SiC And GaN Suppliers Are The Next AI Rack Bottleneck

SiC And GaN Suppliers Are The Next AI Rack Bottleneck

Direct:

NVIDIA's (NVDA) shift to 800V HVDC at the rack level moves the AI bottleneck off the GPU. It lands on the power semiconductor side, where SiC and GaN suppliers are already tight. The Kyber rack platform that pairs with Rubin Ultra targets 800V end to end, reducing or bypassing the legacy 54V intermediate bus and forcing a new conversion stack from substation to processor.

Infineon (IFNNY) guides $12,000 to $15,000 of power semi content per current 130kW AI rack. onsemi (ON) puts content per next-generation 1MW rack near $100,000, roughly double its prior figure. TrendForce projects SiC and GaN penetration in data center power systems reaches 17% in 2026 and clears 30% by 2030, alongside +20% YoY AI server shipments and roughly 47% liquid cooling penetration.

Each named partner occupies a different conversion stage. STMicroelectronics (STM) demonstrated an 800V GaN LLC converter at 1MHz running above 98% efficient. Texas Instruments (TXN) showed an 800V-to-6V DC/DC with integrated GaN at 97.6% peak and over 2,000 W/in³. Navitas Semiconductor (NVTS) is the named GaN and SiC partner on Kyber feeding Rubin Ultra. Wolfspeed (WOLF) posted around 30% sequential growth in AI data-center revenue in Q3 FY26 and launched a 10kV SiC MOSFET aimed at the front-end solid-state transformer slot.

SiC wafer capacity stays tight. Infineon, onsemi, ST and Wolfspeed are signaling allocation-constrained high-voltage portfolios, with announced fab expansions still ramping. Allocation outpaces design wins through 2026 to 2028. Taiwanese power and thermal vendors (AcBel Polytech, YS Tech, Sunon, Song Chuan, Teco) own the downstream busbar, cold plate and CDU buildout, while the binding constraint is upstream wafers and power devices.

For consumer SSD readers this matters indirectly. The hyperscaler capex cycle that pulled foundry and packaging capacity toward HBM and CoWoS now extends to power semis, and that environment keeps controller pricing and lead times tighter than they would otherwise be through 2026.

Drafted with AI assistance against parallel reporting.

Sources:

  • NVIDIA, “800 VDC Architecture for AI Data Centers” (2026)
  • TrendForce, “2026 Outlook: The Era of GW-Level AI Data Centers” (Dec 2025)
  • TrendForce press, “AI to Reshape the Global Technology Landscape in 2026” (Nov 2025)
  • Wolfspeed Q3 FY2026 8-K
  • STMicroelectronics newsroom, 800 VDC portfolio expansion with NVIDIA (2026)
  • Texas Instruments newsroom, 800 VDC architecture with NVIDIA (Mar 2026)
  • onsemi press, 800 VDC collaboration with NVIDIA (2025)
  • Navitas / NVIDIA collaboration SEC filing (May 2025)
  • DigiTimes, “Demand surge for power semiconductors reshapes data-center power and cooling” (May 2026)

Reddit: https://ift.tt/m4KfqSB

Scroll to Top