Author: Xia Mo, Editor: Shui Jing
"The end of AI is computing power, and the end of computing power is electricity." This industry consensus is restructuring the global power ecosystem.
Goldman Sachs' latest SST research report shows that electricity is becoming a harder bottleneck than computing power. The explosive power consumption of AI systems has tech giants deep in thought, as increasingly demanding high-performance hardware, training and inference processes, data center operations, and cooling systems all require massive power support. The traditional power chain in data centers is turning into an energy disaster—for every 10 kWh of electricity entering a data center, 1 to 1.5 kWh is wasted in the conversion process.
It is precisely this shift that has propelled a technical prototype born over 50 years ago to become a strategic high ground that global computing giants are scrambling to occupy.
Today, NVIDIA has included it in its official technology roadmap, Google and Microsoft have incorporated it into the power distribution schemes for next-generation data centers, and top global investment banks have identified it as a market worth over CNY 100 billion. It is the SST (Solid State Transformer). Notably, a few years ago, outside of computing scenarios, SST was still mainly confined to demonstration projects on the grid side and was regarded by most as a long-term technical direction.
Driven by the global explosion of AI, a new hundred-billion-level market—SST—is on the eve of a boom. This article aims to systematically review the SST technology trends, commercialization process, and the layout of core players, and to discuss with readers the topic of "future growth space and practical challenges of the SST industry." It is intended to spark discussion and is for reference only. Readers are welcome to supplement, correct, or even criticize it.
Why SST Has Become the Most Feasible Solution for AIDC
A set of data is sufficient to demonstrate the growth potential of the AIDC industry ecosystem.
According to forecasts by the CAICT (China Academy of Information and Communications Technology), the global total computing power scale is expected to exceed 16 ZFlops by 2030, with a compound annual growth rate of over 40% from 2022 to 2030. The proportion of intelligent computing power, mainly driven by AI training and inference workloads, will reach over 30%.
The sustained high growth of global computing power indicates that computing infrastructure is becoming a strategic real estate allocation as important as office space and production factories. JLL (Jones Lang LaSalle) estimates that by 2030, the total global data center capacity will increase from 103 GW to 200 GW, nearly doubling.
Market research firm Dell'Oro Group predicts that global data center capital expenditure will approach the milestone of USD 1 trillion in 2026 and reach USD 1.7 trillion by 2030. Goldman Sachs has a more long-term estimate: between 2026 and 2031, the cumulative capital expenditure around computing power, data centers, and electricity could reach approximately USD 7.6 trillion.
The industry vision is already very clear, but the industry still faces severe challenges. The power supply chain in traditional data centers is a lengthy conversion chain. Medium-voltage AC is first stepped down to low-voltage AC via power-frequency transformers, then undergoes multiple "AC-DC-AC" conversions through UPS systems, and is finally converted to DC by server power supply modules to supply the chips. Each additional conversion adds another layer of loss.
Goldman Sachs' latest SST research report shows that as the power consumption of NVIDIA's DGX NVL8 surges to 3.7 times that of the A100, and the power density of a single rack rushes from 30 kW to 1,000 kW by 2030, the traditional data center power chain of "grid-transformer-UPS-PDU-server" is turning into an energy disaster—for every 10 kWh of electricity entering a data center, 1 to 1.5 kWh is wasted in the conversion process.
What is changing is not just the technology itself, but the entire computing industry's understanding and cognition of electricity. Starting from first principles, SST is highly likely to be the most feasible solution for the sustainable development of the AIDC (Artificial Intelligence Data Center) industry.
A set of data is sufficient to summarize the full significance of this transformation. First is the leap in efficiency. According to industry data, the five-stage conversion efficiency of traditional solutions is only 75%-80%, while the single-stage conversion efficiency of SST can reach 98%-99.63%. Data from STMicroelectronics shows that its GaN (Gallium Nitride) technology prototype can operate at 800V and 1MHz, achieving 98% efficiency and a power density of 2,600W per cubic inch. Sungrow's 800V high-voltage DC power supply system launched for data centers adopts SiC MOSFET and soft-switching technology, with a system efficiency of up to 96.5%-97.2%, and can achieve over 98% efficiency when supporting "offline mode" (direct mains supply).
Taking a 2.5 MW system as an example, switching from UPS to the SST solution can save nearly 600,000 kWh of electricity per year. Estimated at an industrial electricity price of CNY 0.5/kWh, only the electricity bill can save about CNY 300,000. For a large data center of tens of megawatts, this figure will be amplified to the level of millions or even tens of millions of CNY.
