The wind farm sits sixteen kilometres from the servers, connected by a line that exists for a single customer. Nothing else can draw on it.
That arrangement is now the template. In China's northern grasslands, a data centre no longer waits for a transmission corridor to reach it; it finances the generation, signs a dedicated line, and treats the weather as a supply risk to be managed with batteries.
On 30 October, the National Energy Administration put a number on why this is happening.
Key takeaways
- The NEA expects national compute electricity use to grow by more than 100 TWh a year through 2030, reaching 800 TWh, or about 6% of all electricity.
- Compute centres used roughly 170 TWh in 2025, 1.6% of national demand — so the plan implies a 4.7-fold increase in five years.
- Only 12 compute projects had won green direct-connection (绿电直连) approval by late September, just over 10% of all such approvals nationwide.
- Ulanqab's first such project pairs 300 MW of wind and solar with 45 MW of storage and a 16 km dedicated line, producing 848 GWh a year.
- Power there runs RMB 0.32–0.36 per kWh (≈ US$0.045–0.050), and the city has approved 14 such projects totalling 5.46 GW of renewables.
What the regulator actually said
At a press conference in Beijing, Tan Hongjiang, deputy director of the NEA's power department, said compute electricity demand will rise by more than 100 billion kWh a year through 2030 — a forecast that treats AI as a structural load, not a cyclical one. Internet data services used 41.6% more electricity in the first three quarters of 2026 than a year earlier, and national installed capacity reached 4.13 TW by September, up 11.2%.
The regulator also described the mechanism it is betting on: green direct connection (绿电直连), where a dedicated line carries power from a nearby renewable plant straight to a compute campus, bypassing the general grid. Twenty-four provinces have issued supporting rules. A joint notice from the National Development and Reform Commission and the NEA, numbered 发改能源〔2026〕688号, extended the model to multiple users.
The arithmetic nobody wants to do
One hundred terawatt-hours a year is 11.4 GW of generation running flat out, all year, forever. Wind in Inner Mongolia runs at perhaps a quarter to a third of that, so the renewable equivalent is closer to 40 GW of nameplate capacity added annually — on top of replacing existing fossil generation and meeting ordinary demand growth.
That is why "compute-power coordination" (算电协同) entered the government work report and the 15th Five-Year Plan outline this year. It is not a slogan about efficiency; it is an admission that the binding constraint on Chinese AI has moved from accelerators to interconnection queues.
The economics favour it anyway. Ulanqab's government puts landed electricity at RMB 0.32–0.36 per kWh and says that, with cool air and a PUE below 1.2, operating costs run 30–50% below eastern tier-one cities. A hall saving a third of its power bill can afford to be 500 km from its users.
How a cluster gets power now
The reference project is Zhongjin Data's campus in Ulanqab, which went into operation in July 2025 as the country's first data-centre green direct-connection project with integrated source, grid, load and storage.
Read it as a template: 200 MW of wind and 100 MW of solar on site, 45 MW of batteries, a dedicated line to a campus 16 km away, 848 GWh of self-consumed green power a year, and a planned 25,000 racks backed by RMB 9.1 billion (≈ US$1.27 billion) of investment.
The engineering problem is not generation. It is matching a load that never sleeps to a supply that does. A control platform forecasts output, dispatches the batteries and holds the campus draw inside a band, which is what makes an intermittent source usable for training runs that cannot be interrupted.
The city has since approved 14 green direct-connection projects totalling 5.46 GW of renewables, and plans 7 GW of new supply capacity during the 15th Five-Year Plan. Latency is no longer the objection: two dedicated fibre routes to Beijing measure about 4.2 ms round trip, close enough for inference serving.
Why only twelve
Twelve sounds small against 800 TWh. It is small, and that is the point. A green direct-connection project needs a renewable site, a load of sufficient scale, a dedicated line with its own protection scheme, storage sized to the load's ramp rate, and a province willing to allocate land and interconnection rights. Each is a negotiation, and the twelve are the projects that closed all of them.
There is a second edge. Self-generated power is cheap, but a campus that makes its own electricity also owns its own reliability. Grid backup still has to exist for the still days, and that capacity is priced separately.
Honest limitations
The demand forecast is a government projection, not a measurement, and projections of AI electricity use have been revised upward repeatedly in every country that has published one. The 11.4 GW and 40 GW figures are my arithmetic on top of the NEA number; they assume capacity factors that vary by province and technology.
The Ulanqab cost figures come from the municipal government and a state media visit, not from audited company accounts. I have not seen the tariffs written into any compute project's power purchase agreement, and the 30–50% operating-cost advantage is a claim about a composite, not a single line item.
Nor does this article assess water. Several of the favoured regions are water-stressed, and cooling choice is becoming a siting constraint in its own right.
What readers can do now
- If you are underwriting a Chinese compute project, treat the power contract as the critical path, not the GPU order. Twelve approvals in a year is the real throughput of the system.
- Ask for the capacity factor and the storage duration behind any quoted green share. A 100% renewable claim with two hours of batteries is a different asset from one with eight.
- For latency-sensitive inference, check the fibre route before the power price. Ulanqab's 4.2 ms to Beijing is the reason the western model works at all.