AI was expected to drive unprecedented demand for GPUs, accelerators and memory. It is now creating a second semiconductor market in the transformers, power supplies, cooling systems and electrical infrastructure surrounding those processors.
The transformer does not usually receive much attention in discussions about artificial intelligence. It lacks the glamour of a new GPU, the strategic intrigue of a custom accelerator or the supply-chain drama of high-bandwidth memory. It is one of those pieces of infrastructure that most of us notice only when it fails or when there are not enough of them.
That may be about to change.
An Ars Technica report describes AI data centers as a potential killer application for solid-state transformers. These systems replace much of the passive copper-and-steel architecture of a conventional transformer with power semiconductors and digital controls. They can change voltage, convert alternating current to direct current, regulate power quality and coordinate electricity from the grid, batteries and on-site generation.
That is a much bigger step than improving a piece of electrical equipment. It is part of a transformation stretching from the utility connection to the processor package. AI is turning the entire power path into a semiconductor system.
The GPU creates the demand, but an increasingly large collection of silicon-carbide switches, gallium-nitride converters, conventional silicon power devices, analog components, sensors, microcontrollers and integrated voltage regulators will determine how much electricity actually reaches it.
A New Kind of Electric Load
Most electric grids were built to support relatively predictable demand. Even large industrial customers generally increase and decrease their electricity use on timescales that utilities understand.
AI clusters can behave differently. Thousands of GPUs may move almost simultaneously between computation and communication, producing rapid changes in power demand. The magnitude of those changes rises as data centers move from racks consuming tens of kilowatts toward systems consuming hundreds of kilowatts and, eventually, more than a megawatt per rack.
Traditional data-center power architecture also requires electricity to pass through a series of conversion steps. Medium-voltage utility power is stepped down and distributed as AC. It is converted to DC in power supplies and stepped down repeatedly until it reaches the sub-one-volt level required by modern processors. Every conversion consumes energy, creates heat and adds equipment that can fail.
At earlier data-center densities, those losses were tolerable. At gigawatt campus scale, a seemingly minor percentage becomes the output of a small power plant.
The industry’s response is taking shape as a semiconductor chain extending from grid to GPU.
Replacing Iron With Intelligence
A conventional transformer changes voltage through electromagnetic induction. It is reliable and well understood, but fundamentally passive. It cannot independently redirect electricity, compensate for rapidly changing loads or decide when to draw power from a battery instead of the grid.
A solid-state transformer is essentially a high-power electronic system. In addition to power switches, it requires gate drivers, digital isolators, voltage and current sensors, microcontrollers, communications devices and protection circuitry. Software can monitor and control its behavior in real time.
Silicon carbide, or SiC, is central to making this practical. Compared with conventional silicon, SiC can operate at higher voltages, switch more rapidly and tolerate higher temperatures while losing less energy as heat. Those qualities allow engineers to design power converters that are smaller, lighter and more efficient than equivalent silicon systems.
Wolfspeed says 1,200-volt SiC MOSFETs used for AC/DC and DC/DC conversion can reduce conversion losses by 25% to 40% within those stages. Upstream from that, higher-voltage SiC makes it possible for solid-state transformers to convert medium-voltage grid power directly into the DC power used by an AI data center. Wolfspeed describes SiC as foundational to the complete path from grid connection to rack.
In March, the company introduced what it calls the industry’s first commercially available 10-kilovolt SiC power MOSFET. The product targets grid modernization, industrial electrification and AI data centers. Wolfspeed has also introduced 3.3-kilovolt power modules for medium-voltage systems.
These are still emerging technologies. Solid-state transformers have not suddenly replaced the enormous installed base of conventional transformers. Reliability, cost, standardization, protection and serviceability all need to be proven at data-center scale. Nevertheless, AI may provide the demand and economics needed to pull high-voltage SiC out of specialized applications and into a much larger infrastructure market.
Why 800 Volts Matters
The power architecture inside the data center is also changing.
Current AI racks generally rely on 48- or 54-volt DC distribution. As rack demand rises, low-voltage distribution becomes physically difficult. Power equals voltage multiplied by current. Supplying more power at a low voltage requires enormous current, heavy copper busbars and additional cooling.
Nvidia has estimated that delivering 1 megawatt through a 54-volt architecture could require roughly 200 kilograms of copper busbar in each rack. Power shelves would occupy space that could otherwise hold computing equipment.
The proposed answer is 800-volt direct current.
Nvidia, Google and Microsoft have been developing an 800 VDC architecture through the Open Compute Project. Instead of repeatedly stepping down and converting electricity throughout the building, a future AI factory could convert medium-voltage AC to 800 VDC near the facility entrance and distribute that power across the data hall.
