Over the past few years, the semiconductor industry has focused heavily on GPUs, HBM, advanced packaging, and advanced process nodes. Together, they form the core narrative of AI computing power. However, as AI servers evolve from standalone units to racks, from racks to data centers, and from data centers to "AI factories," another more fundamental layer is becoming increasingly critical: power devices.
Power devices are not chips responsible for computation; rather, they are chips dedicated to power conversion, control, and distribution. Their mission is straightforward: deliver electricity from the grid, batteries, or power systems to GPUs, CPUs, motors, inverters, and various load ends with the highest possible efficiency, minimal loss, and maximum stability.
In the past, when the market mentioned power semiconductors, it often thought first of NEVs, PV inverters, industrial motors, and EV chargers. However, entering the AI era, the industrial logic of power devices is shifting. It is no longer just a supporting industry for electric vehicles and new energy, but a crucial component of AI infrastructure.
To summarize in one sentence: behind computing power lies electrical power, and behind electrical power lie power semiconductors.
I. What Are Power Devices?
Power devices, also known as power semiconductors, are primarily used for switching, rectifying, inverting, transforming, protecting, and regulating electrical power. While ordinary digital chips process information, power devices process energy.
Common power devices include MOSFETs, IGBTs, diodes, thyristors, power modules, driver ICs, and the new generation of wide bandgap power devices manufactured using SiC (Silicon Carbide) and GaN (Gallium Nitride) materials.
In terms of applications, power devices are ubiquitous. They are required for fast charging in smartphones, power adapters for laptops, main inverters in NEVs, PV and energy storage inverters, motor drives in factories, and power systems in AI servers.
Their value lies not in "computing fast," but in "converting efficiently, controlling stably, and withstanding high voltage and high current."
To evaluate the capabilities of power devices, several metrics are typically considered: voltage rating, on-resistance, switching loss, thermal resistance, switching frequency, reliability, package thermal dissipation, and system cost. The weighting of these metrics varies across different applications. Consumer fast charging prioritizes high frequency and miniaturization; NEVs emphasize high voltage, high current, and automotive-grade reliability; while AI data centers simultaneously demand high efficiency, high power density, stability, and scalable supply.
II. Why Have Power Devices Suddenly Become More Important?
The most direct reason is that AI servers are pushing power systems to new boundaries.
The power density of traditional servers is relatively controllable, and their power architectures are quite mature. However, AI servers are entirely different. With GPUs, AI ASICs, high-speed switching chips, and liquid cooling systems concentrated in a single rack, rack power is rapidly evolving from the previous tens of kilowatts to hundreds of kilowatts or even higher.
The higher the power, the greater the current, the more severe the losses, and the harder it is to control heat generation. If power systems continue to rely on the traditional low-voltage, high-current architecture, they will encounter issues such as thick cables, excessive copper usage, multiple conversion stages, high thermal losses, and large spatial footprints.
This is why the 800V HVDC (High Voltage Direct Current) architecture is beginning to attract attention.
Traditional data center power chains often require multiple stages of AC/DC and DC/DC conversion. Each conversion incurs losses and heat. The concept behind 800V HVDC is to increase the distribution voltage, reduce the current, and minimize conversion stages, thereby lowering copper losses, reducing cable volume, improving system efficiency, and supporting AI racks with higher power densities.
In this process, power devices have transitioned from background components to foreground variables. Without high-efficiency, high-reliability, and high-power-density power devices, it will be difficult for AI servers to continue increasing rack power or keep data center energy consumption within reasonable limits.
Therefore, AI is driving not only GPUs and HBM but also PMICs (Power Management ICs), MOSFETs, SiC, GaN, driver ICs, power modules, solid-state transformers, solid-state circuit breakers, and energy storage inverters.
III. Silicon, SiC, and GaN: Not About Replacement, But Division of Labor
The most common misinterpretation of power devices is portraying silicon, SiC, and GaN in a simple substitution relationship. In reality, they are more like a toolbox distributed across different voltage, power, and frequency ranges.
1. Silicon-Based MOSFETs: Mature, Cost-Effective, and Still Irreplaceable
Silicon-based MOSFETs are among the most mature and widely used power devices, particularly suitable for low-voltage, medium-low voltage, and high-volume shipment scenarios. Server motherboards, consumer electronics, power adapters, low-voltage DC/DC conversion, and power supply modules near GPUs still heavily utilize silicon-based MOSFETs.