Second is the leap in power density. Mass-production-level SST based on SiC can reduce volume and weight to 1/4 to 1/6 of traditional solutions, increasing power density by more than 4 times. A traditional 10kV/1MW transformer requires 50 square meters of space, while SST requires only 1-5 square meters, a 90% reduction. The traditional solution for a 1 MW cabinet requires 200 kg of copper busbars, whereas the 800V DC architecture can reduce copper usage by 45%, decrease power distribution space occupation by 80%, and increase available cabinet space by 30%.
Third is the leap in intelligence and response speed. The response speed of traditional transformers is at the second level, while SST can achieve millisecond-level response, accurately matching the severe fluctuations of AI computing workloads. Meanwhile, SST supports bidirectional power transmission, harmonic suppression, and active voltage regulation, and can seamlessly adapt to renewable energy and energy storage systems, providing the possibility for data centers to realize a "flexible grid."
In summary, for a 2.5 MW data center, switching from traditional UPS to the SST solution can save over CNY 500,000 in electricity bills annually. The released power distribution space can accommodate the deployment of over a dozen additional cabinets, and copper usage is reduced by tons. For any AIDC project, this is not just an incremental improvement, but a systematic reconstruction starting from the underlying architecture.
In October 2025, NVIDIA released the white paper "800V DC Architecture for Next-Generation AI Infrastructure," listing medium-voltage rectifiers and solid-state transformers as the ultimate solutions for future out-of-rack power supplies. According to its roadmap, the single-rack power of Rubin Ultra will reach 600 kW in 2027, and the Feynman platform will exceed 1 MW in 2028 (Note: Reports in July 2026 indicated that the Kyber rack has been delayed to 2028, and there is uncertainty in the actual promotion pace). When the cabinet power density surpasses 600 kW and evolves towards the MW level, the traditional 54V/415V power supply architecture approaches its limits in three dimensions: expansion elasticity, copper loss control, and cable carrying capacity. Relying on its ability to step down medium voltage directly to DC and a power density several times that of traditional solutions, SST has become the only feasible technical choice in this range.
The deeper change lies in the fact that SST naturally possesses multi-port access capabilities. DC power sources such as photovoltaics and energy storage can be directly connected to its DC bus without the need for independent AC/DC conversion stages. This means that AIDC is no longer a mere power-consuming terminal, but can become an integrated node of source, grid, load, and storage. This is precisely the most scarce interface capability for resolving the contradiction between power supply and demand in the AI era.
Along with the acceleration of the commercialization process, SST is highly likely to grow into a new hundred-billion-level market. According to estimates by Huaan Securities, the global SST market size will rapidly climb from CNY 2.82 billion in 2026 to CNY 101.04 billion in 2030. Industrial Securities' forecast is more optimistic, believing that the global out-of-rack power supply market space will reach CNY 271 billion in 2030, of which the SST penetration rate will reach 40%, corresponding to a market size of about CNY 176 billion.
Deconstructing the Five Barriers and Three Challenges of SST Commercialization
SST is not a brand-new concept that emerged out of nowhere; its technical prototype can be traced back to 1970. However, for the following half-century, it has remained at the stage of laboratories and demonstration projects, failing to truly enter the commercial market.
It is generally believed in the industry that the large-scale deployment of SST must overcome at least five technical hurdles. The first hurdle is medium-voltage insulation. SST needs to achieve high-frequency conversion of 10 to 100 kHz at medium-voltage levels of 10 kV to 35 kV. The partial discharge index needs to be controlled below 10 pC, and some high-end scenarios even require it to be below 5 pC. The dv/dt caused by high-frequency PWM pulses exceeds 10 kV/μs, further exacerbating the burden on the insulation system. How to meet the requirements of low leakage inductance and high insulation simultaneously within limited space is a tough nut to crack in product design.
The second bottleneck lies in high-frequency magnetic components. The eddy current loss of traditional silicon steel sheets rises sharply at high frequencies, necessitating the switch to new magnetic materials such as nanocrystalline or amorphous alloys. The loss of nanocrystalline cores under high-frequency operating conditions can be reduced by 60% to 80% compared to silicon steel, but the global supply is highly concentrated, forming a duopoly by Germany's Vacuumschmelze and Japan's Proterial, making the supply chain not very abundant.
There is also a detail that sufficiently illustrates the engineering difficulty of SST. When a leading manufacturer was conducting insulation tests on a 10 kV-level prototype, it broke down three consecutive times when the voltage rose to about 80% of the rated value. After two weeks of investigation, it was discovered that the culprit was not the power devices, but a slight displacement of an insulation gasket between the windings of the high-frequency transformer under high voltage. A misalignment of merely 0.3 millimeters was enough to trigger partial discharge. The final solution was not to use more expensive materials, but to completely redesign the insulation structure, iterating through six versions. Similar engineering problems repeatedly occur in every technical module of SST, including magnetic components, protection, and control.
