Everything About AI
NextFin News - More than half of China's data-center project pipeline now sits across the country's north and northwest, far from the capital markets and fiber-rich megacities that built the nation's digital economy, according to a new research analysis published Tuesday. Beijing's bet that artificial intelligence would redraw China's economic geography is no longer a policy ambition. It is a construction reality, and it is moving fastest in places like Ulanqab, a grassland city of fewer than 2 million people in Inner Mongolia.
The Situation: A Buildout With a Name and a Map
The shift has a name in Chinese policy circles: dongshu xisuan, or "Data in the East, Computing in the West." Launched in 2022, the campaign established eight national computing hubs and ten data-center clusters, five of them in the resource-rich west. Four years on, the rebalancing is measurable. By the end of March 2026, intelligent-computing capacity built within the eight national hubs accounted for more than 80% of the national total, China's state media reported, citing the National Data Administration.
The scale of the buildout is the story. China is on pace to more than double the live IT capacity in its data centers over the next five years, with total power demand jumping 151% to 329 terawatt-hours by 2030 - about 2.4% of the country's total electricity consumption, the analysis found. Rystad Energy frames the physical expansion in starker terms: installed data-center power capacity of 32 gigawatts at the end of 2025 is set to reach 40 GW by the end of 2026 and top 60 GW by 2030, a near-doubling in five years, with power consumption rising at an annual rate of 19% between 2025 and 2030.
Ulanqab illustrates the transformation. The city has lined up 66 major computing-center projects and passed 71,000 petaflops of aggregate capacity as of August 2026. Its streets are named after Apple and Huawei. An Alibaba Group data center sits in the rolling grassland. And DeepSeek, the AI startup that stunned Silicon Valley by training powerful models at a fraction of Western cost, is planning a roughly 1-gigawatt AI data center there, with partial operations expected by late 2027 or early 2028. In Hohhot, the regional capital, more than 30 data centers were under construction within a 3-kilometer area. In Qingyang, Gansu, 22 new data-center buildings are planned for 2026, with computing capacity expected to exceed 200,000 petaflops.
"Every day, there are hundreds of workers," said a woman surnamed Yin, who runs a food stall serving plates of chicken, egg and tofu skin to construction crews at a busy Ulanqab intersection for 10 yuan, about $1.50. "It never stops."
The logic is simple physics and simple economics. AI training clusters do not need to sit next to Shanghai's financial district; they need cheap, abundant electricity and cool air to dissipate heat. The megacities have the connectivity. The north and northwest have the power. And for the first time, the power matters more.
Why the Grid, Not Capital, Decides Where AI Gets Built
The first-order read of China's data-center migration is cost arbitrage. The second-order truth is that the grid has become the binding constraint on AI everywhere, and China is answering it with geography rather than just generation. That distinction matters because it determines whether this relocation is durable or merely a subsidy cycle.
Modern AI racks demand 100 kilowatts or more, versus roughly 10 kW for a traditional server rack. A single gigawatt-scale AI facility draws as much power as a mid-sized city. In the Yangtze River Delta - which still tops the analysis's competitiveness ranking on connectivity, land permits and energy availability - that load cannot be absorbed. Electricity costs are higher, and energy-consumption quotas cap expansion. The region's data-center capacity is expected to be constrained over time precisely because it is the most developed.
That is the transmission mechanism behind the migration: power-intensive, latency-insensitive workloads - model training, batch inference, scientific computing - move to where electrons are cheap and cool air is free; latency-sensitive services - payments, ride-hailing, live recommendation engines - stay near users. China is not relocating its data centers wholesale. It is splitting the compute stack by workload, and that split is what makes the geography durable rather than cyclical. A cost-arbitrage cycle reverts when subsidies end. A workload split does not, because the physics of latency and the economics of power are both permanent.
The capacity numbers bear out a structural shift rather than a spending wave. China's intelligent-computing capacity reached 1.59 million petaflops by the end of 2025 - the world's second-largest pool of high-end AI compute, behind only the United States - and climbed to 1.882 million petaflops by the end of March 2026, up 18.2% in a single quarter. A research note by International Data Corporation projects a five-year compound annual growth rate of 46.2% through 2028. A cyclical buildout does not compound at that rate. A regime change does.
