Independent analysis of data centre power architecture, autonomous operations, and the infrastructure behind AI. By Adi Kumar.

AI power semiconductors

TOPIC HUB · AI POWER SEMICONDUCTORS

AI power semiconductors

The reference on the semiconductor devices at the point-of-load and along the AI power delivery chain: silicon carbide (SiC) and gallium nitride (GaN) wide-bandgap devices, direct-to-pin factorised power, and the vendor landscape from utility-scale SST silicon down to sub-1V GPU point-of-load.

01Direct answer

AI power semiconductors covers the silicon devices that convert power at each layer of the AI compute stack, from utility-scale silicon carbide (SiC) devices inside solid-state transformers, to gallium nitride (GaN) devices at the rack-level 800V-to-48V conversion, to point-of-load (POL) silicon delivering sub-1V power directly to GPU packages. Wide-bandgap materials (SiC, GaN) enable higher switching frequencies, higher efficiency, and higher operating temperature vs traditional silicon. Vendor landscape spans specialists (Wolfspeed for SiC, Navitas + EPC for GaN) plus broad-portfolio incumbents (Infineon, ADI, TI, MPS, ROHM, STMicro, Renesas, onsemi, Power Integrations). Vicor's direct-to-pin factorised power architecture is a distinct point-of-load approach that competes with the 48V-intermediate bus.

02Key concepts

SiC
Silicon carbide; wide-bandgap material for MV + HV power devices. Wolfspeed is the pure-play reference.
GaN
Gallium nitride; wide-bandgap for higher-frequency LV switching. Navitas + EPC lead the pure-plays.
Point-of-load (POL)
Sub-1V conversion at the GPU package; specialist silicon from Vicor, MPS, ADI, TI et al.
Factorised power (Vicor)
Direct-to-pin architecture that skips the intermediate 48V bus.
MOSFET / HEMT
Metal-oxide-semiconductor field-effect transistor / high-electron-mobility transistor.
Substrate + epi
The base wafer + epitaxial layer; upstream supply constraint for SiC (Wolfspeed + Coherent are dominant).
Package
The physical form + thermal interface; often more constraint than the die itself for high-current.
Wide-bandgap
Materials (SiC, GaN, Ga2O3, diamond) with bandgap wider than silicon; enable higher voltage + temperature.

03Structural decomposition

The AI power semiconductor stack (by layer)

  1. Layer 1-2 (MV). SiC devices in utility-scale converters + solid-state transformers
  2. Layer 3 (LV distribution). SiC/GaN hybrids in DC-DC converters + power supplies
  3. Layer 4 (rack conversion). GaN + SiC hybrid for 800V-to-48V; Delta, Advanced Energy, Bel Power
  4. Layer 5 (in-rack). GaN devices in 48V busbar management + rack PDU
  5. Layer 6 (POL). Silicon POL + Vicor factorised power + Navitas direct-to-pin

04Deep-dive research

06Commercial + investment implications

Commercial + investment angles

SiC + GaN vendors are attractive PE acquisition targets given consolidation activity. Wolfspeed's restructuring is the reference cautionary case for pure-play capital intensity. Broad-portfolio incumbents (Infineon, TI, ADI, MPS) offer more diversified exposure. See Wolfspeed retrospective for the specific cautionary lens.

07Frequently asked

SiC vs GaN: what is the practical difference?

SiC handles higher voltages (up to MV) and higher power; used in solid-state transformers, EV traction, grid infrastructure. GaN switches faster at lower voltages (typically <=650V); used in rack-level DC-DC and POL. They are complementary, not competing, in most AI power applications.

Is Vicor's direct-to-pin approach going to displace the 48V intermediate bus?

Not fully. Direct-to-pin has real efficiency + density advantages for the highest-density AI accelerator racks. But the 48V intermediate bus is entrenched across the OCP ecosystem, has broader vendor support, and doesn't require the same silicon-integration effort. Expect both to coexist through the 800V DC transition.

Why did Wolfspeed restructure?

Capex + demand-timing mismatch. Wolfspeed built the Mohawk Valley $5B SiC fab ahead of demand ramp. Auto/EV demand softened + industrial ramp was slower than assumed. See tech spotlight 05 for full retrospective.

What is the biggest supplier-concentration risk at Layer 6?

Vicor for direct-to-pin architectures. MPS for rack POL. But both are backed by broader silicon portfolios, so a single vendor failure doesn't break the ecosystem.

How does Chinese silicon (Innoscience for GaN, etc.) fit?

Fast-growing domestic pure-plays with material cost advantages. Global availability constrained by US + adjacent-jurisdiction export controls + national-security review, but Chinese-domestic + SEA + Middle East + emerging-market presence is meaningful.

08Further reading

See also: 800V DC (application layer), M&A (semiconductor acquisitions). Full series: Tech Spotlights (15 essays).
Definition AI power semiconductors are the wide-bandgap devices that enable AI data centre power architectures above the traditional silicon envelope. Silicon carbide (SiC) MOSFETs occupy the 650V to 1700V range covering grid-tie, HVDC and 800V DC bus rectification. Gallium nitride (GaN) HEMTs occupy the 60V to 650V range covering point-of-load conversion at the accelerator package. Five vendors hold roughly 90% of merchant SiC revenue: STMicroelectronics, onsemi, Infineon, Wolfspeed and ROHM.