A September 28 AIDC Daily/BUPT research briefing, drawing on recent academic papers and external energy analysis, highlights a constraint that can be easy to miss in the race to add AI capacity: servers are only one part of the infrastructure equation. Power availability, grid connections, workload timing, and cooling systems are becoming linked operating questions for data centers.
The briefing groups this research under coordination between computing and electricity. Its framing points to rising attention around AI data-center load, power density, and energy constraints. It also carries an important qualification: its keywords track research themes and are not, by themselves, factual conclusions about conditions across China. That makes the material more useful as a map of the engineering agenda than as evidence of a nationwide shortage or a single new government restriction.
For operators, investors, and equipment suppliers, the practical message is still significant. Building more accelerated computing capacity requires decisions about where that capacity sits, when it runs, how it receives electricity, and how its heat is removed. Those decisions can shape the economics of AI infrastructure before a new server is switched on.
Why AI capacity now depends on grid access and timing
The research highlighted by the AIDC Daily/BUPT briefing treats AI data centers as large and dynamic electrical loads. A September 10 entry on power architecture describes energy availability and grid-connection capacity as primary deployment constraints in its technology perspective. The same research discussion connects grid integration with workload orchestration, cooling systems, on-site resources, energy storage, and power-delivery architecture.
That combination matters because AI demand is not necessarily a fixed, evenly distributed load. A center may have computing hardware available while electricity conditions vary by site or time. If a deployment assumes that spare capacity can absorb traffic instantly, it may overlook the operational limits imposed by power availability and grid conditions.
A September 9 preprint examines the problem through large-language-model inference workloads that can be routed across sites. Its modeled cases assume electricity can constrain an AI data center at different times. The paper proposes jointly deciding where models are committed and where requests are sent, using power limits and electricity-price signals as inputs.
Its case studies, based on real-world data, report a 100% service rate under time-varying grid conditions and a 29.0% reduction in total operating cost. Those figures are modeled results from the paper’s cases, not measured achievements from a Chinese data-center fleet or a named commercial operator. Their importance is the operating logic: planning where models are committed and how requests move can be part of capacity management when power conditions change.
The briefing’s focus differs from EastFrontier’s separate report on Chinese prefab data centers targeting the U.S. AI buildout. That earlier article concerns a different deployment subject; the research briefing here focuses on power conditions, workload routing, and cooling.
Cooling is a material share of data-center electricity use
Cooling is often discussed as a facilities detail, but the International Energy Agency places it directly inside the electricity picture. The IEA says servers account for around 60% of electricity demand in modern data centers on average. Cooling and environmental control range from about 7% of consumption in efficient hyperscale facilities to more than 30% in less-efficient enterprise facilities.
That range does not establish a China-wide cooling crisis, and it does not show that one cooling technology has been adopted across Chinese AI sites. It does show why cooling efficiency can materially affect the amount of electricity available for computing. A facility with the same server fleet can have a different total power requirement depending on its environmental-control performance.
The IEA reports that data centers used about 415 terawatt-hours globally in 2024. In its Base Case, China’s data-center electricity consumption rises by around 175 terawatt-hours, or about 170%, from 2024 to 2030. These are IEA estimates and model projections, not Chinese government statistics and not an observed 2026 total.
The geographic concentration of data centers adds another layer. The IEA cautions that concentrated facilities can make grid integration more difficult even when data centers remain a limited share of total global electricity demand. Its Base Case puts global data-center electricity use at about 945 terawatt-hours by 2030 while noting uncertainty and bottlenecks.
For the industry, the implication is straightforward but not simple: a new AI site must be evaluated as a combined electrical and thermal system. Power-delivery architecture, storage, grid conditions, and cooling performance interact. Improving one component may not solve a constraint elsewhere, particularly if the site’s connection cannot support the planned operating profile.
China’s energy planning links compute demand with clean power
A China-specific policy and siting lens comes from Carnegie Endowment for International Peace analysis published July 31. Carnegie argues that China’s 2026–2030 energy plan explicitly links computing power with the electricity system and calls for green direct connection. The analysis interprets that language as an apparent strategic shift toward connecting compute demand with clean-energy supply.
Carnegie also points to major clusters in places such as Inner Mongolia and Ningxia. That observation is about public-document interpretation and the location of major clusters. It does not prove that every project has grid access or that all projects use renewable power. Nor does it establish a blanket rule for data centers nationwide.
The policy direction is relevant because the location of compute affects both energy economics and network operations. If computing demand is coordinated with available electricity resources, workload placement may become an infrastructure decision rather than a purely technical choice. That is consistent with the preprint’s focus on advance model commitment and cross-site routing, although the paper is a research proposal and not evidence of a live national operating system.
For separate historical background, EastFrontier has covered Damao’s AI-energy coordination financing and AI in energy operations. Neither article supplies evidence about a particular data center discussed in this briefing.
The central test for China’s AI buildout is therefore not whether more servers can be ordered. It is whether projects can secure dependable grid connections, manage time-varying electricity conditions, and control cooling demand while keeping services available. The briefing and cited research do not establish a nationwide restriction, a confirmed national shortage, or a completed cooling rollout. They identify a more practical question for the next phase of capacity growth: can power, compute, and thermal systems be planned as one operating asset?