Their advantages include low cost, mature supply chains, stable processes, and thorough customer validation. Even with the rapid development of SiC and GaN, silicon-based devices will remain for the long term, especially in low-voltage, high-current, and cost-sensitive scenarios.
2. IGBTs: The Veteran Workhorse in Industrial and NEV Applications
IGBTs combine the easy-drive characteristics of MOSFETs with the high-current capability of bipolar devices, and have long been applied in NEV main inverters, industrial motors, rail transit, wind power, PV, and high-voltage frequency conversion fields.
Their advantages lie in mature reliability, high power ratings, and strong modular capabilities. However, driven by demands for higher frequency, higher efficiency, and higher power density, SiC is gradually entering some high-end scenarios previously dominated by IGBTs.
3. SiC: The Key Material for High-Voltage, High-Temperature, and High-Power Scenarios
SiC, or Silicon Carbide, belongs to wide bandgap semiconductors. Compared to silicon, it offers higher breakdown field strength, better high-temperature performance, and lower loss potential, making it suitable for high-voltage, high-power, and high-reliability applications.
In NEVs, SiC is primarily used in main inverters, OBCs (On-Board Chargers), DC/DC converters, and high-voltage platforms. Adopting SiC in 800V EVs presents opportunities to improve efficiency, reduce losses, increase driving range, or shrink system size.
In AI data centers, SiC is more suited for the high-voltage end and high-power conversion stages, such as from the grid side to the high-voltage DC bus, solid-state transformers, energy storage inverters, and high-power power supply modules. Its value lies in withstanding high voltage, reducing losses, improving thermal dissipation, and enhancing system power density.
However, SiC is not cheap. Its challenges lie in upstream crystal growth, substrate defect control, epitaxy, device yield, module packaging, and reliability. SiC is not a smooth upgrade from ordinary silicon devices, but an upgrade of an entire material and process system.
4. GaN: High Frequency, High Power Density, and Miniaturization
GaN, or Gallium Nitride, is also a wide bandgap material. Compared to SiC, GaN's prominent advantages lie in high-frequency switching, high power density, and miniaturization, making it particularly suitable for medium-low to medium-high voltage and high-frequency conversion scenarios.
Fast charging was the first application to familiarize the public with GaN. GaN fast chargers are compact, highly efficient, and generate less heat precisely because they can operate at higher frequencies, thereby reducing the size of magnetic components and the overall system.
In AI data centers, GaN opportunities mainly lie in server PSUs, 48V/54V power supplies, DC/DC conversion, in-rack step-down conversion, and high-frequency power modules. It may not handle the highest-voltage grid-side stages, but it holds advantages in high-frequency conversion inside the rack and close to the load.
Therefore, SiC and GaN are not in a zero-sum competition. More accurately, SiC leans towards high voltage and high power, GaN favors high frequency and high power density, while silicon-based devices continue to play a foundational role in low-voltage, large-scale scenarios. The future power chain in AI data centers will most likely see silicon, SiC, and GaN coexisting, each handling the conversion stage best suited to its strengths.
IV. How Do AI Data Centers Change the Value of Power Devices?
Past growth in power devices primarily stemmed from NEVs and new energy power generation. NEVs increased the per-vehicle value of power semiconductors, PV and energy storage drove demand for inverters and modules, and industrial automation brought about motor drive requirements.
What differentiates AI data centers is that they are transforming the entire power chain "from the grid to the chip."
The power system of an AI data center can be divided into four layers.
The first layer is the grid side. As AC power from the grid enters the data center, it requires distribution, protection, rectification, and conversion. This involves high-voltage power devices, solid-state transformers, solid-state circuit breakers, and industrial-grade power equipment.
The second layer is the data center distribution side. To reduce losses, higher-voltage DC distribution will become a crucial direction in the future. This requires support from high-voltage-tolerant and highly reliable power devices.
The third layer is the rack and server power supply. As the power density of AI racks increases, the power ratings of PSUs, power modules, and intermediate bus converters rise, demanding higher efficiency and smaller form factors. This is a critical area where both SiC and GaN play their roles.
