Orient: In the short term, computing-power coordination focuses on supply guarantee and power quality; in the medium to long term, it will move toward mutual support between computing and power.

Orient released a research report stating that the short-term contradiction in computing-power coordination is "supply guarantee and power quality." IT equipment in intelligent computing centers accounts for 45%60% of total electricity consumption, with AI accelerators being the largest single power-consuming source. Electricity costs account for about 10% of AIDC costs; however, AI servers are extremely sensitive to voltage sags. Computing centers are not only high-energy-consuming loads but also high-value loads, and power supply assurance takes priority over electricity cost optimization. In the short term, computing-power coordination will mainly rely on grid electricity purchases and electricity serving computing; in the medium to long term, it is expected to move from "electricity serving computing" to "mutual support between computing and power." Orient's main views are as follows: The short-term contradiction in computing-power coordination is "supply guarantee and power quality." IT equipment in intelligent computing centers accounts for 45%60% of total electricity consumption, with AI accelerators being the largest single power-consuming source; electricity costs can account for about 40% of IDC business costs and about 10% of AIDC costs. However, AI servers are extremely sensitive to voltage sags. In actual tests, a voltage drop of tens of milliseconds on the grid can cause GPU training tasks to report errors and exit, resulting in loss of intermediate data and extremely high restart costs. Therefore, computing centers are not only high-energy-consuming loads but also high-value loads, and power supply assurance takes priority over electricity cost optimization. In the short term, the opportunity cost of computing power is far higher than electricity price fluctuations, and computing power will not yield to the grid. On the training side, the fixed-cost loss from idle GPUs far exceeds short-term electricity price fluctuations, and model capability takes priority over electricity cost optimization; on the inference side, rigid tasks such as L0 and L1 are latency-sensitive, while L2 and L3 can shift peaks but currently account for a limited share, and pricing guidance mechanisms have not yet been systematically formed; at the same time, the grid as a whole provides reliable backstop support, and user-side electricity bills do not fully reflect grid congestion costs. Therefore, in the short term, computing-power coordination will mainly rely on grid electricity purchases and electricity serving computing, with the focus on reliable power supply, voltage sag governance, and upgrades to backup power and distribution. The power assurance system is under pressure from the bottom up: cabinet power distribution, system backup power, and campus power supply all face systemic upgrades. On the cabinet side, single-cabinet power has risen from 25kW in traditional IDCs to 2050kW, and the GB200 supernode full cabinet power consumption reaches 120kW, with 800VHVDC becoming the future route; on the backup power side, traditional UPS plus diesel generators are transitioning toward BESS and grid-forming energy storage; on the campus side, large intelligent computing centers are moving from 10kV access to 110kV and even 220kV direct connection, grid operation and maintenance assurance costs are rising, and green power direct connection has become an important alternative. Spatiotemporal matching is subject to both economic and physical constraints, with economic constraints taking priority in the short term. Computing load adjustability is graded from L0 to L3: traditional IDC loads are stable but weakly adjustable; training can theoretically resume from checkpoints and reduce frequency, but the opportunity cost is high; online inference is weakly adjustable, while offline inference and some training can shift peaks. Fluctuations in renewable energy output are transmitted as high-frequency electricity prices through the spot market, but current peak-valley price spreads and pricing mechanisms are still insufficient to drive large-scale concessions by computing power. In the short term, computing-power matching is more about "power guaranteeing computing" rather than "computing yielding to power." In the medium to long term, after the cost of computing facilities declines, computing power will gradually become sensitive to electricity prices, and "renewable energygrid-forming energy storagecomputing power" is a potential integration direction. As GPU depreciation pressure eases, market-based peak-valley price spreads expand, and the costs of grid-forming energy storage and green power direct connection decline, the adjustability of computing power will gradually be released. On the path forward, advanced enterprises should first pilot computing-power price linkage and dynamic pricing, then promote the integration of green power direct connection and grid-forming energy storage, and ultimately move from "electricity serving computing" to "mutual support between computing and power."