NVIDIA is driving a major transformation of AI data centers, and Chinese manufacturers do not want to be sidelined.

Residents of Pennsylvania, USA, may have a lot of complaints when the AI boom hits. Because over the past 7 years, their average electricity price has risen by nearly 50%.

AI data centers are known as "power hogs", because a data center with a peak demand of 1 gigawatt consumes as much electricity as the annual electricity consumption of about 700,000 households, roughly equivalent to a small city. A 1GW data center consumes as much power as a medium-sized country.

But the trend is unstoppable. Sam Altman, President of OpenAI, once stated that if AI computing power increases 10,000 times in the next 20 years, its demand for energy will also increase 10,000 times.

AI companies are starting construction of data centers with higher power and better performance across the globe. In May 2025, NVIDIA announced the launch of the high-density computing power platform Kyber rack architecture and the 800V power supply architecture. This is a solution designed for future gigawatt-level AI factories, striving to maximize GPU density.

This is a new type of data center in the era of AI boom. NVIDIA defines it as a "fundamental architectural transformation". The core of the transformation can be summarized into two points: Traditional data center rooms usually use 480V alternating current into the cabinet, and step down and rectify it into 54V/48V direct current in the cabinet to supply power to servers. NVIDIA adopts 800V high-voltage direct current (DC) directly into the cabinet, completely replacing the traditional 54V DC bus.

According to NVIDIA's official technical documents, after switching from 415V AC to 800V DC, conductors of the same size can transmit 85% more power, and the demand for copper cables is directly reduced by 45%; at the same time, by eliminating multiple unnecessary AC-DC conversion links in the traditional link, the overall energy efficiency of the system can be significantly improved, which is also an important solution to support the power consumption of a single cabinet to move from 100kW to the megawatt level.

Countless suppliers have been involved in this huge transformation of AI infrastructure, actively or passively. On NVIDIA's partner list, in addition to established European and American giants such as Eaton and Infineon, there are also local enterprises such as Innoscience and Megmeet.

It is undeniable that although a small number of enterprises have entered NVIDIA's supply chain, the technology of many domestic companies still lags behind European and American enterprises in the short term. A CTO of an enterprise with many years of deep experience in the power electronics industry told Hard Krypton that he judges this gap will take 3 to 5 years of catching up.

But on the other hand, the construction of domestic intelligent computing centers and the follow-up of domestic computing chips have still released local market demand. The founder of an enterprise that started making megawatt-level SST earlier, interviewed by Hard Krypton, said that this technology did not attract much attention in the industrial field before, and initially only one or two domestic enterprises engaged in microgrid and fast charging paid attention to the megawatt-level SST solution. However, since the second half of 2025, with the development of intelligent computing centers, the industry has begun to expand, and the company has contacted a total of two to three hundred potential customers, "all of which are domestic enterprises".

In the secondary market, the semiconductor sector continues to rise, and the market performance of enterprises related to concepts such as SST and GaN is also very hot. In the first half of 2026, 20 energy storage-related enterprises have launched 27 new SST products, and institutions such as China Securities have also clearly favored the industrialization trend of SST.

If you want to occupy a place in the future power supply system of data centers, according to the deduction of the cycle from R&D to supply chain establishment, there is not enough time left for enterprises. "Although it is the early stage of industry development, it is already a key node that we must enter the market at," an investor who has long focused on AI Infra told Hard Krypton.

Three Power Conversion Reconstructions Under the New Architecture

To build a large-scale data center adapted to megawatt (MW)-level computing power, the power supply needs to undergo full-chain reshaping from the power generation end, transmission end, power storage end to power conversion end, among which the power conversion end has the most far-reaching impact. The reason why the industry focuses on solid-state transformer (SST), silicon carbide and gallium nitride in the huge power distribution chain is that they represent the core energy conversion points in this 800V DC revolution.

Specifically, from the high-voltage power of the grid to the final injection into the GPU chip, electricity needs to go through three core power conversions inside the data center:

The primary power supply means that thousands of volts of municipal electricity is converted into 800V high-voltage direct current through a transformer, which is the starting link of data center power distribution.

The solid-state transformer plays a key role here. It directly replaces the traditional power frequency transformer that weighs several tons and occupies several rooms at the gate of the data center, converting 10kV/35kV alternating current into an 800V DC bus, with a volume reduced by more than 50% compared with the traditional transformer. The physical characteristics of silicon carbide, such as high voltage resistance, high temperature resistance and high thermal conductivity, provide key support for SST not to break down under tens of thousands of volts of high voltage and megawatt-level high power.

