Author: Daewoo; Source: X, @BTCdayu

In 1962, an American physicist named Thomas Kuhn published a book called "The Structure of Scientific Revolutions", which put forward a very famous argument called "paradigm shift". This word later became a fixed phrase in English, and his works were also cited in a wide range of fields besides the history of science. This book was later called "one of the most influential historical and philosophical works of the twentieth century" by the Encyclopedia Britannica, with more than 140 copies sold worldwide. Thousands of copies, it is as popular as "Harry Potter" in this type of hard-core book.

Kuhn’s core observation is: Science does not accumulate slowly in a straight line. The old explanatory system first explained the world well, and then accumulated anomalies one by one. At a certain point, it could no longer hold up, and the entire system was overthrown at once and replaced with a brand new system that was fundamentally incompatible with the old one.

Kuhn was talking about science, but this rhythm appears more intensively in the industry: film cameras are replaced by digital, PC Internet is eaten up by mobile Internet, and fuel power is overtaken by electric drive. It's the same pattern every time: the old system actually works in its original application, but it can no longer hold up in the face of new demands, either physically, economically, or both.

The AI data center is currently undergoing another switch. The protagonist of today’s conversation is not computing power or the transition from copper to light, but electricity.

1. The old paradigm can no longer hold on

Data center power supplies have experienced a steady internal upgrade over the past decade—from 12 VDC to 48/54 VDC. This is a routine improvement. It can save money and the old structure can still be used.

By 2026, this architecture is becoming more and more restrictive.

The GB200 cabinet released in the first two years has a power of about 120 kilowatts, and 54 volts DCcan still hold up; the Rubin VR200 platform, which will be put into production at the end of 2026, has a single cabinet power of about 200 kilowatts and is beginning to struggle; the Rubin Ultra Kyber platform, which will be released in the second half of 2027, has a single cabinet power of 600 kilowatts, rising to 1 megawatt, 54 Fucan’t hold on any longer.

Because power equals voltage times current. For the same power, the higher the voltage, the smaller the current, and the loss decreases with the square of the current. A 1 MW cabinet would need to draw 18,500 amps if it continued to use 54 volts DC. NVIDIA disclosed: A 1 MW cabinet requires about 200 kilograms of copper for copper bars, but the problem is not that the copper cannot be used, but that there is not enough space. If the 1 MW cabinet continues to use 54 volts, the power supply will occupy 64 cabinet units in the cabinet, and there is no place for the GPU.

Nvidia mentioned this problem in May 2025 and issued a white paper in October. The solution given is 800 volts high voltage direct current (HVDC). At the boundary of the data center, the medium voltage AC is directly converted into 800 volt DC, and then sent to the cabinet using an 800 volt bus. The voltage in the cabinet is then stepped down to 50 volts, 12 volts, and finally to the GPU core voltage of 0.8 volts. This is a bit like first principles. After doing this, the conversion efficiency is increased from 83-87% to 92-96%, the copper consumption is reduced by about 45%, the total end-to-end ownership cost is reduced by about 30%, and the maintenance cost is reduced by 70%.

If this becomes an industry standard, then thedesign standards for almost all power devices, transformers, busbars, protection, and backup systems on the entire power link will have to be rewritten.

The essence of studying AI from an investment perspective is to study capital expenditures and the flow of expenditures.

The International Energy Agency reported in April 2026: Global data center electricity consumption in 2030 will be approximately 950 terawatt hours, almost twice that in 2025; among which, AI data center electricity consumption will increase to 3 times; McKinsey estimates for the same period: By 2030, Cumulative capital expenditures for AI data centers will be US$5.2 trillion.

Where the two curves intersect is the power infrastructure. So,how much of that is electricity, and how has it changed?

2. How big is the switch and in what directions

McKinsey breaks $5.2 trillion into three pieces:

• Flowing to technology developers, chips and computing hardware: 60%, about 3.1 trillion;

• To energy, transmission, cooling, electrical equipment: 25%, about 1.3 trillion;

• Flow to builders - land, materials, site development: 15%, about 800 billion;

The US$1.3 trillion in the middle represents the total cost of power infrastructure switching.

