFlops by mid-2026, with a year-end target of 300,000 PFlops. A total of 571 digital economy enterprises have registered in the city. The China Mobile Qingyang data center represents 4.24 billion yuan in investment and 4,484 physical racks, with the broader park carrying over 100,000 PFlops of computing capacity.
The national investment commitment is equally large. During the 15th Five-Year Plan period (2026-2030), China's computing power networks are set to receive 4 trillion yuan (about $589 billion) in new direct investment, according to the National Development and Reform Commission. The national integrated computing network monitoring and scheduling platform has already integrated over 60 percent of the nation's computing power.
Why the West
The reasoning is rooted in physics and economics.
Modern AI racks consume 100 to 140 kilowatts — ten times the power of a traditional server rack. A single smart computing center with 10,000 GPUs consumes roughly as much electricity per hour as 300 households use in a full day. Cooling that heat requires either massive amounts of water or a climate that does the cooling for you.
The western provinces offer both advantages.
Ulanqab sits at 42 degrees north latitude — the "golden latitude" for data centers. Its average annual temperature is 4.3°C, with roughly ten months per year suitable for natural cooling. Its fiber optic connection to Beijing has a one-way latency as low as 2 milliseconds.
Zhongwei, on the southeastern edge of the Tengger Desert, has an average annual temperature of 8.8°C, over 280 days of good air quality per year, and a geological structure so stable that the probability of a magnitude 7 earthquake in the coming centuries is near zero.
Qingyang has an average annual temperature below 8°C. Its green power direct supply price is 0.398 yuan per kilowatt-hour — roughly half the industrial electricity price in eastern China.
Hohhot, the capital of Inner Mongolia, has built a low-latency computing service circle: 2 milliseconds to Hohhot-Baotou-Ordos-Ulanqab, 5 milliseconds to Beijing-Tianjin-Hebei, and 20 milliseconds to the Yangtze River Delta. Its delivered electricity price for computing transactions is stable at 0.32 to 0.36 yuan per kilowatt-hour — among the lowest in the country.
The power advantage is not just about price. It is about availability. Western China has abundant wind and solar resources. Inner Mongolia's new energy installed capacity reached 20.288 million kilowatts, with green power accounting for 67 percent of the regional total. The national standard for new data centers in hub nodes requires a green power share of no less than 80 percent and a PUE below 1.2.
Chinese data centers in the west are not just cheaper to power. They are greener.
The Companies That Moved
The migration is not a government-directed relocation of anonymous capacity. Named companies are building named facilities.
DeepSeek plans to build approximately 1 gigawatt of AI data center capacity in Ulanqab, and has listed the city as a recruitment location for data center delivery managers and operations engineers.
ByteDance is in early negotiations to add 5 to 6 gigawatts of capacity in Ulanqab, targeting delivery by early 2028 — an investment of roughly 800 billion to 960 billion yuan at an estimated 160 billion yuan per gigawatt.
China Unicom Data has landed a 18.2 billion yuan project in Chayouqian Banner, expected to form a computing cluster exceeding 67,000 PFlops upon completion.
GDS has signed a five-year agreement with Ulanqab worth over 30 billion yuan.
In Guizhou, the China Telecom park's B1B2 cluster has built 14,000 P of intelligent computing capacity, becoming the country's first 10,000-card fully domestic, autonomous and controllable smart computing cluster.
In Zhongwei, 13 large models — including Tencent's Yuanbao and Baidu's Wenxiaoyan — are trained and inferenced on local infrastructure. Eight national ministries, including the Publicity Department and the Supreme People's Court, have located data operations there.
The compute is not just being built. It is being used.
The Electricity-First Logic
The most important shift in China's AI infrastructure strategy is not about compute. It is about electricity.
"East Data, West Computing" was initially framed as a computing power strategy — move the servers to where the power is cheaper. Four years in, the framing has inverted. The determining factor in China's AI competition is no longer how many chips you can stack, but whether you can put the compute where the power already is.
This is visible in the mechanics of the program. The national standard for hub node data centers requires a green power share of at least 80 percent. The "computing-electricity coordination" model has become a formal policy priority, with projects explicitly designed around the principle that the load follows the source, not the other way around.
The Ulanqab project pairs 300 megawatts of renewable generation (200 MW wind, 100 MW solar) with 45 megawatts of battery storage, operating on a "load determines source, source follows load" principle. The Ningxia Zhongwei project — phase one totaling 2 gigawatts with 8.7 billion yuan in investment — combines 500 megawatts of photovoltaic with 1.5 gigawatts of wind, using a dual-track system of "physical direct supply" and "bilateral trading."
