Orient Securities has released a research report highlighting that the immediate challenge in coordinating computing power and electricity supply centers on "supply security and power quality." Data centers are not only high-energy-consumption loads but also high-value loads, making the priority for power reliability greater than that of optimizing electricity costs.

In the near term, this coordination will primarily rely on purchasing electricity from the grid, where power serves computing needs, focusing on reliable supply, voltage sag mitigation, backup power, and distribution upgrades. Over the medium to long term, the integration of "renewable energy—grid-forming energy storage—computing power" emerges as a promising direction for synergy.

The immediate conflict in computing-power-power coordination is "supply security and power quality." IT equipment within intelligent computing centers accounts for 45%–60% of total electricity consumption, with AI accelerators being the largest single source of power usage. Electricity costs can represent approximately 40% of an IDC's operational expenses but only about 10% of an AIDC's costs. However, AI servers are extremely sensitive to voltage sags; a grid voltage dip lasting just tens of milliseconds can cause GPU training tasks to fail, resulting in data loss and prohibitively high restart costs. Thus, data centers are both energy-intensive and high-value loads, where ensuring power supply takes precedence over minimizing electricity expenses.

In the short term, the opportunity cost of computing power far exceeds fluctuations in electricity prices, and computing power will not yield to the grid. The fixed costs of idle training-side GPUs far surpass short-term electricity price variations, making model capability a higher priority than power cost optimization. For inference, rigid tasks at the L0 and L1 levels are latency-sensitive, while L2 and L3 tasks can shift to off-peak periods but currently hold limited share, and pricing guidance mechanisms are not yet systematically established. Meanwhile, the grid's overall reliability acts as a safety net, and user-side electricity bills do not fully reflect grid congestion costs. Consequently, near-term coordination will focus on grid-purchased power and electricity serving computing, with emphasis on reliable supply, voltage sag mitigation, backup power, and distribution upgrades.

The power assurance system faces mounting pressure from the bottom up, requiring systemic upgrades in rack power distribution, backup power, and campus-level supply. At the rack level, single-rack power is rising from 2–5kW in traditional IDCs to 20–50kW, with GB200 supernode full-rack consumption reaching 120kW, making 800V HVDC the future standard. For backup power, traditional UPS and diesel generators are transitioning toward BESS and grid-forming energy storage. At the campus level, large-scale intelligent computing centers are moving from 10kV connections to 110kV or even 220kV direct connections, increasing grid operation and maintenance costs while making green power direct supply a key alternative.

Temporal matching between power supply and computing demand is constrained by both economic and physical factors, with economic constraints taking priority in the short term. Computing load adjustability is tiered from L0 to L3: traditional IDCs have stable but weakly adjustable loads; training tasks theoretically support checkpointing and frequency reduction but carry high opportunity costs; online inference is weakly adjustable, while offline inference and some training can shift to off-peak times. Renewable energy output volatility transmits through spot markets as high-frequency electricity prices, yet current peak-valley price spreads and pricing mechanisms are insufficient to drive large-scale computing power concessions. In the near term, the matching is more about "power securing computing" rather than "computing yielding to power."

Over the medium to long term, as computing infrastructure costs decline, computing power will become increasingly sensitive to electricity prices, and "renewable energy—grid-forming energy storage—computing power" will become a viable integration path. As GPU depreciation pressure eases, peak-valley price spreads in the electricity market widen, and costs for grid-forming storage and green power direct supply fall, computing load adjustability will gradually be unlocked. The roadmap should begin with pilot programs at advanced enterprises for dynamic computing-power-electricity price linkage, then advance to integrating green power direct supply with grid-forming storage, ultimately evolving from "electricity serving computing" to "computing and power mutual support."