The time window is very concentrated. GB300 is still the last generation of the 54-volt era and will be mass-produced and distributed in 2026; the Rubin VR200 platform will be put into production in the second half of 2026, and some American cloud vendors have equipped it with independent high-voltage DC power cabinets as the first test of the 800-volt architecture; the Rubin Ultra Kyber platform will be released in the second half of 2027. This is 800 The complete architecture of 1.5V high-voltage DC has become standard for the whole cabinet for the first time.

The second half of 2026 to the second half of 2027 is the substantial starting point for industry chain switching.

This $1.3 trillion is not evenly distributed. At a glance, changes mainly occur in four directions:

• Transformers and electricity are the most out of stock and the most critical, with the largest single order amount. Players are concentrated in the hands of several century-old power companies.

• Wide bandgap semiconductor is a physical necessity for running high voltage and high frequency at 800 volts.

• The independent power supply cabinet next to the cabinet is where the value of a single cabinet power supply jumps to a new level.

• Automotive and photovoltaic industry chain, two cross-border players will enter the scarce position of AI data center.

The degree of bottlenecks in these four directions varies greatly, among which the transformer is the biggest bottleneck.

3. Bottleneck and betting position

1. Transformer: "TSMC" in the chain

In the entire AI industry chain, NVIDIA is very important, AMD and Intel are very important, storage is very important, and optical interconnection is very important, but only TSMC is the place where "the world is united", so we must look for it.

In the 800-volt high-voltage DC industry chain, transformers play the same role as TSMC.

To use electricity, the data center must first connect to the medium-voltage AC of 13.8 kV to 35 kV from the power grid. The medium-voltage AC is reduced to the low voltage that the data center can use (or directly reduced to 800 VDC), which requires a large power transformer. The supply and demand situation for this thing isridiculous.

On the one hand, the delivery cycle is extremely long. Wood Mackenzie's data for Q2 2025: Generator step-up transformer (GSU) average lead time is 143 weeks, almost 2.8 years, with extreme cases 210 weeks (4 years); power transformers 128 weeks, about 2.5 years.

On the one hand, demand is growing too fast. Since 2019, the demand for generator step-up transformers has increased by 274%, and the demand for power transformers has increased by 116%. During the same period, the unit price of power transformers has increased by 77%, and some distribution transformers have increased by 95%.

As a result, structural differences in the AI link emerged: GPUs iterate one generation in about a year, and it takes four years for an upstream transformer to be delivered.

In addition, geography exacerbates bottlenecks. China controls about 60% of the world's transformer production capacity; the United States imports 80% of its power transformers, and the procurement cost doubled after a 25% tariff was imposed. In April 2026, the U.S. government invoked the Defense Production Act to acknowledge that transformer production capacity was "hazardously limited". The "Defense Production Act" of the United States is usually used in wartime or major public crises. Transformers have been included in this list, and the level of industrial chain tension has been "Red Alert".

Wood Mackenzie's public calculations in the second half of 2025 show that the U.S. transformer market has entered a serious mismatch between supply and demand: the supply gap for power transformers in 2025 is about 30%-40%, and the demand for generator step-up transformers (GSU) is close to twice the available supply that year, with the gap approaching 100%. By 2026, this tension has not been lifted. Reuters said in December that the average delivery time for GSUs in the United States was still 143 weeks, and the average delivery time for power transformers was 128 weeks, with the delivery cycle still being two to three years.

By 2030, Wood Mackenzie expects the GSU shortfall to fall below 10%. The reason is not that demand has weakened, but that new production capacity has been put into operation one after another, and supply has begun to catch up. But the absolute incremental demand and price increases in the years 2026 to 2028 are still clear.

When demand has jumped to a new level and supply is still slowly climbing at the old production capacity, players in the track will benefit from this.

China’s position in the transformer industry chain is somewhat similar to that of semiconductors. Its production capacity ranks first in the world. Even if it is affected by the tariff wall, it is still growing rapidly.

Between 2022 and October 2025, U.S. imports of transformers from China soared from less than 1,500 units/year to more than 8,000 units/year. With the 25% tariff and multiple barriers, American customers are still grabbing Chinese goods because local production capacity simply cannot supply them. During the same period, the average price of China's transformer exports rose from US$12,000/unit to US$20,800/unit. China's transformer exports totaled US$9.037 billion in 2025, +36% year-on-year.

The main players:

TBEA is the de facto leader in China's transformer industry.

