Power supply structure begins to determine computing power speed

From 48V to 800V, and then to solid-state transformers, the bottlenecks in AI data centers are continuing to move from chips, networks, and liquid cooling to more upstream power electronics and energy supply.

800V is not an isolated voltage, but a system reconstruction that reduces conversion, reduces copper consumption, and frees up cabinet space.

SST truly puts medium-voltage power grids, power electronics, SiC high-voltage devices and data centers into one picture.

Energy data centers are beginning to compete like industrial loads for power, transformers, gas turbines, energy storage and grid connection speed.

Picture: Self-made diagram. High-voltage DC puts the grid side, cabinet side and power devices into the same picture.

1. The real change is not the voltage, but the upward shift in power.

The hottest words in AI data centers in the past few years have been GPU, CPO, liquid cooling, and HBM. Looking one step further, the question becomes simpler: in what form does the electric energy enter the computer room, after several conversions, where is the loss, who will handle the transient load, and who can deliver transformers, circuit breakers, gas turbines and energy storage at the project site on time.

BEDROCK’s judgment is that the bottleneck of AI infrastructure is moving from “computing power equipment” to “power structure”. 800V HVDC is the first layer of this round of changes, SST is the next layer of deeper power electronics, and the change in energy structure is the slowest and most difficult to bypass the base.

This line is best read in four steps: cabinet power first pushes the old power supply path to the wall; 800V moves the conversion forward and reduces the current; SST electronically digitizes medium-voltage power distribution; in terms of investment, look at how the content per MW extends from cabinet power supply, facility-level HVDC, high-voltage SiC, all the way to energy supply.

2. Why now: Cabinet power pushes old architecture to the wall

Traditional data centers can tolerate multi-level AC/DC and DC/DC conversion because the power of a single cabinet is low, and copper rows, PSU space, and conversion losses are all within acceptable limits. AI cabinets are different. After GB200/GB300, 120kW to 160kW is no longer an exaggeration; the next step is 480kW, 600kW, and 1MW cabinets starting to enter the road map.

NVIDIA's public statement is very straightforward: traditional 54V cabinet power supply is not prepared for MW-level cabinets. At the 1MW level, if low voltage and high current are used, copper, space and heat loss will become hard limits. The 800V architecture goal given by NVIDIA is to support cabinets from 100kW to more than 1MW, reduce the number of conversion stages, reduce copper usage, and increase end-to-end efficiency by up to about 5%.

Figure: After the cabinet power is increased, the power chain is forced to shift gears from an in-server problem to a facility-level problem.

After the cabinet power exceeds 200kW, the space and copper consumption pressure of the low-voltage and high-current path will quickly increase.

There is a more realistic trigger for power architecture changes: data centers are not evenly distributed loads on the power grid, but huge loads concentrated in a few campuses and nodes. The IEA estimates that global data center electricity consumption will be approximately 415TWh in 2024, accounting for approximately 1.5% of global electricity consumption; it may more than double to approximately 945TWh by 2030. The caliber of the U.S. DOE/LBNL also shows that U.S. data center electricity consumption will be approximately 176TWh in 2023, accounting for 4.4% of total electricity consumption, and may rise to 325-580TWh in 2028, accounting for 6.7%-12%.

Therefore, the AI ​​data center is not a problem that can be solved by buying a few more UPSs. It begins to require a new power structure and a new way of organizing energy.

3. 800V changes the "number of conversions" and "current scale"

The current mainstream path is roughly as follows: the medium-voltage power grid enters the campus, passes through traditional transformer and UPS, and then to AC power distribution in the computer room. Finally, it is converted from PSU to 48V/54V on the cabinet side, and then enters 12V and 0.xV. At every level you go, there are conversion losses, equipment space, points of failure, and thermal management issues.

