China's AI data centers' green power targets face real-world challenges: the conflict between computing demand and grid stability.

AI Computing Power Expansion Meets Energy Bottleneck: The Real Challenges Behind China's Green Data Center Targets

China is advancing its AI infrastructure construction at an unprecedented pace, but the resulting surge in electricity demand is creating a structural conflict with its green transition goals. The 2026 government work report explicitly pledged to strengthen the coordination between computing power and the power grid, setting an ambitious target: by 2030, 80% of electricity consumption in the data center industry should come from renewable sources, compared to just 11% in 2023.

However, the latest industry discussions show that achieving this vision is far more difficult than anticipated. The electricity demand characteristics of AI data centers—especially the pulsed peak loads and the high cost of GPU assets—make them incompatible with the flexibility requirements of renewable energy supply. At the same time, grid operators' concerns that direct-supply models will erode their revenue have become a hidden obstacle to policy implementation.

Background: Computing Power Is Electricity, Green Promises Hit Hard Reality

China's AI data center construction is rapidly driving electricity demand. According to Pei Shanpeng, director of the State Power Investment Group, revealed at a recent industry conference, China's data center electricity demand will increase by 300 billion to 500 billion kilowatt-hours between 2026 and 2030—the lower bound of this figure is already close to the total annual electricity consumption of the United Kingdom. During this period, the incremental electricity consumption of data centers will account for 18% of the total national electricity consumption growth.

To address this trend, the Chinese government is vigorously promoting direct green electricity trading and dedicated transmission network construction, attempting to connect renewable resources in the west with computing power centers in the east. But the reality is that the unique pattern of AI loads is breaking the traditional assumption of "adjustable load" in industry.

Technical Analysis: Why AI Data Centers Are "Inflexible" Electricity Consumers

Unlike heavy industries such as electrolytic aluminum and steel, which operate continuously and stably, the electricity consumption of AI data centers is highly dynamic. GPU training tasks are typically submitted in batches to a queue; once started, operators want them to run at full capacity to complete as quickly as possible, avoiding expensive GPU idling. Pei Shanpeng pointed out: "GPUs are very expensive. Once purchased, operators want to use them as fast and as intensively as possible." This means data centers have almost no willingness to adjust their loads.

In addition, the peak burstiness of AI inference workloads is even stronger—for example, when traffic to a large model application suddenly spikes, electricity consumption can double within minutes. This unpredictability makes power grid dispatch extremely difficult. In contrast, traditional industries can participate in demand response (such as reducing production) from the grid side, while AI data centers can hardly cut loads without interrupting services or extending training cycles.

Wang Zelin, deputy director of the State Grid Jibei Electric Power Research Institute, estimates that if data centers could achieve a 15% load adjustability, it would significantly reduce the pressure on grid expansion over the next three to five years. However, the industry generally believes that even this level is difficult to achieve.

Enterprise Impact Analysis: Cost, Operations, and Compliance Pressure### Cost Impact - CAPEX: Building direct power supply or dedicated green energy lines requires additional investment in transmission infrastructure. For cloud service providers and hyperscale data center operators, this cost may be passed on to rent or cloud service prices. - OPEX: Without subsidies, green electricity is usually more expensive than thermal power. Especially when backup thermal power is needed for peak shaving at night or on cloudy days, overall electricity costs may rise. Pei Shanpeng pointed out that the current push for green electricity is more about emission reduction than cost reduction.

Deployment Impact - Site selection constraints: Enterprises may be forced to build data centers near renewable energy-rich areas (e.g., the west), but this increases network latency, which is detrimental to latency-sensitive businesses like AI inference. - Power supply reliability: Reliance on green electricity may cause power supply fluctuations. Enterprises need to deploy additional energy storage or backup generators, further increasing CAPEX.

Operations and Maintenance Impact - Load management: Enterprises need to develop smarter load scheduling algorithms to arrange training tasks when green power is abundant, but this may require large-scale modification of existing AI workflows. - Compliance pressure: They may face local government assessments on the proportion of green electricity usage. Failure to meet standards could affect business licenses or carbon quotas.

Safety and Compliance - Grid stability risk: Large-scale integration of renewable energy may increase grid frequency fluctuations, posing hardware safety risks to precision GPU equipment. - Carbon accounting requirements: Enterprises need to accurately measure the carbon emission factor per kilowatt-hour to meet ESG reporting and export compliance (e.g., EU Carbon Border Adjustment Mechanism).

Market Competition Analysis: The Game Among Cloud Providers, Data Centers, and the Grid

Cloud Providers and Data Center Operators - Beneficiaries: Hyperscale data centers that have early deployed direct green power supply and energy storage facilities (e.g., GDS, Qinhuai Data) may receive policy priority support; cloud providers with western hubs (e.g., Alibaba Cloud Ulanqab Data Center) have an advantage in green electricity access. - Parties under pressure: Older data centers relying on eastern thermal power face high carbon costs or must undergo transformation; small hosting service providers struggle to absorb the green electricity premium.

Grid Enterprises - Grid operators are cautious about the direct power supply model. Experts point out that direct supply means the grid loses revenue from that portion of electricity sales, while still bearing the fixed investment recovery pressure of the transmission and distribution network. If demand slows, the grid's finances suffer. Therefore, State Grid and China Southern Power Grid may create implicit costs by increasing backup capacity fees or limiting grid connection capacity.

International Comparison - US tech giants like Google and Microsoft have achieved or committed to 100% renewable energy matching, partly through Power Purchase Agreements (PPAs) and Virtual PPAs. The uniqueness of the Chinese market lies in grid monopoly and regional barriers, making the PPA model difficult to replicate.

Industry Trend Observation: The Chinese-style Dilemma of Green Data CentersIn the short term, the goal of achieving 80% green electricity for AI data centers is almost impossible to fully realize before 2030. A more likely path is “orderly progress”: first requiring new data centers to allocate a certain proportion of green electricity or purchase green certificates, while simultaneously promoting the maturity of energy storage and intelligent dispatching technologies.

In the long run, China must address the institutional barriers in the electricity market: allow third-party power purchase agreements, establish cross-provincial green electricity trading mechanisms, and improve the capacity market to incentivize the grid to support distributed green electricity integration. The flexibility of AI workloads is not entirely unchangeable—through batch processing windows and latency-tolerant training tasks, they can partially align with green electricity supply curves. However, this requires a fundamental transformation in data center operation models.

Furthermore, while liquid cooling technology, high-efficiency UPS, and modular data centers can improve energy efficiency, they cannot solve the problem of unpredictable loads. The core contradiction lies in the conflict between the real-time demands of AI services and the intermittency of renewable energy.