CITIC Securities: Computing power breaks through power boundaries, SST starts a new cycle of silicon entering and withdrawing copper

The Zhitong Finance App learned that CITIC Securities released a research report saying that the increase in AI chip performance and density has led to a jump in intelligent computing center (AIDC) power consumption. Traditional uninterruptible power supply (UPS) architectures have hit bottlenecks in various aspects such as energy consumption and floor space, and AIDC power supply and distribution are accelerating the evolution towards high voltage DC. In the short term, high voltage DC transmission (HVDC) systems and the Panamanian power supply are expected as transition solutions. In the long run, 800VDC+ solid state transformers (SST) may become the ultimate solution for high-density computing power supply due to the advantages of medium voltage direct entry, high frequency isolation, and “silicon to copper retreat”. It is recommended to focus on SST components and core components such as SiC, high-frequency magnets, capacitors, and solid state circuit breakers.

CITIC Securities's main views are as follows:

Computing power expansion is shifting from “chip constraints” to “power constraints”, and high-density cabinets force power supply and distribution to be placed in front of Capex

The AI computing power upgrade is also driving up single-chip power consumption, cabinet power consumption density, and load fluctuations. Traditional data center cabinets of about 10 kW are leaping towards the MW level of Rubin Ultra/Kyber. The power supply equipment outside the cabinet has been upgraded from back-office support to the core infrastructure that determines GPU deployment density, launch speed, and effective computing power output, and is expected to be one of the most definitive aspects of value increase in this round of AIDC capital expenditure.

Nvidia's 800VDC direction has been established, and the power supply architecture is evolving rapidly along “high voltage, DC, modularization, and minimization”

Traditional UPSs require multi-stage transformations such as medium voltage step-down, AC/DC rectification, DC/AC inverter, and rack PSU rerectification. The series loss from the power grid to the cabinet is greater, and low voltage and high current simultaneously drive up copper consumption and gray area occupancy. By increasing the transmission voltage and reducing the same power current, 800VDC can significantly improve copper consumption, efficiency, and mitigate land occupation conflicts in gray areas. With the rise in cabinet power density in the Rubin era, upgrading the power supply architecture has become a necessary condition for the delivery of computing power.

Power Rack/Sidecar outside the cabinet was the first to be cashed out, and 800VDC+SST is the direction of long-term convergence

In the short term, existing data centers have improved solutions for UPS sidecar and HVDC sidecar/independent cabinet power supplies, which free up cabinet space and reduce transmission losses while retaining the existing AC power distribution system; some newly built data centers may choose a transition plan between HVDC systems and Panamanian power. SST can directly convert medium voltage AC to 800V DC, integrate voltage transformation, rectification, electrical isolation, power quality control and bidirectional energy management to reduce low voltage distribution levels and fault nodes. It is expected to become the ultimate solution for future AIDC cabinet power architectures.

SST uses “silicon to copper retreat” to restructure electric energy routes, energy efficiency, reduced ash area, and full life cycle economy form the underlying driving force of industrialization

According to Nvidia's “White Paper on 800V DC Power Supply Technology for Data Centers”, SST relies on power semiconductors and high-frequency magnetic components to raise the operating frequency from 50/60Hz to the kHz level, greatly reducing the transformer volume. The full link efficiency can be increased by about 3% compared to traditional UPS, and the system footprint can be reduced to less than 50% of the traditional solution. Saving electricity and land, shortening the deployment cycle, and reducing SiC costs are expected to continue to improve the economy throughout the life cycle. The inflection point of industrialization will shift from “whether it can be achieved” to “when to pass customer verification and form large-scale orders”.

Power semiconductors and high-frequency magnetic components are the core functional components of SST

According to Nvidia's “White Paper on 800V DC Power Supply Technology in Data Centers”, power semiconductors and high frequency/medium frequency transformers account for about 32% and 16% of SST costs, respectively, and are the most valuable incremental links; among them, SiC determines voltage resistance, frequency and conversion loss, high frequency magnetic components determine isolation capacity and power density, and film capacitors, supercapacitors, and solid state circuit breakers are responsible for bus regulation, instantaneous power support, and DC fault protection, respectively. Device cost reduction and reliability verification will jointly determine the inflection point of the SST penetration rate.

The entire industry chain is in the product launch+customer verification stage. 2027 is expected to be the first year of large-scale deployment

According to the official websites and announcements of various companies, Sifang Co., Ltd. has launched 10kVAC/800VDC and 2.4MW digital smart SST 1.0, which is currently in the stages of market promotion, scenario implementation and overseas certification promotion; China Western Power's 10kV/2.42MVA products have been used in the Gui'an “East Digital Western Computing” data center; Jinpan Technology has completed the 10kV/2.4MW prototype and continues to iterate, and the mass production time has not yet been announced. At this stage, the focus should be on tracking customer entry, certification results, demonstration run time, official orders, and revenue confirmation, and the release of prototypes should not be directly equated with large-scale scaling.

The market space depends on the five variables of hashrate chip shipment, unit power consumption, architecture penetration rate, unit value, and redundancy. SST has high elasticity from 0 to 1

This report is calculated using a scenario method. Under the optimistic/neutral/pessimistic scenario, the SST penetration rate for 2030E is assumed to be 40%/36%/32%, respectively. At the same time, it is assumed that the average SST price will drop to 3.0 yuan/W after large-scale cost reduction. The bank estimates that the 2030E SST market is expected to exceed 460 billion yuan under optimistic conditions and reach 416.6 billion yuan under neutral conditions.

Risk factors:

1) AIDC construction falls short of expectations; 2) technological progress risk; 3) tariff risk; 4) AI related capital expenditure falls short of expectations; 5) competition increases risk; 6) supply chain risk; 7) geopolitical and trade barrier risk.