Orient Securities released a research report indicating that the immediate challenge in computing-electricity coordination centers on supply security and power quality. IT equipment in intelligent computing centers accounts for 45% to 60% of total electricity consumption, with AI accelerators representing the largest single source of power draw. While electricity costs make up roughly 10% of AIDC expenses, AI servers are extremely sensitive to voltage sag incidents, making data centers both high-energy and high-value loads where power reliability takes priority over cost optimization. In the near term, coordination will rely on grid power purchases with electricity serving computing needs, while the medium-to-long term could evolve from electricity serving computing to mutual support between the two.

Orient Securities Company Limited highlights that the short-term contradiction is balancing supply security with power quality. IT equipment in intelligent computing centers consumes 45% to 60% of total electricity, driven primarily by AI accelerators. Electricity costs can reach approximately 40% of IDC operational expenses, though only about 10% in AIDC settings. However, AI servers are acutely sensitive to voltage sags — a millisecond-level grid voltage drop can cause GPU training tasks to fail, losing intermediate data and incurring prohibitively high restart costs. Consequently, computing centers are not just energy-intensive but high-value loads, making power supply assurance more critical than reducing electricity bills.

Short-term opportunity costs of computing far exceed power price fluctuations, so computing will not yield to grid constraints. Idle GPUs on the training side incur fixed-cost losses far greater than brief price volatility, prioritizing model capability over electricity expense optimization. On the inference side, rigid L0 and L1 tasks are latency-sensitive, while L2 and L3 tasks can shift off-peak but currently represent limited volume, with pricing incentive mechanisms not yet systematically established. Meanwhile, the grid provides reliable overall backstop, and user-side tariffs do not fully reflect grid congestion costs. Therefore, near-term coordination will emphasize grid electricity purchases, focusing on reliable supply, voltage sag mitigation, and upgrades to backup power and distribution systems.

The power assurance framework faces pressure from the bottom up: cabinet-level distribution, system-level backup, and campus-level supply all require systematic upgrades. At the cabinet level, per-rack power has risen from 2 to 5 kW in traditional IDCs to 20 to 50 kW, with GB200 supernode racks reaching 120 kW, making 800V HVDC the future pathway. On the backup side, conventional UPS plus diesel generators are transitioning to BESS and grid-forming energy storage. At the campus level, large intelligent computing centers are moving from 10kV grid connections to 110kV or even 220kV direct links, raising grid operation and maintenance costs while green power direct connection emerges as a key alternative.

Temporal and spatial matching faces dual constraints from economic and physical factors, with short-term economic limits taking precedence. Computing load flexibility is classified into L0 through L3 tiers: traditional IDC loads are stable but lack adjustability; training can theoretically resume from checkpoints or downclock, though opportunity costs are high; online inference has weak adjustability, while offline inference and partial training can shift schedules. Renewable output fluctuations transmit into the spot market as high-frequency price signals, but current peak-valley spreads and pricing mechanisms remain insufficient to drive large-scale computing concessions. In the short term, computing-electricity matching is more about electricity ensuring computing rather than computing deferring to electricity.

Looking further ahead, once computing infrastructure costs decline, computing will become progressively price-sensitive, with "renewables — grid-forming storage — computing" emerging as a promising integration path. As GPU depreciation pressure eases, market-based peak-valley spreads widen, and costs for grid-forming storage and green power direct connection fall, computing flexibility will gradually unlock. The roadmap starts with advanced enterprises piloting computing-electricity price linkage and dynamic pricing, then advances to green power direct connection integrated with grid-forming storage, ultimately moving from electricity serving computing to mutual computing-electricity support.