The demand backdrop tightens the utilization math further. Wood Mackenzie forecasts China's data centers will quadruple their electricity consumption to 774 terawatt-hours by 2030, and by 2060 they could account for 17% of the country's total electricity use. A study published in Nature estimates that China and the United States together will drive nearly 80% of global data-center electricity-consumption growth, with worldwide demand nearing 900 TWh by 2030, up from about 300 TWh in 2020. Even at the conservative end of these forecasts, the inland hubs are being built into a demand curve that is still steeply upward.
The Policy Hand Is Visible - and Getting Heavier
This is not a market discovery. It is a state-directed relocation with market incentives layered on top. Hub designation brings investment and preferential access to digital infrastructure, but it also comes with strict targets on facility utilization, energy efficiency and clean power. New data centers in hub regions must source more than 80% of their annual electricity use from renewable sources, either through direct connections or by purchasing green energy certificates.
Beijing has also moved to prevent the overcapacity that has plagued other state-priority industries. In June 2026, three agencies including the National Data Administration restricted the development of small data centers and required new projects to locate in centralized hubs. Large projects now need national approval before construction can begin. The policy is explicitly trying to channel local enthusiasm into coordinated capacity rather than duplicative shells.
The 15th Five-Year Plan, covering 2026 to 2030, codifies the next phase: "compute-power synergy," coordinating computing infrastructure with green electricity supply. Wood Mackenzie describes the shift as a move from "Power Follows Compute" to "Compute Follows Power." That phrase captures the regime change in one line - the scarcest input is no longer silicon, it is electrons, and infrastructure now goes where the electrons are.
Efficiency targets reinforce the point. The average power usage effectiveness of China's ultra-large computing facilities fell to 1.34 by the end of 2025, according to the National Data Administration's digital-development report. New large and mega data centers must already achieve a PUE of 1.25 or lower, while projects in national computing hubs face a stricter threshold of 1.2. Cool Inner Mongolian air is not just cheaper. It is regulatory arbitrage - a natural cooling advantage that translates directly into a permitted project.
The Companies Racing to Build
The pipeline is filling with named players, not anonymous capacity. Alibaba Cloud announced a three-year AI and cloud infrastructure investment program worth 380 billion yuan (about $52 billion), which could rise to $69 billion, expanding campuses across Hangzhou, Nantong, Heyuan, Ulanqab and Shaoguan. ByteDance is targeting more than 1.2 GW of AI data-center capacity in 2026 and, according to people familiar with the matter, plans to spend $14 billion on Nvidia chips this year. Kuaishou is investing about 10 billion yuan ($1.4 billion) in a Ulanqab campus designed to accommodate roughly 300,000 servers.
The AI startups are not waiting for the cloud giants. DeepSeek's planned 1 GW Ulanqab facility would be far larger than anything now operational from a Chinese company, though still behind the 3 GW and 5 GW projects announced by American hyperscalers. Z.AI, formerly Zhipu, is working on its own 1 GW venture and plans to fill it with only Chinese-made chips - a constraint imposed by U.S. export controls, and a signal that the domestic supply chain is being stress-tested at a scale no export regime intended.
"Lower-cost power, cooler climates and greater land availability in these regions are attracting developers, while power-intensive, less latency-sensitive workloads are migrating from established demand centers."
The researchers wrote, calling north and northwestern China "the main growth pole for new development, despite having a much smaller operating fleet."
The equipment chain is adapting to the new density. Chindata has unveiled a combined power-and-cooling solution for AI data centers designed to support 200 kW per rack, pairing it with plans for a gigawatt-scale campus. Liquid cooling is rapidly displacing traditional air cooling as per-rack density leaps from 10 kW to 30 kW to 100 kW and beyond. The buildout is not just concrete and servers; it is a forced upgrade of the entire power and thermal stack.
The Counter-Thesis: Overcapacity and the Megacity Pull
The strongest case against this narrative is that it is a policy-driven capex cycle, not a durable reallocation. China has a documented record of local governments racing to build capacity ahead of demand - in solar panels, in electric vehicles, in property. Data centers are the latest candidate. North and northwestern China could account for a third of national live capacity by 2030, but that still leaves two-thirds elsewhere, and the megacities retain better connectivity even without the energy abundance. If AI demand growth slows, or if the domestic chip supply chain cannot deliver at scale, the inland hubs risk becoming underutilized shells - expensive monuments to a policy target rather than productive infrastructure.