The fourth layer is the point-of-load (PoL) power supply near the chips. Ultimately, GPUs require low-voltage, high-current power supply close to 1V. The closer to the chip, the lower the voltage and the higher the current, which places higher demands on the system-level synergy of MOSFETs, drivers, packaging, thermal dissipation, inductors, and capacitors.
This means that AI servers are not simply buying a few more power devices, but are driving an upgrade of the entire power architecture. The value of power devices will increase alongside the rise in rack power, conversion stages, efficiency requirements, and reliability demands.
V. Why Is the 800V Architecture Important?
The 800V HVDC architecture is important because it solves the physical challenges of high-power AI racks.
Electrical power equals voltage multiplied by current. Given a constant power, increasing the voltage can reduce the current. Once the current drops, cable heating, copper losses, and copper material usage all decrease. For high-power racks, this is not a minor detail but a fundamental issue determining whether the system can continue to scale.
The traditional low-voltage, high-current architecture works for racks in the tens of kilowatts, but when rack power moves towards hundreds of kilowatts or higher, cables, copper busbars, thermal dissipation, space, and conversion efficiency all become bottlenecks. The 800V architecture delivers power to racks and servers in a more efficient and compact manner by increasing the distribution voltage and reducing conversion stages.
This also explains why SiC, GaN, solid-state transformers, and solid-state circuit breakers are placed in the same narrative. 800V is not a single-device innovation but a system architecture innovation. It requires the collaborative effort of high-voltage devices, high-frequency devices, driver ICs, isolation, protection, packaging, thermal management, and complete machine power supply manufacturers.
For power semiconductor companies, the real opportunity lies not just in selling individual MOSFETs or diodes, but in entering the system design chain from the grid to the rack, from the rack to the server, and from the server to the chip.
VI. Who Benefits in the Industry Chain?
The power device industry chain can be roughly divided into five layers.
The first layer is upstream materials. This includes SiC substrates, epitaxial wafers, GaN epitaxy, silicon wafers, metal materials, packaging materials, and ceramic substrates. The upstream barriers for SiC are particularly high, with slow crystal growth, difficult defect control, and a long cycle for yield improvement. GaN-on-Si, on the other hand, relies more on large-size silicon-based substrates and epitaxial process control.
The second layer is equipment and processes. SiC crystal growth furnaces, epitaxial equipment, ion implantation, high-temperature annealing, etching, thin-film deposition, and testing equipment all determine the speed of capacity release and yield. Unlike digital chips that solely focus on advanced process nodes, power devices have stringent requirements for material defects, thermal stress, and reliability.
The third layer is device design and manufacturing. This includes silicon MOSFETs, IGBTs, SiC MOSFETs, GaN HEMTs, diodes, and driver ICs. Leading companies typically possess long-term process accumulation and customer certification data.
The fourth layer is modules and packaging. In high-power scenarios, packaging is not simply "wrapping the chip," but a critical step determining thermal conductivity, inductance, reliability, and system lifespan. Power modules, double-sided cooling, silver sintering, ceramic substrates, and low-parasitic packaging all impact final performance.
The fifth layer is system manufacturers. This includes server power supplies, data center power equipment, vehicle electronic control, PV and energy storage inverters, industrial power supplies, and EV charger manufacturers. In the future, power device manufacturers will find it difficult to win high-end clients relying solely on discrete devices if they fail to understand system architectures.
VII. Where Are the Opportunities for China's Power Devices?
China is one of the world's largest markets for NEVs, PV, energy storage, consumer electronics, and industrial power supplies, making it also one of the most concentrated markets for power device demand. The opportunities for domestic power devices genuinely exist, but they must be viewed in layers.
First, the domestic substitution of silicon-based power devices has already established a certain foundation. For medium-low voltage MOSFETs, diodes, and some IGBT and module products, domestic manufacturers possess strong supply capabilities and are continuously integrating them into consumer, industrial control, new energy, and some automotive applications.
Second, SiC is the most watched direction for domestic substitution. China's massive demand in the NEV and PV energy storage markets provides local customer scenarios for SiC substrates, epitaxy, devices, and modules. However, the true barriers for SiC lie in high-quality substrates, device reliability, automotive-grade certification, and scaled yield, rather than mere capacity expansion.
