By Guy Duport |
Standfirst China has shown that planning computing and power together can steer where AI data centers are built. But building computing capacity does not necessarily produce commercially useful demand. Utilities should borrow the coordination and insist on measured outcomes.
Capacity is not load
China reports nearly 2.5 million petaflops of intelligent-computing capacity, most of it in designated national hubs. Yet researchers have reported utilization rates of only 20% to 30% at some western data centers. That gap between installed capacity and commercially productive use is the most relevant lesson for electricity utilities. China demonstrates that computing and power-system planning can be coordinated. It also shows that coordination is no substitute for testing whether new facilities deliver metered, paying demand rather than merely meeting construction targets.
The International Energy Agency (IEA) estimated in its 2026 update that global data-center electricity use would rise from about 485 terawatt-hours (TWh) in 2025 to roughly 950 TWh in 2030 under its central outlook. That is a global forecast, not a connection schedule for any regional grid. Installed servers, secured land and requested megawatts are not metered demand.
What China built
China launched its “East Data, West Computing” initiative in 2022, identifying eight national computing hubs and ten data-center clusters. The objective is to move suitable workloads away from demand-heavy eastern cities toward regions with greater land and energy resources. In principle, this connects the location of computing facilities with electricity development and communications networks. However, the geographical logic of generation does not automatically match the technical or commercial needs of every computing workload.
In August 2026, China’s National Data Administration reported 2.45 million petaflops of intelligent-computing capacity, measured at 16-bit floating-point (FP16) precision, as of the end of July. It said more than 85% was located in eight national hubs and three computing–electricity coordination areas, while 1.45 million petaflops fell within the national monitoring platform. These are reported capacity and monitoring figures, not measured utilization or grid load. The International Energy Agency (IEA) estimated that Chinese data centers consumed approximately 100 terawatt-hours of electricity in 2024, while noting substantial data gaps. For utilities, the units matter: petaflops measure computing capacity, not megawatts of demand.
A second National Data Administration account described computing networks as closely linked with the power grid, water networks and communications infrastructure. Energy is no longer treated as an input purchased after a site has been selected; it is part of the location decision itself. That is the part of China’s model worth copying.
What happened: capacity outran use
Results have been uneven. Researchers writing for the Australian Strategic Policy Institute reported that some western data centers operated at utilization rates as low as 20% to 30%, compared with an initial policy ambition above 60%. These figures describe particular reported facilities, not a verified utilization rate across China. Data transfer costs, latency, customer proximity and access to technical expertise can offset the apparent savings from cheaper electricity. A western location cannot serve every eastern user simply because national planners designate it a computing hub.
For grid companies, the question is not whether a region has surplus annual generation. It is whether the location can supply sufficiently reliable power when needed and connect the data center to its customers at commercially acceptable latency. The same analysis reports renewable-energy curtailment above 30% in some western areas. That is a location-specific warning, not a national average. Neither installed renewable capacity nor nominal computing capacity establishes the hours, cost or reliability of actual electricity demand.
The IEA’s 2025 Energy and AI analysis estimated that roughly 20% of planned data-center projects globally could face delays unless grid-related constraints were addressed. It identified long connection queues, equipment lead times and opportunities to place computing capacity where grids are less congested. Coordinated siting helps, but it does not remove those engineering or economic constraints.
Domestic chips raise the electricity bill
Hardware matters as well. International Data Corporation (IDC) shipment figures reported by Reuters put Nvidia at approximately 34% of China’s data-center AI accelerator market in the first half of 2026 and Huawei at 28%. Those market shares measure shipped units, not comparable computing performance. For an illustration of energy trade-offs, specialist analyst SemiAnalysis estimated a roughly 559-kilowatt power budget for Huawei’s CloudMatrix 384 configuration, against about 145 kilowatts for Nvidia’s GB200 NVL72, with materially lower computing efficiency per watt in that particular comparison. These are model-based system comparisons, not independently metered operating data, and they must not be extrapolated to all domestic chips. The relevant utility question is what hardware the developer will actually install, how it performs under the intended workload and what electricity demand it creates.
For a utility, that should shape contractual diligence. A developer might request a large connection based on a hardware deployment that is delayed, downsized or redesigned. Conversely, a switch to less efficient systems could raise demand at an existing campus. Forecasting should therefore combine construction milestones with evidence of equipment procurement, power ramp-up plans and credible customer demand rather than accepting nameplate computing capacity as a proxy for consumption.
How North American grids are responding
North American regulators are confronting similar challenges through connection and market rules rather than national computing-hub targets. In June 2026, the US Federal Energy Regulatory Commission issued show-cause orders directing six regional grid operators to justify or reform large-load connection tariffs, including safeguards for existing customers. In Texas, Senate Bill 6 established new obligations for very large loads, including disclosures and potential demand reductions under defined grid conditions. Governor Greg Abbott also ordered an audit of data-center projects in the interconnection process in August 2026. Alberta’s system operator responded to an earlier queue exceeding 16 gigawatts by imposing an interim 1,200-megawatt connection limit, which it subsequently allocated. These are distinct legal approaches, but all expose the weakness of treating an application queue as a reliable demand forecast.
Four questions for utilities and regulators
China’s experience, good and bad, suggests four questions to settle before major network expenditure is committed, whether national infrastructure is largely state-directed or market-led.
1. Is the load real? The Chinese experience shows why capacity targets are not a substitute for paying customers. Distinguish early expressions of interest from funded projects with secured land, permits, equipment orders and phased electricity requirements. Publish alternative planning scenarios rather than a single optimistic number.
2. Is the location electrically rational? Lower electricity prices may not overcome latency and connectivity constraints. Compare available generation, substation capacity, transmission congestion, reliability and water availability with the requirements of specific workloads. Test alternative sites before committing to expensive reinforcement.
3. Who pays if the project changes? Underused computing facilities can leave costs with investors, governments or utilities. Connection charges, milestones, financial security and transparent cost-allocation rules should keep speculative projects from shifting disproportionate risks to households and existing industry. The precise safeguards must reflect local law and regulation.
4. Can demand become flexible? Closer coordination between computing and electricity can potentially shift some work away from periods of network stress. Assess which computing loads, onsite storage or backup resources can respond safely to system conditions. Not all AI operations are interruptible, so compensate flexibility according to its real technical and commercial costs.
Coordinate, then meter the results
The same accounting applies to countries attracting foreign-financed data centers. An installation within national borders can depend on imported accelerators, foreign cloud operators and distant customers, so the grid costs may stay local even if much of the commercial revenue does not. Regulators should establish who finances the enabling grid, who uses the computing and who bears the long-term risk before calling any of it sovereign.
The central lesson from China is organizational rather than ideological. Electricity grids, computing sites, communications links and industrial policy interact whether institutions coordinate them or not. China shows both the value of coordination and the cost of building toward capacity targets. Utilities should welcome credible data-center investment, but plan around connected, contracted and ultimately metered demand—not announced computing capacity.
Research and related analysis: Research Papers•