The secondary power supply occurs at the cabinet. 800V DC enters the computer room and is converted to 54V/12V near the cabinet, existing in the form of cabinet power supply. Gallium nitride has fast switching speed, low high-frequency loss, and can reduce the size of magnetic components.

The tertiary power supply refers to that the 54V/12V power is sent into the server motherboard and further converted to an ultra-low voltage of about 1V to directly supply power to the GPU/CPU core.

If traditional low-frequency silicon-based devices are still used, the huge volume and heat generated by 800V high voltage will overwhelm the computer room. From the perspective of technological evolution, SST is not a brand new physical principle, and it was previously used in new energy vehicle supercharging and rail transit scientific research. However, introducing it into data center power supply on a large scale is a new attempt that has just moved towards commercial implementation.

This shift in application scenarios has also rewritten the competitive landscape of the semiconductor industry. The third-generation semiconductors represented by silicon carbide and gallium nitride have higher voltage resistance, high temperature resistance and high-frequency working capabilities than traditional semiconductor materials represented by silicon, and are especially suitable for applications such as new energy vehicles, photovoltaic energy storage, and 5G communications. They are important materials for the new generation of power semiconductors (power semiconductors refer to semiconductor devices used for power conversion, transmission and control, such as IGBTs and power diodes), but they have not been widely used for many years due to cost and technology constraints.

Chen Han, an insider in the semiconductor industry, reviewed the changes in the industry over the past years to Hard Krypton: In 2021, the industry experienced a chip shortage and price surge, giving domestic power semiconductors the opportunity to try and select models to enter the market; from 2022 to 2023, due to the de-stocking caused by the excessive release of production capacity, low- and mid-end devices fell into a fierce price war, and the profitability of manufacturers was worrying; it was not until 2024 that the inventory was cleared, and the industry began to make breakthroughs in high-quality products such as automotive-grade products.

With the advancement of the new generation of power centers, power semiconductors can finally jump out of the past vicious competition.

"Before the end of 2025, people generally felt that third-generation semiconductors were difficult to develop, and the new energy vehicle segment had been rolled into a dead end with negative gross margins," investor Yuan Yue told Hard Krypton. For a long time, investors have abandoned this track, but in the second half of 2025, AI computing power surged, power semiconductors shifted from new energy vehicles and traditional energy to the field of AI data center power supply, and power semiconductors returned to the center of the stage again.

"Essentially, all investors looking at this track are investing in AI Infra," Yuan Yue concluded. "Everyone has been looking for opportunities along the computing power industrial chain — the first wave invested in GPUs, the second wave invested in storage and CPUs, and then invested in switches, liquid cooling, and PCBs, and now it is the turn of power supplies and power infrastructure. In this exponentially growing industry, the physical upgrading of the power supply system has just kicked off, and the outbreak of the local industrial chain is only a matter of time."

 The Trust Barrier of the "High-Voltage Heart"

Although the demand is booming, from the current status of the industrial chain, the technologies of SST and third-generation semiconductors have not yet reached maturity. The technical level of domestic companies still lags behind overseas giants.

At present, domestic enterprises engaged in SST R&D are mainly cross-border from traditional transformer manufacturers and original new energy vehicle device manufacturers. However, many interviewed investors emphasized that SST has a high technical threshold, and the experience of automotive production lines cannot be fully reused.

Taking the key silicon carbide device as an example, the withstand voltage upper limit of a single silicon carbide transistor is limited, and it must be used in series and parallel under high voltage. However, the yield and consistency of a single domestic device are still insufficient at present, which directly affects the reliability of the overall system.

Benefiting from the vigorous construction of silicon carbide production lines for new energy vehicles in the past few years, the domestic silicon carbide industrial chain has been relatively mature. But Chen Han told Hard Krypton that although there are many participants, very few of them can achieve profitability — on the one hand, the penetration rate is still in the rising period, on the other hand, domestic products are concentrated in the low- and mid-end market with severe involution, and the high-end market is still dominated by overseas giants.

