Weekend Industry Research — Analyzing the Market Share of SiC, GaN, and Silicon MOSFETs in AI-Driven Power Electronics Infrastructure

The rapid construction of AI data centers is driving a major upgrade of the power grid, bringing another long-underestimated field back into the spotlight: power semiconductors.
The core of the power system is efficiently controlling current. The most essential device for current control is the MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor).
Over the past few decades, almost all global power devices have been built on silicon MOSFETs. Silicon is cheap, mature, and has a complete supply chain, thus dominating the industry for a long time. But as AI server power demands surge, EVs enter the 800V era, data centers evolve toward higher voltage, and high-frequency power supply needs increase, traditional silicon begins to hit physical limits. Therefore, SiC (Silicon Carbide) and GaN (Gallium Nitride) are emerging.
SiC is more like heavy industry. Its core advantages are high voltage and high power. SiC has a higher breakdown voltage and better thermal conductivity, making it significantly more efficient than traditional silicon IGBTs in high voltage and high current scenarios. As a result, fields such as EV main drive inverters, photovoltaic inverters, energy storage, industrial high-voltage drives, power grids, and high-voltage UPS are rapidly adopting SiC. Especially Tesla’s push for the 800V platform, which is essentially a crucial turning point for the entire SiC industry explosion. In recent years, new energy vehicles have been the biggest driver for SiC. Companies like Wolfspeed, onsemi, STMicroelectronics, Infineon Technologies, ROHM, and Mitsubishi Electric have all benefited from this cycle.
But SiC is not perfect. Compared to GaN, it generally has slower switching speeds, higher Qg, weaker high-frequency performance, and magnetic devices are harder to miniaturize at high frequencies. So GaN has taken a different route. GaN’s real strength lies in high frequency. GaN has lower Qg, lower output capacitance, and almost no reverse recovery issues, making it especially suitable for high-frequency DC-DC, AI server power supplies, GPU VRMs, fast charging for mobile phones, high-frequency PSUs, and miniaturized power supplies.
AI might be the true long-term cycle for GaN. Because AI data centers are driving the entire power architecture toward higher frequency, higher current, smaller size, and higher efficiency. Especially after the 48V architecture, a large number of high-frequency DC-DC converters are becoming core bottlenecks, which is GaN’s sweet spot.
Traditional server racks may only have 5-10kW, but AI racks are already entering 50kW, 100kW, and potentially approaching MW levels in the future.
AI data centers are gradually transforming from IT facilities into “power facilities.” Between the grid and GPUs, there are many power conversions: high-voltage transmission, transformers, UPS, PSUs, AC/DC, DC/DC, VRMs, and close-to-load power supplies. Each conversion results in energy loss. When a single AI campus consumes gigawatt-level power, a 1% efficiency improvement can translate into enormous economic value. Thus, power semiconductors are shifting from supporting roles to core bottlenecks.
As a result, GaN is increasingly used in AI server PSUs, high-frequency DC/DC, GPU VRMs, and power modules. Many systems even feature “SiC + GaN” hybrid solutions—high-voltage main trunks with SiC, high-frequency ends with GaN. In data centers, the high-power high-voltage parts from the grid to the data center are more suitable for SiC, while high-frequency power supplies inside servers are better suited for GaN.
In the future, the entire power semiconductor industry may form a three-layer structure: low voltage and low cost — silicon MOSFET; high frequency and high efficiency — GaN; high voltage and high power — SiC.
Around 650V is the area where GaN and SiC face direct competition. Below 650V, GaN has a clear advantage. Above 650V, SiC’s advantages grow stronger. Near 650V, both can be developed.
At the same time, many critical systems worldwide operate around 400V~800V DC bus voltages.
650V devices typically correspond to 400V AC rectification, 380V HVDC, upstream of 48V architectures, data center PSUs, industrial power supplies, photovoltaics, OBC, and AI server power supplies.
This is one of the most core voltage ranges in modern industry and data centers.
Therefore, competition shifts from purely device parameters to system costs, EMI, drive complexity, heat dissipation, yield, reliability, customer validation, lifespan, thermal cycling, ppm failure rates, and long-term supply capabilities.
This is also why the power semiconductor industry’s moat is so deep, especially for SiC. The real difficulty with SiC is not just device design but wafer growth, epitaxy, defect control, yield, and high-temperature reliability. These capabilities require long-term process accumulation. As a result, the truly dominant players are often companies with over ten years of experience. Different companies have different strengths: Wolfspeed excels in materials, STM in EV, Infineon in modules and system integration, onsemi in automotive clients, Rohm in reliability.
GaN, on the other hand, has not yet fully entered maturity. Currently, Texas Instruments, Navitas Semiconductor, Infineon Technologies, and Efficient Power Conversion are advancing GaN in different directions. TI may be underestimated in the long run because the most important factors for major clients are not just PPT parameters but reliability, qualification, and long-term supply, which are TI’s strengths.
Overall, AI is increasing the “power semiconductor content” in entire systems. The future competition in AI infrastructure may not only be about computing power but also about power supply, distribution, cooling, and power efficiency.
In the past, the core of the semiconductor industry was computing. Over the next decade, power control itself may become one of the new core bottlenecks.
Disclaimer: I hold assets mentioned in this article. Views are biased and do not constitute investment advice. Do your own research.