Understanding the Next Big Military Race in Power Semiconductors Under the AI Data Center Cycle, Beyond GPUs, It's Power

AI data centers are getting larger and larger, with energy consumption of a single data center often comparable to a medium-sized city.
In the past, data centers used 10-20kW per rack, now it's 80kW, 120kW, or even 600kW per rack. Power consumption of large AI clusters has entered the GW level.
The bottlenecks are no longer just GPUs, CPUs, and storage, but also current, heat, power distribution, copper losses, power conversion efficiency, grid access, and HVDC.
AI data center industry chain:
Power grid → Transformer → UPS → HVDC → PSU → VRM → GPU.
Traditional servers mainly used 48V because, during the internet era, rack power was not high. But in the AI era, the issues with low-voltage systems are becoming fully exposed. Because:
P = VI
For the same 1MW power, 48V requires over 20,000A of current, 400V about 2,500A, and 800V further drops to around 1,250A.
Lower current means thinner copper cables, reduced copper losses, less heat, smaller busbars, less stress on PSUs, less liquid cooling pressure, easier construction, and lower costs.
800V is a high-voltage platform validated by electric vehicles. Why do EVs adopt 800V? Because fast charging, high power, reduced line loss, and decreased heat loss.
Today, AI data centers face the same issues. As a result, SiC, high-voltage MOSFETs, high-voltage DC/DC converters, high-voltage PSUs, HVDC, busbars, and solid-state transformers—originally associated with new energy vehicles—are beginning to spill over into AIDC.
But 800V might just be the beginning; the real trend is HVDC (high-voltage direct current).
This is why traditional industrial power companies are suddenly being revalued by the market. Companies like Vertiv, Eaton, Schneider Electric, ABB, Siemens are becoming key beneficiaries in the AI industry chain.
This is also why power semiconductors are being re-evaluated by the market.
Infineon is a typical example of a company seamlessly transitioning from automotive power semiconductors to power infrastructure semiconductors.
Infineon may be one of the few global companies truly capable of “Grid-to-Core” power semiconductor platforms. Covering everything from high-voltage grid, HVDC, PSU, GPU power supply, high-frequency GaN, drivers, controllers, MCUs, to power modules, MOSFETs, SiC—all-encompassing.
This is also its biggest moat.
More importantly, Infineon is not fabless but an IDM. It designs, manufactures, packages, and tests its own products. This is extremely important in the power semiconductor industry because power semiconductors differ from CPUs/GPUs. Logic chips compete on EUV, FinFET, GAA, and transistor density. Power semiconductors truly compete on thermal management, high-voltage stability, long-term reliability, materials, packaging, epitaxy, and yield. Especially since future AI data centers will operate under long-term full load, high current, high heat density, and high voltage. Manufacturing itself is a technology.
Infineon’s current key assets include Villach, Dresden, Kulim. The most critical are the 300mm power fab and 200mm SiC. The market underestimates one point: 300mm power semiconductors are actually very difficult. Because of thermal stress, yield, high-voltage device challenges, and defect control, they are far more complex than ordinary mature processes.
In the AI era, the demand for power devices is beginning to expand on a large scale. The ability to manufacture advanced power semiconductors is starting to become a new moat.
If we only look at the “purest” AI high-voltage power players, companies like Navitas Semiconductor and Wolfspeed stand out. Especially Navitas, which is essentially pure Beta for GaN + high-efficiency AI power.
Wolfspeed represents another logic. If AI data centers fully adopt SiC PSUs, HVDC, and high-voltage power architectures, they could usher in a second growth curve.
There are also large industrial power platforms, such as Eaton, Schneider Electric, and ABB. Because they control power distribution, medium voltage, low voltage, circuit breakers, power management, and data center power topology. These switching costs are extremely high. Ultimately, AI will find that GPUs can be replaced, but once the power architecture is set, its lifecycle is very long.
In summary, whoever can continuously solve the challenges of current, heat, efficiency, power distribution, reliability, and grid access in the era of ultra-high power density AI will likely stay ahead in this race. Because the next bottleneck for AI has shifted from GPUs to Power. And this industry chain has not yet been fully priced in by the market.
Disclaimer: I hold assets mentioned in this article. The views are biased and do not constitute investment advice. DYOR.