There is also a latency ceiling. Not every workload can migrate. Financial transactions, real-time recommendation engines and consumer applications that require single-digit millisecond response will remain anchored to eastern demand centers. The western buildout is real, but it serves a specific slice of the compute stack - training and batch work - and that slice may not grow as fast as the headline capacity numbers imply if model efficiency improves faster than usage.
The answer to the overcapacity charge is twofold. First, the approval gate is new and binding: small, speculative projects are now blocked, and large ones need national sign-off. That is a materially different policy posture from the open-door buildouts of the solar and EV eras. Second, the demand signal is not domestic alone. China and the United States are projected to drive nearly 80% of global data-center electricity-consumption growth, and China's own data-center electricity use is expected to rise fivefold to more than 250 TWh by 2030. Even at the low end of forecasts, the utilization math is tighter than in previous state-led buildouts.
The falsifying signal is specific and observable. The National Data Administration has set a national capacity-utilization target of at least 60%, with hub-level PUE thresholds of 1.2. If the reported hub utilization rate falls materially below 60% for two consecutive reporting periods, or if the average PUE of new hub projects drifts above 1.25, the structural-relocation thesis is wrong and this is a subsidy-driven cycle. Watch the administration's utilization disclosures and the hub-level PUE audits - not the ribbon-cuttings.
Who Wins, Who Is Exposed, and What Comes Next
The beneficiaries and the exposed are now legible. Winners include the inland hub provinces - Inner Mongolia, Gansu, Ningxia, Guizhou - and the equipment chains that serve high-density, liquid-cooled facilities: cooling systems, power distribution, and prefabricated modular construction. The domestic AI-chip supply chain is being handed a forced march: Z.AI's commitment to Chinese-made silicon means Huawei Ascend, Hygon Information and Cambricon will be tested at a scale no export regime intended. Cambricon achieved its first annual profit in 2025, and Huawei Ascend held approximately 49% of domestic AI-accelerator shipment share - the inland buildout is the stress test that will determine whether that share is durable.
Exposed are the eastern megacities' data-center economics. As energy quotas bind and electricity costs diverge, new hyperscale capacity in Shanghai, Beijing and Shenzhen becomes harder to permit and more expensive to run. Third-party operators with portfolios concentrated in tier-one metros face a choice: follow the workloads inland, or accept a shrinking share of the growth market. GDS Holdings, China's largest third-party data-center operator, operates mainly in Beijing, Shanghai, Shenzhen, Guangzhou and Chengdu - precisely the markets where land, power quotas and government approvals are scarcest. Its newer facilities run at a PUE of roughly 1.3 to 1.45, above the 1.2 hub threshold, which is a competitive disadvantage as the policy tightens.
Time horizons point in different directions, and collapsing them into one verdict would be a mistake. In the short term, the buildout is capex-heavy and sentiment-driven - a policy cycle that can overshoot, and local enthusiasm can outrun near-term utilization. In the medium term, utilization rates and power costs will separate the viable hubs from the political ones; the 60% utilization target and the 1.2 PUE threshold are the filters. In the long term, the structural call stands: compute follows power, and China's power is in the north and west. That is a decade-long reallocation, not a two-year spending wave.
Three scenarios frame the path. The base case is that the eight hubs absorb the majority of new AI training capacity through 2030, with north and northwest live capacity reaching a third of the national total while the megacities retain latency-sensitive inference. The upside case is that domestic chip yields improve faster than expected, unlocking the 1 GW-class projects from DeepSeek and Z.AI ahead of schedule and pulling more of the compute stack westward. The downside case is that AI demand growth disappoints and export controls tighten further, leaving inland campuses underutilized and forcing a policy pivot back toward eastern consolidation.
What to watch, concretely: the 60% national capacity-utilization target, hub-level PUE readings against the 1.2 threshold, and whether DeepSeek's Ulanqab facility hits its late-2027 or early-2028 partial-operation milestone. Those three data points will tell you, before the headlines do, whether this is a regime change or a rally.
China's AI race was supposed to be won with chips. It may be decided with electrons - and the electrons are in the grassland.
Explore more exclusive insights at nextfin.ai.