Third, GaN holds potential for cost-effectiveness and scaling. GaN-on-Si can leverage parts of the silicon industry chain foundation, making it suitable for high-frequency, miniaturized, and high-power-density scenarios. Chinese companies have opportunities in consumer fast charging, data center power supplies, and some industrial power supplies, but high-end applications still require breakthroughs in reliability, patents, and customer certification.
Fourth, equipment and materials deserve attention. Many industrial opportunities lie not in the most conspicuous device segment, but in SiC crystal growth, epitaxy, packaging materials, ceramic substrates, testing equipment, and reliability verification platforms. The higher up the value chain, the more the long-term value of materials and processes is demonstrated.
However, risks must also be acknowledged: China's power device industry is prone to the structural contradiction of "rapid capacity expansion at the low end and slow breakthroughs at the high end." Once mid-to-low-end capacity becomes excessive, price wars will compress profits; meanwhile, high-end SiC/GaN requires long-term certification and reliability data, making it difficult to fully replace overseas leaders in the short term.
VIII. Power Devices Are Not Just About a Simple Price Hike Logic
Recently, the market has focused on price hikes for power devices, but the price increase is merely the result, not the root cause.
If the price hike stems from rising raw material costs, its sustainability may be limited. If it results from short-term channel inventory replenishment, it may also drop quickly. Price hikes truly worth attention are driven jointly by product structure upgrades, increased capacity utilization rate, locked orders from high-end customers, and supply bottlenecks.
Power devices are essentially still a cyclical industry. The slowing growth of NEVs, price wars in PV and energy storage, weak industrial demand, and fluctuations in AI capital expenditures could all affect the pace of demand. Even with strong demand from AI data centers, it does not mean all power device manufacturers will benefit.
To determine whether a power device company is truly benefiting, four questions need to be examined.
First, have its products entered high-value scenarios, such as AI server power supplies, automotive-grade SiC, energy storage inverters, and high-end industrial power supplies?
Second, is its capacity effective capacity, rather than nominal capacity with low yield, low certification, and low gross margins?
Third, does it master key processes and packaging capabilities, rather than merely engaging in low-end packaging or trade-based supply?
Fourth, can the price hike be transmitted to gross margins, cash flow, and long-term orders, rather than merely stopping at price hike notices and market rumors?
IX. Three Main Themes for the Future
The future of the power device industry can be observed along three main themes.
The first theme is the upgrade of AI power architectures. 800V HVDC, high-power racks, liquid cooling, energy storage support, and near-chip power supply will jointly drive the increase in the value of power devices. The focus here is not on discrete devices, but on system-level power solutions from the grid to the GPU.
The second theme is NEVs and energy infrastructure. The 800V platform for EVs, EV chargers, PV inverters, energy storage converters, industrial motors, and smart grids remain crucial markets for SiC and IGBTs. AI is a new variable, but traditional electrification has not disappeared.
The third theme is domestic substitution and industry chain security. SiC substrates, epitaxy, GaN devices, power modules, packaging materials, and testing equipment all have room for domestic substitution. However, the evaluation criteria must upgrade from "can it be done" to "can it be mass-produced stably, pass certifications, and generate profits."
X. Conclusion: Power Devices Are the Energy Efficiency Engine of AI Infrastructure
In the semiconductor narrative of the AI era, the focus cannot be solely on computing chips. GPUs handle computation, HBM handles bandwidth, advanced packaging handles interconnection, and power devices are responsible for delivering electricity efficiently, stably, and reliably to these chips and systems.
As the power density of AI servers steadily rises and the electricity pressure on data centers keeps increasing, power systems are no longer just background support, but a critical link determining the efficiency, cost, and scalability of AI infrastructure.
The core value of power devices is shifting from "discrete device cost" to "system energy efficiency value." SiC, GaN, and high-performance silicon-based devices will not simply replace each other, but will collaborate with a division of labor across different voltage, power, and frequency levels. In the future, truly competitive enterprises will not only be able to manufacture devices but also understand system architectures, material processes, packaging and thermal dissipation, and customer certifications.
For tech readers, the most important thing in understanding power devices is to grasp one sentence: the end of AI is not just about computing power, but also electrical power; and power semiconductors are precisely the critical bridge connecting electrical power and computing power.
If this article has been helpful to you, please consider tipping the blogger at your discretion. Blogger's contact (add WeChat: tigerchip)