The third challenge comes from DC protection. The fault current in a DC system has no natural zero-crossing point, making it impossible to extinguish the arc by utilizing current zero-crossing as in AC. The complexity of the protection logic increases exponentially. The protection chain requires five layers of coordination, and currently, only a few manufacturers such as Eaton and ABB have mature solutions.
The fourth hurdle is mass production consistency. It is not difficult to build a few prototypes with over 98% efficiency in the laboratory, but during mass delivery, each unit must stably pass a series of stringent tests such as withstand voltage, partial discharge, and temperature rise, with no unit allowed to fail. Data centers typically require the Mean Time Between Failures (MTBF) of equipment to be over 50,000 hours.
The fifth aspect lies in system reliability. A single SST usually contains 20 to 40 power modules, each of which requires independent control. The switching frequency is between 10 and 100 kHz, and its topology and algorithm complexity are 5 to 10 times that of traditional UPS. At the system level, the power factor must be brought above 0.99, and the Total Harmonic Distortion (THD) must be controlled below 5%.
The above five tests must be passed simultaneously, and none can be omitted. The significance of a single-point breakthrough is limited. AIDC's requirement for power supply availability is over 99.999%. Any unplanned power outage could interrupt ongoing large model training and cause data corruption, with losses counted in tens of millions of US dollars. Therefore, system-level engineering capability is the true competitive barrier.
Beyond technology, there are also three practical thresholds to face.
First is economics. Saving copper, space, and electricity bills—these advantages belong to long-term operational benefits, while the procurement price of SST equipment is currently several times that of traditional solutions. The costs are mainly concentrated in power electronic devices, high-frequency magnetic components, and complex drive control and protection circuits. However, with the advancement of the localization of SiC devices and the initiation of scaled production, costs are dropping rapidly; it just takes time to bridge this current price gap.
The second threshold is the lack of standards. There are currently no unified design specifications, interface standards, or testing methods for the application of SST in data centers. Products from different manufacturers are difficult to interoperate, and the standardization process has just begun.
The third threshold is the reliability verification cycle. The power system has extremely little fault tolerance for continuous power supply, and SST must undergo long-term grid-connected operation verification to gain customer trust. AIDC customers have zero tolerance for power outages; without a sufficiently long safe operation record, it is difficult to make any procurement decisions.
Technical barriers and practical challenges are essentially asking the same question: Has the new technology been proven reliable enough in real-world scenarios? The answer can only be accumulated through time and data.
In the Race of Industrial Competition, Who is Leading?
Undeniably, in the current global commercialization process of SST, overseas enterprises are running a bit faster.
Among them, the enterprises with the fastest industrialization progress have moved from the prototype stage to real-world scenario verification. Delta Electronics is currently the one with the clearest pace. In November 2025, Delta Electronics, together with Meituan and Chindata, globally launched an intelligent DC power supply system solution based on SST, becoming the first publicly disclosed commercial landing case. Its single power cabinet outputs up to 1 MW, occupies only 1 square meter, and achieves an efficiency of 98.5%. In March 2026, Delta Electronics further introduced a containerized SST DC mobile intelligent computing center, integrating CPUs, GPUs, and SST into an integrated delivery unit. Currently, Delta's SST has completed long-term grid-connected testing in the data centers of leading cloud vendors.
It is also understood that SST-related products from another giant, Eaton, have been operating stably in a large data center in North China for over a year, with system availability reaching 99.999% and dynamic response below 5 ms. Currently, multiple pilot projects by Eaton are underway, covering several leading cloud vendors at home and abroad. In June 2026, it also released the MV SST Solid State Transformer 2.0, adopting a 10 kV medium-voltage direct input and 800V DC one-step conversion architecture, with a full-chain efficiency of 98.5%, a power density of 278 kW/m², and a delivery cycle shortened by 50% compared to traditional solutions.
Of course, domestic enterprises have also shown a fierce catching-up momentum.
China XD is one of the domestic enterprises with the deepest layout in the SST field. In April 2026, its 10 kV/2 MW/750V SST was put into grid-connected operation in the Xi'an XD Smart Park. Two months later, XD Power Electronics won an order for four 13.8 kV AC/800V DC SSTs for an overseas data center, becoming a landmark event for domestic SST products entering the overseas market.
Another heavyweight player, Sungrow, officially released its fully self-developed solid-state transformer EnerNeo for AIDC scenarios on July 9. According to the introduction, for AI data center scenarios, EnerNeo is based on an 800V DC power supply architecture, adopting medium and low-voltage zoning design, intelligent module rotation rest, and integrated vacuum insulation technology. Combined with intelligent technologies such as digital twins, it achieves 99.999% system availability, 98.5% conversion efficiency, and a power density of 312 kW/m².