The company's full-year revenue in 2025 will be 97.318 billion yuan (-0.6% year-on-year), and net profit attributable to the parent company will be 5.954 billion yuan (+43.69% year-on-year). Revenue has not increased, but profits have increased by half. This structure usually means that the product structure is moving towards the high end.

On August 28, 2025, the company won the bid for the Saudi Electricity Company's (SEC) localized procurement framework order for ultra-high voltage and high-voltage transformers and reactors, with a total amount of approximately 16.4 billion yuan (a guarantee of 11.5 billion yuan based on a minimum execution volume of 70%), and a contract execution period of 7 years. This is the largest single overseas order in the company's history. Supply list: 179 extra high voltage transformers, 108 extra high voltage and high voltage reactors, 391 high voltage transformers, a total of 678 pieces of equipment. In November of the same year, it won the EPC contract for a 380kV substation in Saudi Arabia worth RMB 1.36 billion.

TBEA cannot directly sell to the US AI data center. There are three barriers: tariffs, political barriers, and lack of UL certification. But the blocked market is not all the market. The Belt and Road Initiative (Saudi Arabia, India, Southeast Asia) absorbs part of the global production capacity mismatch; domestic AI computing power construction requires its high-end products; after the Saudi factory is put into operation in 2027-2028, it will form a radiation radius in the Middle East. In the first half of 2025, the company's power transmission and transformation product contracts in the international market exceeded US$900 million, +88.10% year-on-year.

Jinpan Technology is going in the opposite direction. It is one of the few Chinese manufacturers that has squeezed into the first-tier supply chain of North American AI data centers. The company's full-year revenue in 2025 will be 7.295 billion yuan (+5.71% year-on-year), and net profit attributable to the parent company will be 660 million yuan (+14.82% year-on-year). It's definitely not big, and the order structure is completely different. In the first quarter of 2026, new overseas orders were 2.252 billion yuan, +280.73% year-on-year, accounting for 67.34% of new orders in the quarter; as of the end of the first quarter of 2026, overseas orders on hand were 5.140 billion yuan, accounting for 57.08% of the total orders on hand.

Jinpan’s product line extends from dry-type transformers to oil-immersed, switch cabinets, power modules, and solid-state transformers—it has transformed from a “single transformer supplier” to an “AI data center complete power solution provider.” The way to bypass tariffs is localized production in Mexico + a Virginia factory in the United States is under preparation. In January 2026, the Yuanshen ONE series solid-state transformer (10kV/2.4MW, efficiency 98%+) was released, aligning with NVIDIA’s 800-volt architecture.

China Xidian (601179.SH) is a national team of UHV state-owned enterprises. The company participates in more than 80% of domestic UHV projects. In 2024, the proportion of single-machine equipment exports exceeded 50% for the first time. In the first half of 2025, overseas business revenue was 2.171 billion yuan, accounting for 19.21% of total revenue. In 2025, subsidiary Xidian International won the bid for data center substation equipment in Malaysia. Xidian Changbian has received a large transformer supply order from GE. This path is a typical way for China Xidian to "indirectly enter the U.S. AI data center." Xidian Power Electronics, a subsidiary, already has the capability to develop 800-volt DC architecture solid-state transformers.

The next level down are several segmented players. Siyuan Electric (002028.SZ) Overseas revenue accounts for more than 30%, covering Europe, the Middle East and North America. Igor (002922.SZ) focuses on small and medium power distribution transformers in North America and has built a factory in Mexico.

Outside of China, the most intuitive player is Eaton.

CEO Paulo Ruiz characterized DC as "One of the biggest changes in the electric power industry since Edison's day" during the conference call. Eaton's financial report for the first quarter of 2026: revenue of US$7.5 billion (+17% year-on-year), of which data center orders increased by approximately 240% year-on-year, and data center-related revenue increased by approximately 50% year-on-year; the electrical Americas division's order reserve was US$14.5 billion, +44% year-on-year. Eaton's self-disclosed total data center order backlog has reached 228 gigawatts, which is equivalent to 12 years of order backlog based on the construction rate in 2025. Orders taken today will bedelivered until 2038. In March 2026, Eaton acquired Boyd Thermal for US$9.5 billion - incorporating the liquid cooling business and positioning itself as a complete solution extending from "electrical equipment" to "from grid to chip".

