Oriental Securities: Telecommunication collaboration focuses on securing supply and power quality in the short term and moving towards mutual assistance in computing and electricity in the medium to long term

The Zhitong Finance App learned that Oriental Securities released a research report saying that the short-term contradiction between power computing collaboration is “guarantee supply and power quality”. IT equipment in intelligent computing centers accounts for 45% to 60% of total electricity consumption. AI accelerators are the largest single power consumption, and electricity costs are about 10% of AIDC costs; however, AI servers are extremely sensitive to voltage drops. Computing power centers are both high energy consumption and high value loads. The power guarantee priority is higher than electricity bill optimization. Short-term electricity computing collaboration will focus on grid electricity purchases and electricity service calculation, while in the medium to long term, it is expected to move from “electricity service calculation” to “mutual assistance in electricity calculation.”

Orient Securities's main views are as follows:

The short-term contradiction between computing and telecommunication collaboration is “guarantee of supply and power quality”

IT equipment in intelligent computing centers accounts for 45% to 60% of total electricity consumption, and AI accelerators are the largest single power consumption; 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 drops. A voltage drop of tens of milliseconds on the power grid can cause the GPU training task to report an error and exit. Intermediate data loss and restart costs are extremely high. Therefore, the computing power center is not only a high-energy load, but also a high-value load, and the priority of power guarantee is higher than electricity cost optimization.

The opportunity cost of short-term computing power is much higher than fluctuations in electricity bills, and computing power will not make concessions to the power grid

The fixed cost loss of idle GPUs on the training side far exceeds short-term electricity bill fluctuations, and model capability takes precedence over electricity cost optimization; rigid tasks such as L0 and L1 on the inference side are sensitive to delays. Although L2 and L3 can be erroneous, the current share is limited, and the pricing guidance mechanism has not yet been systematically formed; at the same time, the overall power grid is reliable, and the user side electricity bill does not fully reflect the cost of grid congestion. Therefore, short-term electricity calculation collaboration will focus on grid power purchase and electricity service calculation, focusing on reliable power supply, voltage suspension control, and power backup and distribution upgrades.

The power guarantee system is under pressure from the bottom up: cabinet power distribution, system power backup, and park power supply are all facing systematic upgrades

On the cabinet side, the power of a single cabinet was raised from 2-5kW of traditional IDC to 20-50kW, and the power consumption of the entire GB200 supernode cabinet reached 120kW, and 800VHVDC became the future route; on the backup side, traditional UPS+ diesel engines are transitioning to BESS and grid-based energy storage; on the park side, large-scale intelligent computing centers are directly connected from 10kV to 110kV to 220kV. Power grid operation and maintenance guarantee costs are rising, and direct green power connections have become an important alternative.

Spatio-temporal matching is limited by both economic constraints and physical constraints. Short-term economic constraints take priority

Computing power load adjustability is graded according to L0-L3: traditional IDC load is stable but the adjustability is weak; training theory can be repeated at intervals and the frequency can be downgraded, but the opportunity cost is high; online reasoning is weak, and offline inference and partial training can be erroneous. Fluctuations in energy output are transmitted to high-frequency electricity prices through the spot market, but the current peak-and-valley price differences and pricing mechanisms are not enough to drive large-scale concessions in computing power. In the short term, electricity matching is more about “electricity insurance” rather than “giving up electricity.”

The medium and long term expect that after the cost of computing power facilities falls, computing power will gradually become sensitive to electricity prices. “New energy - grid energy storage - computing power” is the direction of potential integration

As depreciation pressure on GPUs eases, price differences between peaks and valleys of electricity marketability widen, and the direct connection costs of grid-based energy storage and green power decline, computing power adjustability will gradually be released. On the path, advanced enterprises should first pilot electricity calculation price linkage and dynamic pricing, then promote the integration of direct green power connection and grid-based energy storage, and eventually move from “electricity service calculation” to “mutual assistance in electricity calculation.”