The electricity comes first. The compute follows.
The American Contrast
The United States faces the same physics. It has not built the same solution.
In Texas, Governor Greg Abbott paused new grid connections in August 2026 after ERCOT's interconnection queue swelled to 474 gigawatts — more than five times the grid's record peak demand. Data centers accounted for approximately 90 percent of that total. BloombergNEF estimated the freeze could delay 49.8 gigawatts of data center load and cost projects up to $15 billion.
In PJM, the largest U.S. grid operator, the capacity auction for 2028-29 cleared at $325 per megawatt-day — the maximum allowed. Data centers contributed roughly $6.3 billion of the $16.4 billion total cost, which was socialized across every ratepayer on the grid.
In Virginia, Oregon, Tennessee, and Wisconsin, states have begun requiring data centers to pay their own way. In Virginia, the State Corporation Commission created a new GS-5 rate class for large loads. In Oregon, data center rates rose 29 percent while residential rates fell 1.3 percent.
The American approach is reactive. It responds to the queue, the ratepayer, the local zoning board. The Chinese approach is designed. It routes the computation to where the power, land, and climate already are.
Neither approach is inherently superior. China's system operates with a level of centralized coordination that would be politically difficult in the United States. It can designate a region as a computing hub and direct investment there. It can mandate green power thresholds. It can align grid planning with data center construction at the national level.
The U.S. system cannot do those things. It can build behind-the-meter gas plants — and it is. It can fast-track interconnection queues — and it is trying. It can pass ratepayer protection bills — and it did, by a vote of 417-3.
But it cannot tell a data center to move to the grasslands. And that is exactly what China did.
What the Grasslands Reveal
The "East Data, West Computing" program is often described as a Chinese infrastructure project. That description undersells what it actually is. It is a demonstration that the geography of AI compute is not fixed. It can be changed — if you have the authority and the will to change it.
China moved its AI compute to where the electricity was cheap, the climate was cool, and the land was available. It built the transmission lines, the fiber optic cables, and the data centers. It required the green power. It coordinated the grid.
The United States has the same cheap electricity in West Texas, the same cool climate in the Pacific Northwest, and the same available land in the Great Plains. What it does not have is a mechanism for routing the computation there. Its data centers are being built where the users are, and its users are being asked to pay for the infrastructure that serves them.
The grasslands of Inner Mongolia are not a natural home for AI. They became one because China decided they should be. The United States can build data centers. Whether it can decide where they go is a different question.
Sources:Ministry of Industry and Information Technology via C114 (July 2026); CNR (August 22, 2026); Guiyang Daily (August 23, 2026); China News Ningxia (May 2026); China News Ningxia (August 31, 2026); CCTV+ (August 9, 2026); China Industry News (August 21, 2026); China Energy News (September 14, 2026); Gansu Economic Daily (September 10, 2026); Shanghai Metals Market (July 24, 2026); 21st Century Business Herald (August 26, 2026); National Development and Reform Commission press conference (August 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.
Forecasts from third-party analysts can change with market conditions.
Cost and pricing examples are point-in-time estimates; actual rates vary.
Country and company comparisons rely on public reporting, not operational data.
This sector moves fast; timelines and deal terms may be updated later.
Company deals and regulatory rulings may evolve; verify current status.
AI infrastructure is changing quickly; claims can become outdated soon.
Sources
- Ministry of Industry and Information Technology via C114 (July 2026)
- CNR (August 22, 2026)
- Guiyang Daily (August 23, 2026)
- China News Ningxia (May 2026)
- China News Ningxia (August 31, 2026)
- CCTV+ (August 9, 2026)
- China Industry News (August 21, 2026)
- China Energy News (September 14, 2026)
- Gansu Economic Daily (September 10, 2026)
- Shanghai Metals Market (July 24, 2026)
- 21st Century Business Herald (August 26, 2026)
- National Development and Reform Commission press conference (August 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.; Forecasts from third-party analysts can change with market conditions.; Cost and pricing examples are point-in-time estimates; actual rates vary.; Country and company comparisons rely on public reporting, not operational data.; This sector moves fast; timelines and deal terms may be updated later.; Company deals and regulatory rulings may evolve; verify current status.; AI infrastructure is changing quickly; claims can become outdated soon.