The idea of ​​800V HVDC is to advance the high-voltage DC bus to the facility side or computer room side, so that the electric energy can enter the rows and cabinets in the form of higher voltage and lower current. The physics is very simple: under the same power, the higher the voltage, the lower the current, and the copper loss and copper usage decrease. The engineering is not simple: circuit break protection, safety standards, insulation, operation and maintenance, power backup, DC/DC topology, and device reliability all need to change accordingly.

Figure: The key differences between the old path and the 800V path are fewer transitions and fewer points of failure.

There's a useful divergence here: ±400V may land sooner, 800V unipolar is more aggressive. ±400V is closer to the existing overseas AC power system, safety standards and engineering experience, so paths such as Google and Meta will enter construction and verification earlier. NVIDIA's 800V route is more representative of future high-density cabinets, but supporting circuit breakers, switches, safety specifications and collaborative verification are slower.

This also explains why the industry chain experience is not completely consistent. The technical direction is clear, but to truly scale up production on a large scale, we need to work together from servers, power modules, row power distribution, power backup, circuit breaker protection, computer room engineering to acceptance standards. Rubin 800V is delayed several months from the original plan and does not change direction, which only shows that this matter is no longer a project that a single supplier can independently advance.

route more like what Benefits Difficulty 48V/54V Server internal optimization Mature ecology, sufficient experience in safety and maintenance Copper consumption, space, and the number of PSUs in MW-level cabinets are under great pressure ±400VHVDC More stable high-voltage DC transition Easier integration with existing facility voltage and safety systems DC protection, backup and row power distribution standards still need to be redone 800VHVDC Main path to 1MW cabinet Reduce current, copper, number of conversion stages, and in-cabinet power space Circuit break protection, insulation, safety standards, and DC/DC device routes must all mature SST Medium voltage power grid direct connection calculation load Smaller size, faster response, and higher system integration High-voltage SiC, solid-state protection, thermal management and reliability verification are more difficult

4. The location of SST: not the device, but the power distribution form

800V solves the problem of cabinet density, and SST solves the problem of campus-level power distribution. Solid-state transformers are most easily misunderstood as "more advanced transformers," but what they really change is the control method: traditional transformers do electromagnetic conversion, and SST puts power semiconductors, control software, protection, isolation, and bidirectional energy flow into it, making the power system as controllable as an electronic system.

Why does the AI ​​data center push SST to the forefront? Because the load is too concentrated and the power changes too quickly, the delivery cycle, size, protection speed and maintainability of traditional large transformers and mechanical circuit breakers begin to appear cumbersome. The IEA noted that new transmission lines in advanced economies typically take 4-8 years, and waiting times for transformers and cables have doubled in the past three years. A structural mismatch is emerging between the speed of data center construction and the speed of power infrastructure.

More importantly, the AI ​​data center is not a docile load in the eyes of the power grid. Training and inference clusters will see power jumps up and down on the millisecond to second level. Hundreds of thousands of GPUs/ASICs are down or up at the same time, which may correspond to transient changes of several MW or even larger. Traditional transformers are essentially passive devices that will transmit this voltage, current, phase and reactive power disturbance back to the upstream power grid; the value of SST lies in active regulation, turning "loads that make the grid uncomfortable" into "loads that the grid can manage".

A rhythm calibration needs to be done here: the long-term value of SST does not mean that it will immediately become the main line of income in 2026-2027. The constraints on the engineering side are very tight: at this stage, the cost of SST may be about 35% higher than the traditional solution, the system efficiency is only improved by about 1 point, and the data on high-voltage devices above 15kV, long-term reliability and grid-connected operation are still insufficient. A more reliable base case is: 2026-2027 to see prototypes, testing, network installation and small batch verification; 2028-2030 to see the introduction of a few high-power parks; after 2030, we will discuss whether to enter the mainstream power distribution solution.

This means that medium-voltage UPS, 800V/±400V HVDC rectifier cabinets, BBU/CBU, and solid-state circuit breakers will receive orders earlier than SST. They are not legacy devices that are immediately replaced by SST, but rather a transitional layer to SST.

