The transmission gantries arrive before the buildings do. At a half-built campus in the west, a substation hums behind a fence against a load not yet connected; two halls beside it are shelled in and empty, cooling pipework installed, racks absent.
The site was planned backwards from the electricity rather than forwards from the servers. Power came first because power takes longest.
That ordering is the story of Chinese compute (算力).
Key takeaways
- The national computing network is a 15th Five-Year Plan priority; NDRC projects >RMB 7 trillion (≈ US$986bn / HK$7.7tn) of investment in it and related fields in 2026 (CGTN).
- SemiAnalysis puts China's data-centre capacity above 24 GW by end-2026, against roughly 56 GW in the US and 14 GW in EMEA (25 September 2026).
- Compute-centre electricity demand is forecast at about 300 TWh in 2026, versus roughly 260 TWh in the US and 150 TWh in Europe (Beijing Daily, January 2026).
- AI data-centre load is estimated above 5 GW in 2026, with 6–8 GW under construction (industry estimate).
- CAICT counted 42 ten-thousand-card compute clusters in China as of March 2026.
A grid, not a market
Most countries let demand pull data centres to cheap power and land. China is doing something closer to rural electrification: build to a plan, connect the provinces, let utilisation catch up.
The 15th Five-Year Plan makes the national integrated computing network a priority. The NDRC projects more than RMB 7 trillion (≈ US$986bn / HK$7.7tn) of investment in the network and related fields during 2026 alone (CGTN, 18 June 2026). That number is not an AI budget. It covers transmission, cooling, buildings, fibre and the compute inside them: model training is one load among many on a national grid.
The tell is a phrase. The 2026 Government Work Report put "computing–power coordination" (算电协同) into the text for the first time (SCIO briefing, 20 March 2026). When a work report coordinates compute with electricity, the object of policy is no longer the chip but the electron.
The capacity is arriving faster than the demand is located
SemiAnalysis estimates China will pass 24 GW of data-centre capacity by the end of 2026, against roughly 56 GW in the United States, 14 GW in EMEA and 15 GW in Asia-Pacific excluding China (25 September 2026). On that measure China is closer to the leader than the chip debate implies.
The money is concentrated: Alibaba, Tencent and Baidu spent about US$20 billion on capex in Q2 2026, more than double a year earlier, all three with negative free cash flow. ByteDance accounts for roughly one-fifth of delivered capacity, mostly leased (SemiAnalysis).
CAICT counted 42 ten-thousand-card compute clusters as of March 2026. A cluster of that size is not a product; it is a piece of civil infrastructure with a substation attached.
The physics arrives before the software
Compute-centre electricity demand in China is forecast at about 300 TWh in 2026, against roughly 260 TWh in the United States and 150 TWh in Europe (Beijing Daily, January 2026). China trails the US on capacity but leads on electricity, implying a different mix of utilisation and efficiency. Industry estimates put AI data-centre load above 5 GW in 2026, with 6–8 GW under construction. Globally, Gartner expects data-centre electricity use of 565 TWh in 2026, up 26%, with AI-optimised servers rising from 20% to 31% of the total.
Three constraints decide whether it gets used:
- Connection. A hall needs a substation, a line and a provincial allocation, and the queue is measured in years.
- Cooling. Water and heat rejection set density limits, and the trade-off between water and efficiency varies by site.
- Interconnection. Training a large model (大模型) across thousands of cards is a network problem before it is a capacity problem, which is why clusters are sized in ten-thousand-card units.
Utilisation is the only number that matters
Capacity is a nameplate. Utilisation is a business. The gap between them is where the plan will be judged.
National token consumption averaged 140 trillion a day by March 2026, up more than 40% from 100 trillion at the end of 2025 (National Data Administration). That is evidence of demand somewhere, not evidence of where. A cluster in a western province can be fully booked by a Beijing lab and still leave the local grid, the local operator and the local industrial park with nothing to show.
If ByteDance takes one-fifth of delivered capacity mostly as a tenant, the owner's utilisation and the operator's utilisation are different numbers, and only one is usually disclosed.
Price is the other signal. In March 2026, Tencent Cloud raised prices on some model inputs by 463%, Alibaba Cloud raised AI compute and storage by up to 34%, and Baidu Cloud by 5–30% (China Development Network; China Economic Net). Rising prices with 24 GW of capacity suggest the constraint is not buildings but the expensive, connected, cooled kind of capacity.
What would falsify the plan
- Grid delays that leave halls dark for a year after completion.
- Provincial utilisation data that never appears in any public disclosure.
- AI load growth running past the 6–8 GW under construction, turning a capacity story into a queue.
Honest limitations
The 24 GW figure is a SemiAnalysis estimate of capacity, not measured consumption, and definitions vary by whether they count shell space or energised racks. The 300 TWh forecast comes from a single cited outlet and is a projection, not a metered total. The RMB 7 trillion figure covers the network and related fields, not AI specifically, so it cannot be read as an AI subsidy. No public series gives utilisation by province or by operator, and the 5 GW AI load figure is an industry estimate without a published method.
What readers can do now
- Operators: site near power and interconnection before you site near customers; in this system the queue, not the rent, sets your timeline.
- Investors: ask for energised megawatts and billed GPU-hours rather than planned capacity, and treat signed leases as a claim rather than revenue.
- Researchers: keep capacity and consumption in separate columns. Most published China compute numbers are the first, and almost none are the second.