In addition, there are ABB, Hitachi Energy, Siemens Energy, GE Vernova, Schneider Electric which together form the world's first echelon of traditional transformers. Most of them are century-old companies, a group of companies that started building electric generators in the second phase of the industrial revolution, and are given two years to digest the 2030 demand curve. They are all expanding production, but even if they expand production at full capacity, the gap will not be filled before 2030.

In addition, the delivery pressure of traditional transformers has forced the alternative path of solid-state transformers.

Heron Power (unlisted) was founded by Drew Baglino, who served as senior vice president of energy business at Tesla. Heron's core product is a 5 MW solid state transformer that directly converts 34.5 kV medium voltage AC to 800 V DC. Completed $140 million Series B in February 2026. The most critical number in this financing announcement is not the $140 million, but that customers have expressed interest in more than 40 GW, including Intersect Power (the clean energy developer acquired by Google for $4.75 billion) and Crusoe (the developer of the 1.2 GW Stargate data center campus in Abilene, Texas). It plans to start trial production in early 2027 and build a US factory with an annual production capacity of 40 GW.

Enphase Energy’s turn is more interesting. Enphase is the global leader in residential photovoltaic microinverters, with cumulative shipments of 87.8 million units. Launched on April 28, 2026, the IQ SST is a 1.25 MW solid-state transformer with single-stage conversion from medium voltage AC directly to 800 VDC, targeting 98.5% efficiency and 99.999% availability. The bottom layer uses GaN power devices and the control chip uses 22nm Kestrel ASIC, both of which Enphase has been doing in microinverters for twenty years. Enphase’s own estimates put the U.S. AI data center IQ SST addressable market at more than 11 gigawatts by 2031.

2. Wide bandgap semiconductor: the “engine” in the chain

Wide bandgap semiconductor The term is a mouthful, but its place in the 800-volt data center is not.

The entire power link can be thought of as three sections:

In the first segment, electricity comes in from the grid, usually medium voltage AC from 13.8 kilovolts to 35 kilovolts.

In the second section, the data center converts it into 800 volts of direct current, and then sends it along the bus bar to the vicinity of the AI cabinet.

In the third paragraph, the power system inside the cabinet or next to the cabinet reduces the 800 volt DC step by step to 50 volts, 12 volts, 6 volts, and finally to about 0.8 volts that the GPU can really consume.

The semiconductor device is in the voltage conversion link of the second and third stages.

It is not a wire, nor a battery, nor a simply "conductive" material. It functions more like a set of high-speed electronic switches: turning the current on and off hundreds of thousands or millions of times per second, and converting one voltage into another voltage through this high-frequency switch. The faster the switching and the lower the loss, the smaller and lighter the power module can be made, the more power-saving and less likely to generate heat.

The problem is that after 800 volts of high-voltage DC enters the AI data center, the power system faces a very demanding combination: high voltage, high current, high frequency, and high power density. Traditional silicon-based power devices can also work in this environment, but at the cost of greater heat generation, lower efficiency, larger size, and more difficult lifespan and stability.

So two materials are needed: GaN and SiC.

SiC, also known as silicon carbide, is better at handling high-voltage and high-power scenarios. It is suitable to be placed at the front end of the link and is responsible for heavy tasks such as medium-voltage AC to high-voltage DC, or high-voltage DC main power conversion.

GaN, also known as gallium nitride, has the advantages of fast switching speed, low loss, and small size. It is more suitable to be placed at the back end of the link, responsible for quickly reducing 800 volts DC to 50 volts, 12 volts, and 6 volts in the cabinet.

A simple sentence:SiC tubes have higher voltage and heavier front end; GaN tubes have higher frequency and are closer to the back end of the server.

Once the 800-volt architecture is implemented, wide bandgap semiconductors will not be a dispensable material story, but an upgrade of the underlying devices that will determine whether the entire power link can run. GPUs need power, but GPUs can't feed off 800 volts directly. Every step of step-down, rectification, inverter, and voltage stabilization requires power semiconductors to open and close precisely like gates.

The implementation of the power war is a war to tame voltage.

Here, there are the following players:

Infineon Infineon is the world’s most certain player in the SiC front-end segment and the earliest player to enter the revenue realization period.