Figure: The value of SST is to make AI loads controllable, not just to replace transformers, but to handle transients and reactive power.

Figure: SST brings the “grid side” and “cabinet side” closer. The real elasticity of SiC is on the high-voltage side.

This step will raise the position of SiC. The amount of SiC used in 800V cabinets may not explode immediately. Many 800V to 50V and 12V solutions can weigh between SiC, GaN, and silicon MOS. What really makes the value of SiC steeper may not be the few devices in the cabinet, but the high-voltage infrastructure outside the cabinet: SST, solid-state circuit breakers, high-voltage DC protection and medium-voltage DC.

In other words, SiC was a high-efficiency option in the 800V era and more of a bottom-tier condition in the SST era. GaN's position is more focused on high-frequency, high-power-density DC/DC links, especially where 800V to 50V, 12V, and 6V conversion requires shrinking magnetic components.

Figure: The division of labor between GaN and SiC: one is biased towards high frequency density, and the other is biased towards high voltage tolerance.

5. The investment implications of SiC: Will the capacity per MW jump?

SiC This paragraph cannot only write about technology. For investment, the most important question is not "will 800V use SiC?" but how much SiC content each megawatt of AI data center will bring. If this number is only a few thousand US dollars/MW, it is difficult for the theme to support a big opportunity; if it reaches hundreds of thousands or more than 200,000 US dollars/MW, it will change from a side branch of automotive SiC to a new main line of AI power infrastructure.

This is also where this round of SiC discussion is most likely to be misunderstood. The narrow aperture only looks at non-SST 800V cabinet devices, and the value of SiC may indeed be very small, and may even be shunted by GaN and silicon MOS; the wide aperture includes UPS/PDU/BESS, 800V busbars, solid state circuit breakers, SST, and high-voltage DC protection, and the content of SiC will span an order of magnitude.

BEDROCK's judgment is: 800V opens the entrance to SiC, and SST and solid-state protection determine the ceiling of SiC. If there are only a few devices in the 800V power shelf in the end, the magnitude is not enough; if high-voltage DC, SST and solid-state circuit breakers become the default architecture of new AI data centers, SiC's TAM caliber will be rewritten.

Figure: SiC is about high voltage and high power, GaN is about high frequency and miniaturization. It is more important to clearly explain the division of labor in materials than to pile up application icons.

Adapted from Wolfspeed's public materials and Yole 2025 caliber; the illustration is used to illustrate the material's applicable range and does not represent a single company's forecast.

Figure: The SiC investment framework looks at different calibers of content per MW, and the conclusion is an order of magnitude different.

The numerical values ​​are scenario frameworks used to illustrate differences in calibers and are not guidance for a single company.

This change in magnitude is critical. According to the narrowest caliber, the non-SST 800V solution may only have a few dozen SiCs per MW, with a unit price of tens of dollars, which works out to only a thousand dollars per MW; this is not a market that can independently support a large market value revaluation. According to the wide caliber, the current insertion of UPS, PDU, BESS and other peripherals can reach $20-40k/MW; to 800V transition and early SST, it may go up to $50-100k/MW; if the new AI data center in 2028-2030 adopts 800V+DC and SST by default, and high-voltage conversion, protection, energy storage and switching equipment are all SiC-based, the content may become $145-275k/MW.

It will be more intuitive to change to the scale of the data center. A 100MW AI data center may only cost hundreds of thousands to millions of dollars in SiC in a narrow caliber; with a flipped architecture, it may become $14.5-27.5M. By 2030, if 25-30GW of new annual AI data center deliveries do occur, with content approaching $145-275k/MW, data center SiC annual TAM could be pushed to $3.6-8.3B. This number shouldn’t be copied mechanically, but it explains why high-voltage SiC assets are back in the picture: The disagreement is no longer just about automotive SiC repairs, but whether AI power infrastructure can open up the second curve.

