The headline numbers are shakier than they look
On October 8, 2026, the Financial Times reported, citing SemiAnalysis's China Datacenter Model, that China is using cheap energy and vast land to expand AI data-centre capacity. It put Ulanqab in Inner Mongolia at 89 signed data-centre project agreements by end-June 2026, with about 12.5 GW of operational and planned capacity. It also put national capacity at over 24 GW against about 56 GW for the US by end-2026. We relied on secondary summaries of the underlying FT piece. Other outlets give different Ulanqab figures: Crypto Briefing reports nearly 100 facilities completed or under construction since 2016, representing 12.5 GW of pledged capacity, not operating capacity. The report itself notes that its figures describe energy infrastructure, not AI computing power.
That caveat matters more than the headline. A gigawatt is a measure of electrical envelope. Whether it becomes useful AI compute depends on how many capable accelerators sit inside it. Nothing new was announced by regulators this week, so the useful question is which policies shaped the buildout and what they imply.
What 'East Data, West Computing' actually does
The project was launched in 2022 and sets up eight national computing hub nodes and 10 national data-centre clusters. Its aim is to move computing from the costly, power-constrained east toward resource-rich western regions, according to a summary of the 2023 implementation plan. The plan was issued jointly by five authorities, including the National Development and Reform Commission and the National Data Bureau. The Gansu provincial government's English-language portal carries the same Xinhua account of the plan. It set a target for new computing power in the national hubs to exceed 60% of the country's new computing power. The state-media summary adds that by the end of 2025, green electricity should exceed 80% of power in newly built hub data centres.
Inner Mongolia is one of the designated hubs, and Ulanqab's appeal follows the policy logic. Crypto Briefing reports that western regions pay roughly half what eastern users do for electricity, though it gives no named source for that price comparison, so treat it as indicative. Cool climate, cheap land and abundant wind and solar do the rest.
The strongest case for the state-directed model
Supporters of this kind of central planning make a serious argument. Data-centre load is lumpy, and uncoordinated siting can strain grids and push up household bills. A planner who ties new capacity to renewable generation and transmission can, in principle, avoid the congestion and local opposition that now slow US projects. China's approach also gives investors a predictable map of where approvals will be easy. On the evidence so far, the siting policy appears to be delivering on the physical side: tenant announcements in Ulanqab are large and numerous.
Where the argument gets weaker
The weakness is that siting policy solves the cheap part of the problem. Land and power are the inputs China has in surplus. The scarce input is advanced chips, and that is where US policy bites. Washington's January 2026 rule, Revision to License Review Policy for Advanced Computing Commodities (91 FR 1684, published January 15, 2026), moved licence review for Nvidia's H200 and comparable chips bound for China and Macau from a presumption of denial to case-by-case review. Exporters must certify four conditions: sufficient US supply, no diversion of foundry capacity needed by US customers, adequate security procedures at the recipient, and independent third-party testing in the US.
That is a narrower and more conditional channel than a market. A hall built for 1 GW can sit partly idle or be filled with less efficient domestic silicon. Either outcome lowers the compute delivered per gigawatt compared with a US site. So the 24 GW versus 56 GW comparison probably flatters neither side in a clean way: it overstates China's usable AI capacity if chips are the constraint, and it says little about efficiency.
What this means for policy
Two lessons follow, and both cut against reflexive responses.
First, the US should not read these figures as proof that it is losing a power race and respond with sweeping new restrictions on its own data-centre builders. The more useful takeaway is that US capacity is limited by permitting, interconnection queues and local approvals, which are domestic and fixable. Faster grid connection and predictable siting rules would do more for American AI than anything aimed at China's pledged gigawatts.
Second, export controls deserve evidence-based scrutiny rather than slogans. If chips are the binding constraint, controls are the lever that matters, and the January rule shows Washington is treating them as a negotiable, conditional instrument rather than a fixed wall. Proportionate policy would state clearly what capability thresholds it is protecting, review them as domestic Chinese chips improve, and avoid blanket measures that push buyers toward self-sufficiency. A tightly targeted regime with transparent criteria is easier to defend, and easier to adjust, than a broad one.
For China, the risk is different. Pledged capacity encourages overbuilding in remote hubs, where latency and a thin local customer base limit use. Planned siting can create stranded assets if the compute never arrives. The state-set 60% hub-share target is a measure of where capacity goes, not whether it is used.
What to watch
The numbers worth tracking are not pledges but utilisation: operating accelerators per hub, delivered compute, and the share of green power actually consumed. Until the FT and SemiAnalysis data can be checked directly, the sound reading is cautious: China has a real siting advantage in energy and land, and a real constraint in chips, and policy on both sides is shaping the gap between them.