China's Data Centers Join Power Market as Virtual Power Plants for the First Time
Computing Power Integration into the Grid: China’s Data Centers Enter Electricity Spot Market as Virtual Power Plants for the First Time, Ushering in a New Era of Bidirectional Coupling between “Electricity Prices” and “Computing Power”
In May 2024, China’s construction of a new power system achieved a landmark breakthrough: three national-level computing hubs—the China Unicom Guangdong–Hong Kong–Macao Greater Bay Area (Shaoguan) Data Center, the China Mobile Yangtze River Delta (Nanjing) Intelligent Computing Center, and the Beijing–Tianjin–Hebei (Langfang) Artificial Intelligence Innovation Base—officially participated in Guangdong Province’s electricity spot market as Virtual Power Plants (VPPs) via the Guangdong Energy Management and Ancillary Services Trading Platform (“Yue Neng Tou” Platform). This marks the first time data centers in China have moved beyond their traditional role as passive electricity consumers to become active market participants capable of autonomous response, flexible adjustment, and value creation. This milestone not only validates the engineering feasibility of the “computing follows electricity” technical pathway but carries deeper significance: AI computing infrastructure has formally entered the core of China’s electricity market-oriented regulation system, triggering a fundamental restructuring of computing cost composition and electricity pricing mechanisms.
Electricity Price Volatility Directly Impacts Large-Model Training Costs—The Era of “Marginal Computing Pricing” Accelerates
Historically, electricity costs account for 60%–70% of total data center operating expenses. Yet pricing models remain highly rigid—primarily relying on fixed catalog tariffs or long-term Power Purchase Agreements (PPAs), largely decoupled from real-time supply–demand imbalances that cause sharp electricity price fluctuations. With the implementation of the VPP model, data centers can now leverage their inherent load flexibility—including thermal margin in hot/cold aisles, buffering windows for non-real-time inference tasks, and fine-grained GPU training job scheduling—to declare adjustable capacity and responsive bidding prices in the spot market, participating in day-ahead, intraday, and even real-time peak-shaving and valley-filling activities. When midday solar generation surges in Guangdong and nodal electricity prices fall below ¥0.2/kWh, IDCs can automatically trigger high-priority model fine-tuning; when evening peak prices soar above ¥1.2/kWh, they intelligently pause batch-processing jobs, dispatch stored energy, or activate backup diesel generators. The electricity cost per large-model training session thus shifts from “fixed amortization” to “dynamic fluctuation,” directly mapping onto the marginal cost per token generated. Internal calculations by a leading cloud provider indicate that, in regions with typical peak-to-valley price ratios reaching 5:1, intelligent scheduling strategies can reduce annual AI training electricity costs by 18%–25%. This shift is driving algorithmic frameworks to embed “power-aware schedulers,” requiring LLM training optimization not only for FLOPs but also jointly for kWh.
Commercial Viability of Energy-Efficiency Technologies Has Reached Its Tipping Point: Dual Acceleration in Liquid-Cooling Penetration and Direct Green-Power Procurement
The explicit signaling of electricity prices has fundamentally reshaped the economic evaluation logic for data center energy-efficiency technologies. Liquid-cooling solutions—which previously relied heavily on policy subsidies or ESG narratives for adoption—now demonstrate clear return-on-investment (ROI) under the spot market regime: once rack power density exceeds 30 kW, air-cooled systems struggle to achieve PUE below 1.5, whereas immersion liquid-cooling consistently maintains PUE between 1.08 and 1.12. Assuming Guangdong’s average spot price of ¥0.52/kWh and annual operation of 8,760 hours, a single rack saves over 120,000 kWh annually—further enhanced by extended equipment lifespan and reduced maintenance costs—compressing the payback period for liquid-cooling retrofits to under 2.3 years. Industry data shows that domestic liquid-cooling server tender volumes surged 210% year-on-year in Q1 2024. Simultaneously, the closed-loop business model of “direct green-power supply + computing load absorption” is gaining rapid traction: the Shaoguan cluster has signed a bundled “green power–computing power” agreement with local wind farms, stipulating that each MWh of green electricity corresponds to 1.2 PFlops·h of intelligent computing services—securing low-cost clean power while providing renewable energy with a stable off-take outlet. This “source-load symbiosis” model is transitioning from pilot projects toward scalable replication.
Revaluation of Telecom Operators: A Paradigm Shift from “Electricity Bill Payers” to “Electricity Service Providers”
For data center operators, VPP status unlocks a second growth curve beyond traditional IDC operations. Beyond reducing their own electricity bills, their aggregated, highly flexible load resources can participate in ancillary service markets such as frequency regulation and spinning reserve. In Guangdong, frequency regulation ancillary service bids currently range from ¥6 to ¥12/MW·min—meaning a single 30-MW data center cluster holds annual ancillary service revenue potential exceeding ¥10 million. Even more profoundly, this triggers a transformation in valuation logic: capital markets are shifting from solely assessing “rack utilization rate” and “bandwidth revenue” to comprehensively evaluating “integrated source-grid-load-storage capabilities.” Operators with self-built PV/storage assets, provincial grid dispatch system connectivity, and deployed AI-based load forecasting models already command EV/EBITDA valuation premiums 35%–50% higher than industry peers. Equipment vendors likewise face structural opportunities: millisecond-response smart circuit breakers, multi-protocol-compatible edge IoT terminals, and SaaS platforms for load aggregation tailored to electricity markets are becoming standard procurement items for new IDC projects.
Systemic Challenges Remain: Cross-Provincial Transaction Barriers, Metering Accuracy, and Safety Boundaries Demand Urgent Resolution
Despite its promising outlook, widespread deployment still faces practical bottlenecks. Currently, spot markets operate on a provincial basis, preventing unified bidding across cross-provincial data center clusters; existing electricity meters—with typical resolution at 15-minute intervals—cannot meet VPPs’ millisecond-level response requirements; most critically, interrupting computing tasks may cause model training failures or violate service-level agreements (SLAs). Defining the technical red line between power-system regulation safety and computing-service reliability remains an open question, with no national standards yet established. The industry urgently calls for accelerated formulation of the Technical Specifications for Data Center Virtual Power Plant Grid Integration, clarifying core metrics including regulation response time, maximum adjustable capacity ratio, and fault-ride-through capability—and advocating for the establishment of insurance mechanisms covering computing-service interruption risks.
When computing power becomes a “flexible, programmable load” on the grid, the deep integration of two national strategic infrastructures—electric power systems and artificial intelligence—is now irreversible. This is not merely about connecting technical interfaces; it represents a paradigm fusion of energy economics and computational economics. In the future, AI competitiveness may well hinge on who best masters the peaks and troughs of the electricity price curve.