This change has also begun to occur on the corporate side. Wolfspeed disclosed in FY26 Q3 that AI data center-related revenue increased by approximately 30% quarter-on-quarter and launched the first commercial 10kV SiC MOSFET for grid modernization, industrial electrification and AI data center infrastructure. This information should not be written as a single company propaganda, but it illustrates that the industry direction has expanded from "automotive SiC" to "high voltage infrastructure SiC."

10kV SiC isn’t just about “higher voltage” either. In a high-voltage power supply, if the voltage of a single device is not enough, multiple devices can only be connected in series; if the current is not enough, they can only be connected in parallel. Series and parallel connection will bring more driving, protection, voltage equalization, heat dissipation and reliability problems. Once a high-voltage device is reliable, its value is not just to sell one more MOSFET, but to reduce the control complexity at the system level. This is also what SST is most concerned about in high-voltage, strong protection and strong reliability scenarios.

Magnitude Approximate content investment implications key verification Non-SST/in-cabinet devices Thousands of US dollars/MW to low 10,000 US dollars/MW It is just a topic entrance and cannot be directly extrapolated into a large TAM.

800V→50V/12V/6V solution finalized, PSU/Power Shelf dismantled.

UPS / PDU / BESS peripheral plug-in $20-40k/MW Can start to enter the revenue, but still partial to incremental devices. Proportion of SiC procurement in data center energy storage, UPS, and PDU. 800V Transition/Early SST $50-100k/MW There is visible elasticity in the company's income, and 2027 is the observation window.

±400V/800V greenfield projects, solid state circuit breakers and DC protection pilots.

Complete 800V+SST architecture

$145-275k/MW

The opportunity to truly rewrite the data center SiC TAM has been moved back to 2028-2030.

SST mass production, high voltage thick substrate, 10kV/6.5kV device reliability.

Therefore, the essence of SiC investment is not "immediate amplification of demand", but "shifting the magnitude distribution later". 2026 still looks likely to be small, 2027 to see real projects for 800V and solid state protection, 2028-2030 to see if SST and high voltage infrastructure push the capacity per MW up. The companies that are really worth studying are not all SiC names, but assets that can obtain high-voltage devices, thick substrates, reliability certifications, and a North American customer system.

6. From an investment perspective, who takes away the value?

Putting this line on investment, we can’t just ask “who benefits?” What is more important are three questions: how much is the value of a single MW in this link, when does it enter revenue, and do you have bargaining rights after entering revenue. Changes in the AI ​​power supply structure do not lift all links at once, but are implemented in layers based on time and architecture.

BEDROCK's rankings are: short-term certainty in rack power, liquid cooling, UPS/HVDC and conventional power equipment; medium-term resilience in 800V HVDC facilities and solid-state protection; and long-term odds in SST, high voltage SiC and thick substrates. The most common mistake is to put these links into the same timetable for valuation.

Deterministic ordering: Power equipment/gas turbines/transformers > 800V/±400V HVDC and medium voltage UPS > Rack power supplies/liquid cooling > high voltage SiC/SST.

Order of Odds: High Voltage SiC/SST > 800V HVDC Facilities & Solid State Protection > Rack Power Supplies/Liquid Cooling Pumps > Traditional Power Equipment.

Looking at the two tables backwards is the difficulty in investing in this line. Benefit link magnitude and rhythm Investment evaluation and key verification In-cabinet power supply and board-level power supply

2025-2027 will be the first to be honored. PSU, power shelf, DC/DC, DrMOS, GaN, and controller are shipped with high-power cabinets.

High certainty, medium elasticity. More like revenue encashment and share competition, not suitable for extrapolation by SST level TAM. Look at the full cabinet shipment, power shelf design, 800V→50V/12V/6V topology, customer socket and ASP.

Liquid cooling and thermal management

It will be uploaded simultaneously with the 120kW-1MW cabinet power, and the redemption time will be earlier than SST.