Its financial report numbers are the strongest anchor for this line: In fiscal 2025, Infineon's AI data center power solution revenue exceeded 700 million euros; in fiscal 2026, the company raised its revenue target from approximately 1 billion euros to approximately 1.5 billion euros; by fiscal 2027, management expects further growth to approximately 2.5 billion euros. Based on the company’s total revenue of approximately €14.7 billion in fiscal year 2025, this is no longer a scrap business, but is rapidly growing into a company-level growth curve. At the beginning of 2026, Infineon increased its annual investment plan by 500 million euros and the total investment amount to 2.7 billion euros. The core purpose is to expand AI data center power supply-related production capacity.

GaN backend: scattered overseas, Chinese players began to stand at the poker table

Among overseas players, onsemi, TI, ST, Power Integrations, and Navitas are all making plans around 800-volt data center power supplies. Power Integrations has joined Nvidia's 800V-related supplier list; Reuters mentioned that this list includes players such as Infineon and InnoScience.

Navitas is the most resilient pure GaN chip, with revenue of approximately $8.6 million in the first quarter of 2026, an increase of 18% sequentially. The company is shifting from consumer electronics to AI data centers and high-power markets. It demonstrated power strips ranging from 800 volts to 6 volts at GTC 2026, which shows that it has made progress in the technology position, but the real large orders still need to be verified after the 800 volt architecture is put into production.

On the Chinese side, Innosec is the purest GaN target, with revenue in 2025 of approximately 1.213 billion yuan, a year-on-year increase of approximately 46%; of which, AI data center revenue is 63.19 million yuan, a year-on-year increase of 50.2%, accounting for approximately 5.2% of total revenue. It’s still young now, but it has gone from concept to early implementation. Sanan Optoelectronics is more like an asset-heavy base. The annual report disclosed that in the field of data center and AI server power supply, it has achieved mass production for leading customers such as Great Wall, Vertiv Technology, Flextronics, Delta, and Lite-On, and participated in the development of HVDC high-voltage DC power supply solutions. Silan Micro is an IDM that is easily overlooked. In 2025, the second generation SiC-MOSFET has been used in the power supply of AI computing power centers and has been shipped in large quantities.

The most worthy of close attention for these companies is their progress in the supply chain in the 800-volt paradigm shift.

3. Independent power supply cabinet: the power supply must be moved out

In the 54-volt era, every AI server cabinet has a drawer-type power supply unit embedded in it, which costs tens of thousands of dollars per cabinet. When it comes to the 1 MW cabinet, this drawer will not fit in. The power supply unit will take up 64 cabinet units in the cabinet, and there is no room for the GPU.

The power supply must be physically moved out.

The solution demonstrated by Nvidia, Delta, Megmeet and other suppliers at GTC 2026 is to move the power supply unit, battery backup unit (BBU), and super-capacity unit out of the computer cabinet and form an independent power cabinet next to the GPU cabinet. The 1 MW GPU enclosure comes with a separate power supply enclosure. In the future, the computer room will no longer have GPUs in cabinets next to each other, but will have one cabinet of GPUs and one cabinet of power supplies.

How much is the value of a single independent power cabinet?

Morgan Stanley’s split on the Rubin Ultra Kyber platform gives about $216,000, of which the power distribution unit (PDU) is about $11,000, the power supply unit is about $115,000, the battery backup unit is about $38,000, the super-capacity module is about $16,000, and other accessories are about $36,000. This is several times the value of a single cabinet power supply in the NVIDIA GB200 era.

Some players:

Vertiv is the purest provider of critical data center infrastructure. It was clearly announced in October 2025 that the full 800-volt product line would be launched in the second half of 2026, in line with the NVIDIA Rubin Ultra platform. Revenue in the first quarter of 2026 was US$2.65 billion, adjusted earnings per share increased by 83% year-on-year, and net profit was US$390 million; the order backlog exceeded US$15 billion, corresponding to 12-18 months of forward revenue. Full-year 2026 guidance was raised to revenue of $13.5-14.0 billion (+34% year-over-year) and adjusted earnings per share of $6.35. The customer list covers almost all US hyperscale cloud vendors.

Eaton After bringing in the liquid cooling business through the acquisition of Boyd Thermal, its positioning has been extended from "electrical equipment" to "from grid to chip", and it has cooperated with Nvidia on the Rubin platform. Several specific numbers have been mentioned in the previous section on transformers - here it appears in another identity.