Nvidia’s roadmap includes 800 VDC power racks for retrofitting existing AC facilities, row-level power centers supporting as much as 2 megawatts per row and, eventually, power blocks converting utility power directly to 800 VDC. More than 80 infrastructure and equipment companies are reportedly developing products around the emerging specification. Nvidia’s roadmap shows a phased migration rather than an overnight replacement of existing facilities.
The architectural shift reduces current, copper and the number of conversion stages. It also creates a common DC backbone to which batteries, renewable generation and other energy sources can connect more directly.
This brings solid-state transformers and 800 VDC together. One changes how electricity enters the data center. The other changes how it moves through the building. Power semiconductors connect the two.
SiC at the Heavy End, GaN Closer to the Rack
Silicon carbide and gallium nitride are frequently grouped under the wide-bandgap semiconductor label. Their properties overlap, but they are likely to fill different roles.
SiC is especially attractive for high-voltage, high-power applications. Those include the grid interface, solid-state transformers, large uninterruptible power supplies, battery inverters and microgrids. It is likely to carry much of the load from the utility connection through the 800 VDC backbone.
Gallium nitride, or GaN, can switch exceptionally quickly. Higher switching frequencies permit smaller transformers, inductors and capacitors, making GaN attractive for the compact, high-density conversion systems installed near racks and processors.
STMicroelectronics has demonstrated a 6-kilowatt design that converts 800 VDC to 12 volts with a claimed 97.5% peak efficiency and power density of approximately 2,500 watts per cubic inch. It has also demonstrated an 800-to-6-volt design intended to eliminate another intermediate conversion stage. ST’s reference designs combine GaN, silicon MOSFETs, mixed-signal devices and advanced packaging.
Texas Instruments has developed a similar two-stage architecture. Its first stage converts 800 volts to 6 volts using integrated GaN power devices. A multiphase silicon power stage then converts 6 volts to the sub-one-volt power used by the processor. TI reports 97.6% peak efficiency and power density above 2,000 watts per cubic inch for its first stage. The TI architecture also includes an 800-volt hot-swap controller that protects equipment and isolates failures.
This will not necessarily be a clean division in which SiC wins one market and GaN another. Cost, voltage, frequency, availability and reliability requirements will determine the material used at each stage. ST, Infineon, onsemi, ROHM and others are building portfolios that combine SiC, GaN and traditional silicon.
The future powertrain will probably use all three.
The Hardest Part May Be the Last Few Inches
Moving a megawatt to the rack does not solve the final power-delivery problem. A GPU may operate below one volt while drawing extraordinary current. Once voltage drops, current rises. Moving thousands of amps laterally across a circuit board creates resistance, heat and voltage instability.
Power delivery is therefore moving closer to the processor.
Vertical power delivery places voltage-regulation components underneath or immediately adjacent to the GPU rather than around its perimeter. Package-integrated voltage regulators go further by bringing portions of the power system into the processor package.
Infineon’s quad-phase power modules can deliver as much as 280 amps and use embedded-die packaging and integrated magnetics to fit beneath a processor. Infineon says the modules can reach current densities of 2 amps per square millimeter.
Marvell’s package-integrated voltage regulator reduces higher supply voltages to the sub-one-volt levels consumed by advanced processors. The company claims that shortening the final high-current path can reduce transmission losses by as much as 85%. Vicor, Empower Semiconductor, Monolithic Power Systems and TI are pursuing their own versions of vertical or highly integrated power delivery.
This adds another dimension to advanced packaging. AI packaging has generally been discussed in terms of connecting GPUs, chiplets and HBM. The processor package must now deliver enormous amounts of current and remove the resulting heat. Embedded power devices, integrated magnetics, silicon capacitors, top-side cooling and direct cold-plate compatibility become part of the package roadmap.
The package is becoming part of the power grid.
Conventional Silicon Still Has Plenty of Work
The attention given to SiC and GaN can leave the impression that conventional silicon is being pushed aside. That is not what the emerging system looks like.
Each wide-bandgap power switch requires supporting components. Gate drivers tell it when to switch. Analog-to-digital converters measure current and voltage. Digital isolators protect low-voltage controls. Microcontrollers manage switching behavior. Electronic fuses and hot-swap devices isolate faults. Communications ICs pass operational data into facility-control software.
Traditional silicon also remains highly competitive in the final low-voltage conversion stages closest to the GPU.
That spreads the AI semiconductor opportunity across a much wider group of suppliers. Nvidia’s 800 VDC ecosystem includes Analog Devices, Infineon, Innoscience, Monolithic Power Systems, Navitas, onsemi, Renesas, ROHM, STMicroelectronics and Texas Instruments. These companies are not trying to displace the GPU. They are supplying the devices that allow more GPUs to operate inside a fixed power envelope.
Much of this analog, mixed-signal, embedded and control silicon is manufactured on mature process nodes. AI capital spending is therefore beginning to extend beyond leading-edge logic, HBM and advanced packaging into parts of the semiconductor industry that have received far less attention.