Deterministic, but not exclusive to 800V. It is more suitable as a supporting main line for high-power cabinets. Look at rack power routing, cold plate/immersion/CDU standards, customer onboarding and O&M costs.

800V/±400V HVDC and medium voltage UPS

Observation window 2026-2028. AC/DC, row distribution, medium voltage UPS, BBU/CBU, solid state circuit breakers and high voltage DC protection are starting to get thicker.

The most important main line in the mid-term. The value of a single MW does not necessarily fall on semiconductors, but the system value and entry barriers will be significantly increased. See Google/Meta's ±400V, NVIDIA/Rubin's 800V, North American safety certifications, OCP/NVIDIA routes, system partner list and delivery cadence.

High voltage SiC, thick substrates and modules

From $20-40k/MW peripheral insertion to $145-275k/MW complete architecture scenario, the time is more towards 2027-2030.

The odds are the best, but cash out is delayed. The real barriers are not ordinary devices, but high-voltage products, reliability, substrate quality and customer certification. Look at 3.3kV/6.5kV/10kV device volume production, yield, long-term reliability, substrate supply and North American customer qualification.

SST and Solid State Protection

More like architectural options circa 2030. Once traditional transformers and mechanical protection are replaced, the system value will increase significantly.

Big odds, slow redemption, and strong project-based attributes. The value is not just transformer replacement, but high voltage conversion, fast protection, reactive power management and load buffering. 2026 revenue cannot be used to judge long-term value, nor can early prototypes be directly extrapolated to volume TAM. Look at SST pilots, 15kV+ device maturity, long-term operational data, solid state circuit breaker failure response, module redundancy, O&M responsibilities, grid access approvals and real-world deployments at major customers.

Power supply and grid connection equipment

Already cashing in now. Gas/medium speed turbines, transformers, switches, power EPC, energy storage and grid connected resources all benefit from power scarcity.

The highest certainty, but the valuation logic is more like industrial backlog and power resource repricing, not pure 800V theme elasticity. Look at the grid connection queue, transformer delivery date, gas turbine orders, PPA, electricity prices, park site selection and self-built power station plans.

In the future, we will see many large SST market scales. For example, it is assumed that the data center power supply penetration rate will reach 10%-15% in 2028, or the global SST market will be pushed to hundreds of billions of yuan in 2030. Such numbers are useful, but are better suited as upper-bound stress tests for the bull case rather than the base case. What really determines whether it can be realized is not how big the TAM is written, but whether the cost can be reduced from about 35% higher than the traditional solution, whether the efficiency improvement can cover the complexity of transformation, whether the 15kV+ devices can mature, and whether major customers are willing to assume the operation and maintenance responsibilities of the new power distribution architecture.

Calculation examples, not investment advice. The following two examples are only used to illustrate how structural changes can expand the company's serviceable market, and do not represent buying and selling recommendations, target prices or any investment operations.

Company example Calculation caliber Potential annual revenue opportunity in 2030 Delta

Traditional PSU/power shelf is about $0.25-0.4m/MW; 800V HVDC puts in-row power, DC/DC shelf, BBU/CBU, e-fuse and busway, it may be $0.6-1.0m/MW; if it is further packaged with SST, CDU, microgrid, the caliber can be $1.2-2.0m/MW.

Base case: 25GW AI data center is newly added, 40% adopts 800V/SST, Delta 25% share, $1.2m/MW content volume, corresponding to about $3.0bn, about NT$90-95bn. This magnitude is approximately 0.16-0.17x Delta's 2025 revenue, or 0.12-0.13x current 2026 market revenue forecasts. The opportunity comes from the system package, not the individual PSU.

Wolfspeed

Non-SST in-cabinet SiC is only at the thousand-dollar/MW level; high-voltage peripherals and early SST can reach $50-100k/MW; under the complete 800V+SST architecture, high-voltage SiC, solid-state protection and thick substrate caliber can reach $145-275k/MW.