Delta (Delta Electronics) is the number one AI server power supply market share, estimated at 60-75%. Delta has jointly developed an 800-volt microgrid solution with NVIDIA, and a Chinese version of the Panama medium-voltage DC solution with Alibaba. Demonstrated at GTC 2026 660 kW inter-row power rack, 98.5% efficient solid state transformer. Morgan Stanley's supply chain research shows that Delta has cooperated with at least three U.S. ultra-large-scale cloud vendors on high-voltage DC pilot projects, and the first batch is expected to be implemented in the second half of 2026.

The above players all have other businesses, most of which are related to energy cooling. Due to space limitations, they cannot be discussed here. We will select targets worth paying attention to and expand on them later.

4. Cross-border position: the road paved by the automobile industry chain

It is not off-topic to talk about AI and the automobile industry. The interesting point is that in fact, the reason why 800-volt high-voltage DC can be transformed from a concept into a complete industrial chain within two years is precisely because the automobile industry chain has already paved the way.

800-volt high-voltage DC in data centers and 800-volt fast charging in electric vehicles are physically homogeneous at the underlying level, using the same type of GaN and SiC devices, the same LLC resonant topology, the same set of high-frequency technology, and the same type of liquid cooling. The supplier network that has been established in the automotive industry chain is almost ready for data center 800 volts.

The 800V penetration curve of China’s passenger car market provides the most direct evidence. Zosi Auto Research data: China's 800-volt architecture passenger car sales will reach 840,000 units in 2024, +185% year-on-year, with a penetration rate of 6.9%; the penetration rate is expected to be 9.5% in 2025, and Minsheng Securities estimates a three-year compound annual growth rate of 270.9%. Leapao B01 has packed an 800-volt high-voltage platform, lidar, and 650-kilometer battery life into a starting price of 89,800 yuan. Passenger cars in the 100,000-yuan class are all using 800-volt fast charging.

The players on this cross-border road are divided into two groups.

Heron Power and Enphase Energy have already been mentioned before. One company has crossed over from Tesla to make solid-state transformers, and the other has crossed over from photovoltaic micro-inverters to make solid-state transformers. Their scarcity lies not in how deep they go, but in the crossover itself. It is difficult for a pure new force starting from scratch to put together a complete product line within two years, and it is much faster to migrate from the existing production capacity of electric vehicles or photovoltaics.

Several A-share companies are another group of reverse crossovers. They were originally engaged in car fast charging, vehicle power supplies, and energy storage batteries, but now they are entering AI data centers using the same type of devices and topology:

• Megmeet: The only mainland Chinese manufacturer in the A-share market that has entered Nvidia’s AI server power supply chain. Small batch deliveries of GB300’s 33 kilowatt high-voltage DC power supplies will begin in the second half of 2025. Delta, Lite-On, and Flex are all Taiwanese or overseas companies, and Megmeet’s scarce position in the industry chain lies here.

• Jianghai Shares: Supercapacitors exclusively supply the 33-kilowatt supercapacity solution for ByteDance’s data center.

• Azure Lithium Cell: The battery backup unit uses small cylindrical lithium cells.

• Youyou Green Energy: A leader in DC charging modules, it plans to enter the AI data center power supply system through ABB channels.

• Shenghong Electric Co., Ltd., Hewang Electric: Key subcontractors for the Vertiv 800-volt system.

These companies currently do not have a high proportion of AI data center-related businesses.

The opening chapter "The Battleship Dreadnought Being Towed for Disintegration" is an oil painting created by the British artist William Turner in 1838. The picture shows the British sailing battleship "Dreadnought", which had made great contributions in the Battle of Trafalgar, being dragged to the disintegration site by a small steam tug emitting black smoke. The sunset, dusk, and fog made the entire river seem to be saying goodbye. The BBC voted it the greatest British painting.

It is the painterly definition of a paradigm shift, the moment when the old era is dragged away by the new. The tugboat is small and inconspicuous, but it is the protagonist of the new era; the battleship is huge and gorgeous, but it is the one that is to be sent away.

I find this kind of image of "the small one pulling the big one and the inconspicuous one deciding the overall situation" very interesting to me.