At the same time, new technologies also face an extremely high "trust threshold". Li Cheng, an investor who has focused on semiconductors for many years, metaphorically said, "Replacing SST is like performing a heart transplant. No data center dares to try it easily. Once SST fails, the expensive GPUs of the entire row or even the entire computer room will be paralyzed."

This also explains why many data centers are still using traditional coil-wound transformers for "patching" at present. Although they are bulky, their reliability has been verified for decades. Data shows that from January to April 2026, the total export value of traditional transformers in China has reached 21.77 billion yuan, a year-on-year increase of 27%, and the export value has repeatedly hit new highs.

"The trend is irreversible. When the power consumption of a single chip and a single cabinet continues to rise and the traditional transformer cannot fit in physically, people have to replace it," Li Cheng added. "It's just that the power of a single cabinet in most domestic intelligent computing centers has not yet risen to the critical point where replacement is inevitable."

In addition to semiconductors, in Li Cheng's view, server power supplies (PSUs) and power racks as the "last meter" of power implementation may still have some opportunities.

According to statistics from Valuates Reports, the global AI server power supply market size will grow from US$2.846 billion in 2024 to US$60.81 billion in 2031, with a compound annual growth rate (CAGR) as high as 45%. The order-of-magnitude leap in power density means that PSUs are no longer traditional low-gross-margin switching power supplies, but core systems integrating high-frequency circuits, liquid cooling heat dissipation and microcontrollers, and the unit price of hardware and profitability will also increase significantly.

From the perspective of the competitive landscape, the high-end power supply market has long been dominated by established European and American giants. "But the advantage of China's power supply industry does not lie in having mastered the most cutting-edge technology of 800V power supplies, but in the complete power electronics supply chain, strong manufacturing and engineering capabilities, and the high-voltage application experience accumulated over years of development of the new energy industry," Li Cheng judged.

The Unreachable NVIDIA Supply Chain and the Future Market

Among the 800V partner lists disclosed by NVIDIA, there are very few mainland enterprises. Enterprises such as Innoscience and Megmeet mainly cut in through cooperation in gallium nitride devices and power modules.

According to investors, NVIDIA's supply chain usually needs to be locked two years in advance: first, the enterprise needs to have a demonstrable hardware Demo to squeeze into the candidate list; after the Demo is verified, NVIDIA will also launch a strict factory audit process to comprehensively assess its manufacturing process, stability of upstream suppliers, monthly production capacity and wafer yield.

Even though the review process is cumbersome, hardware manufacturers around the world are still flocking to it. The core reason lies in the extremely high premium income — "The gross margin reserved by NVIDIA's supply chain for suppliers is generally 60%~70%, while the gross margin of the ordinary domestic power supply supply chain is often only 20%~30%."

In addition to the factory audit threshold, what is more solid is the barrier brought by geopolitics and security trust. For the high-power infrastructure of data centers that is most closely related to people's livelihood and security, North American data centers still find it difficult to accept domestic companies.

This also leads to the fact that if domestic enterprises want to go global, they often need to transit through intermediaries such as Singapore, or step back to provide OEM of components, and it is extremely difficult to penetrate directly with the whole machine brand.

So, if they cannot enter NVIDIA's supply chain, where are the opportunities for domestic SST and third-generation semiconductor enterprises?

Many investors said that although the vast majority of domestic start-ups cannot reach overseas top-level computing power chains in the short term, the evolution of domestic GPU chips such as Huawei Ascend, Cambricon and Moore Threads is also very rapid. As the power consumption of domestic chips continues to rise, local intelligent computing centers will also face the problem of power supply. Domestic power supply and semiconductor enterprises can cooperate with them, pre-bind local leading chip and server manufacturers, and jointly pre-research the next-generation 800V power supply solution.

On the other hand, the domestic substrate industrial chain has brought great cost and scale advantages. Although there is a gap with Europe and the United States in high-end device design, China has the world's largest and most mature silicon carbide substrate and epitaxy industrial chain, such as Tianyue Advanced and Sanan Optoelectronics. If large-scale implementation is to be realized in the future, cost reduction of silicon carbide and gallium nitride devices is the key.

In the second half of this year, as the first batch of application scenarios of NVIDIA's 800V high-voltage DC power supply architecture, the Rubin series will be officially shipped in batches and launched to provide services. The standards it has explored and the actual application situation will also guide the direction of the entire industry. For many domestic semiconductor and power supply enterprises, the east wind they are waiting for is actually not far away.