As AI Pushes Data Centres to Breaking Point, Chinese Chipmakers Bet on Silicon Carbide

China's semiconductor industry is turning to silicon carbide (SiC) as artificial intelligence pushes data center power systems toward their physical limits, with a cohort of domestic chipmakers betting the wide-bandgap material becomes essential infrastructure for the AI era.

The Power Wall

Modern AI racks are power-hungry in ways traditional data center design never anticipated. Where a conventional server rack drew 10 to 15 kilowatts, the latest GPU racks demand 120 kilowatts or more — and next-generation systems are headed toward megawatt-class racks. At those densities, conventional silicon power devices lose too much energy as heat, and operators are moving to higher-voltage direct-current architectures where SiC excels.

Silicon carbide handles higher voltages, temperatures and switching frequencies than silicon, cutting conversion losses meaningfully at each stage of the power chain. Those percentage points matter enormously when a single AI campus draws hundreds of megawatts.

China's Angle

Chinese chipmakers spent the past five years building SiC capacity for the electric vehicle boom, only to face brutal price competition as capacity outran demand. AI data centers now offer a second act. Companies across the supply chain — from substrate producers to device makers — are repositioning toward power devices for AI infrastructure, an application where China's cost advantages and domestic demand give local firms a protected runway.

The bet aligns with Beijing's broader push for self-reliance in every layer of AI infrastructure that does not depend on restricted lithography. Power semiconductors, unlike leading-edge logic, can be manufactured entirely on domestic equipment.

The Stakes

If megawatt racks become standard, the power electronics inside data centers become a strategic market in their own right — one where Chinese suppliers could hold the kind of position in AI infrastructure that US export controls were designed to prevent in compute. The chips that feed the GPUs, it turns out, may be as contested as the GPUs themselves.