Base case: 25GW new additions, 40% entering high-voltage SiC/SST caliber, WOLF 25% share, corresponding to approximately $0.36-0.69bn, equivalent to 0.47-0.91x of FY25 revenue, or 0.60-1.15x of FY26 Q3 annualized revenue. Bull case If most of the 25-30GW per year enters the complete structure, the industry TAM can reach $3.6-8.3bn, and 30% share of WOLF corresponds to $1.1-2.5bn, which is equivalent to 1.45-3.30x of FY25 revenue.

Delta's example shows that the significance of 800V/SST to system vendors is to expand from "power supply in the cabinet" to "the entire power path from medium voltage to the chip." Delta’s own showcase products already include 1.1MW class 800V In-Row Power, 90kW DC/DC Power Shelf, 660kW Power Rack, BBU, 2.4/3MW CDU, SST and microgrid. Using the revenue of approximately NT$555bn in 2025 as a reference, the annual revenue opportunity of NT$90-95bn in the base case is equivalent to 16-17% of the revenue in 2025; based on the current market revenue forecast for 2026, it is approximately 12-13%. Its flexibility does not come from a single power shelf ASP, but from the customer's willingness to hand over power, cooling, backup and energy management to the same system vendor.

Wolfspeed's example is more on the odds side. The company has launched commercial 10kV SiC MOSFET, and in FY26 Q3, it also revealed that AI data center applications have increased by approximately 30% month-on-month. Taking FY26 Q3's annualized income of $150m as a reference, which is about $0.6bn, the base case's $0.36-0.69bn is already 60-115% of the current annualized income, while the bull case may be several times higher. But we can’t just look at TAM here, but also look at mass production yield, customer qualification, balance sheet, gross profit repair, and whether high-voltage devices really enter the SST/solid-state protection mass production project. The opportunity for WOLF is great, but the path to cash out is later, narrower, and more dependent on technology and financial execution than Delta.

The most important magnitude judgment here is: in the short term, look at revenue realization, and don’t expect the capacity per MW to jump immediately to the forward scenario; in the long term, look at the architecture reversal, and don’t deny high-voltage SiC and SST just because the numbers in 2026 are small. In other words, the cabinet power supply and liquid cooling are responsible for determinism, 800V HVDC is responsible for mid-term flexibility, high-voltage SiC and SST are responsible for forward odds, and the power supply is responsible for whether the entire AI infrastructure can be online on time.

7. Behind the power structure is the reordering of the energy structure

800V and SST solve the problem of how to make electric energy more efficient, more compact and more controllable after it enters the cabinet. But there’s another slow variable in AI data centers: where the electricity comes from.

The IEA's judgment is very restrained: by 2035, the new power consumption of data centers will be met by multiple sources such as renewable energy, natural gas, and nuclear power. Renewables will account for a significant increase, but dispatchable power is equally critical. The practical implication here is that AI data centers will not address power delivery with just one “clean narrative.” It will pull park site selection, PPA, natural gas units, energy storage, nuclear power, geothermal, demand response and grid transformation into the project model.

The industry can also see this direction. Large data centers have begun to consider off-grid or semi-off-grid solutions in some areas, and the combination of natural gas medium-speed engines, gas turbines, backup power, and energy storage has been re-discussed. Energy storage may not be the protagonist in the pure gas main power supply scenario, but it will have a place in new energy access, load fluctuations, grid connection constraints and redundancy strategies.

Figure: The energy structure of AI data centers is no longer a single issue. Line speed, reliability and low-carbon goals must be calculated simultaneously.

So, 800V is not the end of the energy issue. It just makes electricity more usable inside the computer room. What really determines whether the data center can go online on time is still the upstream power supply, grid connection queue, equipment delivery date and local electricity price. In the future, competition in AI infrastructure may become more and more like a competition in “power acquisition capabilities.”

In the end, what is really scarce may not be a certain power device, but the ability to deliver power grids, power semiconductors, backup power, thermal management and cabinet engineering together.