STMicroelectronics offers an example of the scale of that opportunity. The company now expects approximately $1 billion in data-center revenue in 2026 and has said that figure could double in 2027 if current demand continues. Its data-center portfolio includes SiC, GaN, silicon power devices, mixed-signal chips, microcontrollers and motor drivers rather than AI accelerators.
From Passive Power to Programmable Power
Semiconductors provide something beyond conversion efficiency: They give software control over electricity.
A semiconductor-based power system can expose real-time telemetry, regulate voltage, isolate faults, dispatch batteries and manage bidirectional power flows. It can coordinate the utility feed with on-site generation and storage. It can respond to rapid fluctuations in GPU demand before those fluctuations propagate into the wider grid.
Analog Devices describes the hot-swap controller as an intelligent power-entry point that can provide telemetry for predictive maintenance, capacity planning and energy forecasting. Solid-state transformers can perform similar functions at a much higher level in the system.
The result is a power architecture that begins to resemble the computing infrastructure it serves. Capacity can be observed, allocated and reconfigured. Failures can be detected and isolated. Batteries and onsite generation can be orchestrated through software rather than treated as independent backup systems.
The data center’s power train becomes programmable.
Cooling and Water Need Chips, Too
The semiconductor chain extends into thermal infrastructure.
Direct-to-chip liquid cooling requires coolant-distribution units, pumps, compressors, valves, fans, chillers and dry coolers. Those systems need motor-control ICs, power modules, embedded processors and flow, pressure, temperature and leak sensors.
Texas Instruments is already positioning its real-time microcontrollers, motor controllers and sensing products for coolant-distribution units and rack-level leak detection. Infineon’s data-center portfolio extends from grid conversion through cooling controls, including the chips governing pumps, compressors, fans and valves.
Water-reuse systems add conductivity, pH and chemical sensing, intelligent pump drives, membrane monitoring and industrial connectivity. The semiconductor content is smaller than it is in the grid-to-GPU power train, but the direction is similar: AI data centers are pushing cooling and water equipment toward continuous measurement and automated control.
Even power-semiconductor packaging is adapting to cooling. Onsemi has highlighted top-side-cooled SiC packages that can connect more naturally to the cold plates used in liquid-cooled systems. Power conversion and thermal management are beginning to be engineered together.
Silicon Photonics Attacks the Demand Side
Semiconductors can also reduce the power problem by cutting the energy required to move data.
As AI clusters grow, networking consumes a larger portion of the facility’s power. Conventional pluggable optical modules contain power-hungry digital signal processors and require electrical signals to travel from the switch chip to the edge of the system.
Co-packaged optics places silicon-photonics components beside the switch ASIC. Nvidia claims its design can provide 3.5 times better power efficiency than traditional pluggable optics. Broadcom reports more than 30% system-level power savings for one production design and substantially larger savings within the optical interconnect itself.
Those are vendor claims and will depend on configuration and deployment. They nevertheless illustrate the broader semiconductor response. Power devices improve how electricity reaches the processor, while silicon photonics reduces the electricity required to keep processors connected.
One Percent Is No Longer Small
A one-percentage-point improvement in a conventional server power supply may appear incremental. The same improvement across a 1-gigawatt AI campus frees 10 megawatts continuously.
That equals approximately 87.6 gigawatt-hours of electricity annually. A 5% end-to-end improvement frees 50 megawatts before accounting for the additional cooling power avoided because less electricity is lost as heat.
At that scale, efficiency is capacity. Electricity saved in conversion can be directed into additional accelerators and additional tokens. It may also reduce the number of transformers, generators, cooling units and utility upgrades required to support the campus.
Those economics explain why hyperscalers are willing to reconsider electrical architectures that remained largely unchanged for decades.
They also explain why the semiconductor opportunity surrounding AI is becoming much larger than the market for AI processors alone.
The Second AI Semiconductor Market
The AI infrastructure boom began with GPUs, accelerators, CPUs, HBM and networking silicon. All remain essential. Beneath that visible compute market, however, another semiconductor market is taking shape.
Silicon carbide is moving into transformers and utility connections. Gallium nitride is moving into high-frequency rack conversion. Power-management silicon is moving underneath the processor. Analog components and microcontrollers are moving into batteries, cooling plants and water systems. Silicon photonics is reducing the power required to move data across the AI factory.
Better semiconductors cannot manufacture electricity. They cannot eliminate transmission constraints or settle who should pay for grid expansion. A more intelligent power system can still be connected to a natural-gas plant, and an efficient data center can still place significant demands on the community around it.
What semiconductors can do is make the power already available more controllable, measurable and productive. They can reduce the share lost during conversion, protect the grid from volatile loads and allow more compute to operate within a fixed power budget.
The electric grid and the data center are beginning to look less like separate systems joined by a utility meter. From grid to GPU, they are becoming one semiconductor-controlled architecture.