Its power coverage ranges from 1.5 MW to 4.5 MW. It can adapt to the renovation of old machine rooms, the construction of new ultra-large-scale intelligent computing clusters, and phased expansion needs through modular combinations. It also supports applications such as backup power, dynamic load regulation, and AC load compatibility. In terms of market expansion, Sungrow stated that its SST products have attracted the attention of overseas cloud vendors and are currently conducting in-depth docking with leading North American cloud vendors such as Microsoft and AWS, participating in the joint design and definition of power supply specifications for next-generation computing infrastructure.
In terms of capacity building, Sungrow has initiated the construction of a dedicated SST production line and plans to form a capacity of 1.2 GW by the end of 2026.
On July 31, according to public reports, Sungrow provided two 2.5 MW EnerNeo SSTs and supporting DC/DC conversion cabinets for a new domestic computing center, with the entire project planned to be put into operation by the end of September. From the external release of the product on July 9 to the confirmation of the landing project, there was an interval of only 22 days.
Meanwhile, players like Sifang Co., Ltd. and Jinpan Technology also cannot be underestimated. Among them, Sifang Co., Ltd. released the Digital and Intelligent SST 1.0 in April 2026, featuring 10 kV AC input, 800V DC output, 2.4 MW power, and 98.5% efficiency, and announced "mass production upon release." The company plans to release a 35 kV prototype in the third quarter of 2026 and a 13.8 kV prototype in the fourth quarter. Sifang Co., Ltd. has early engineering accumulation in the SST field and is one of the earliest domestic enterprises to achieve SST engineering landing.
In January 2026, Jinpan Technology released the Yuanshen ONE series SST, with a typical capacity of 2.4 MW, an efficiency of over 98%, and a footprint saving of about 60% compared to traditional solutions. It adapts to 10 kV, 13.8 kV, and 35 kV grid systems and supports multi-port access for wind, solar, and energy storage. This product adopts a full SiC design scheme, and the company has formed dozens of core patents around SST.
According to the analysis and judgment of Guosheng Securities: The stable operation of SST requires the accumulation of practical experience. The leap from 10 kV to 35 kV is difficult. Enterprises with rich practical experience in delivering SST system projects have a better chance of seizing the initiative in AI data center demand.
It is widely recognized that in the fields of UHV DC transmission and flexible DC distribution networks, China has the most engineering practices and grid-connected operation records globally. The core technologies required for SST, such as grid-forming control, medium-voltage rectification, and DC protection, have been pre-verified on the grid side by domestic enterprises, which is the natural advantage of Chinese enterprises.
In a research report, Guojin Securities deconstructs the competitive barriers of SST into three levels: the ability to participate in defining product specifications, medium-voltage system engineering capability, and high-frequency power electronics productization capability. The time scales corresponding to the three are different: in the short term, it depends on the speed of prototype verification; in the medium term, it depends on the accumulation of system-level reliability; and in the long term, it depends on modular mass production capability.
The 24Chao Industry Research Institute (TTIR) analyzes that the winning point in the current competition is not at the hardware level. What truly determines the market position of enterprises is a more upstream issue: who enters the architectural design system of AI giants first.
Just as Pulitzer Prize-winning journalist Gary Rivlin revealed a cruel reality in his book "The Silicon Valley AI Dream": the cost of training and inference for foundation models is astonishingly high; this is a game that only giants can afford to play.
The facts are indeed so. During the earnings call in August last year, Jensen Huang stated that the cost of building an AI data center with a capacity of 1 GW is approximately USD 50 billion. According to data from authoritative institutions, in 2025, global data center investment reached USD 770 billion, a figure that has already surpassed the investment in the entire photovoltaic industry, as well as the investment in upstream oil and gas exploration and development.
According to other public reports, in the first three quarters of 2025, the capital expenditure of the top four US cloud vendors reached USD 284.1 billion; domestic enterprises Alibaba and Tencent Holdings spent over CNY 150 billion, among which Alibaba announced in 2025 an investment of CNY 380 billion over the next three years for AI and cloud infrastructure construction, with a focus on computing centers...
In the AIDC ecosystem, whoever can build an "industrial community" with leading AI enterprises can seize the initiative in the new industrial competition, and even possess the ability to change the industrial landscape and direction.
When NVIDIA released the 800V white paper, it simultaneously disclosed its collaborative relationships with more than 20 partners. Delta, Eaton, and Vertiv are all on the list. This list draws a dividing line; being able to participate in the definition of the power supply interface for NVIDIA's next-generation cabinets means that competitors have been locked out of the standard before the parameters are even finalized. This may be the most hidden and hardest-to-reverse disadvantage in the SST competition.