What runs the power, not the math?

Why silicon hits a ceiling

Silicon's 1.12 eV band gap was the goldilocks property that made it the substrate for computing. For switching high voltages, however, that same band gap is a limitation. Above about 600 volts, a silicon power transistor begins leaking current through its off-state in significant quantities, and at the temperatures that high-voltage switching produces, the leakage runs away thermally. Silicon power devices exist up to about 6 kV in specialized geometries, but the practical efficient ceiling for mainstream silicon power is closer to 1 kV.

Above that, two wide-band-gap materials take over. Silicon carbide (SiC) has a band gap of 3.26 eV — almost three times silicon. Gallium nitride (GaN) has a band gap of 3.4 eV. Both can hold off thousands of volts in the off-state, switch faster than silicon, and tolerate higher operating temperatures. They are also harder to grow, harder to dope, more expensive per wafer, and require different fab equipment. None of these were good enough trade-offs to displace silicon for sixty years; the trade-offs only flipped when the customer base — primarily EVs, and now AI data centers — needed the higher voltages enough to pay for the cost premium.

What changed between SiC and GaN

SiC won the high-voltage, high-current sweet spot: 600V to 6.5 kV, currents up to hundreds of amps, applications dominated by EV traction inverters and on-board chargers. The 2018 launch of the Tesla Model 3 with a STMicroelectronics SiC inverter was the inflection point. Within five years almost every premium EV used SiC in the inverter and on-board charger; Tesla switched all its vehicles by 2020, and BYD, Volkswagen, Hyundai, and BMW followed. The SiC wafer fab buildout that this triggered is still running ahead of demand.

GaN won the lower-voltage, higher-frequency niche. GaN-on-silicon devices switch at megahertz frequencies and handle a few hundred volts efficiently, which makes them the right device for fast chargers (the small 65-100 W laptop chargers that have replaced bricks), data-center power supplies, and emerging applications like vertical power delivery for chips. A 2025 GaN-based 240W laptop charger is roughly one-third the volume of the equivalent silicon brick from 2015. The fast-charger displacement is essentially complete in the high-end segment.

The split is not always clean. SiC and GaN compete at the boundary — roughly 400-600V — and the choice often comes down to who has supply and who has the design relationship. For AI data centers operating at 800V DC bus architectures (a transition currently underway from 48V), the device choice is contested between SiC and GaN, and the eventual answer will determine the structure of the next decade of data-center power.

The supply situation

The SiC supply chain has five major players. Wolfspeed (formerly Cree) is the largest pure-play SiC company, growing wafers and selling devices, with the most aggressive capacity build in the US (Mohawk Valley, NY). Infineon is the largest device-side player in Europe, has invested heavily in its own SiC wafer capability, and dominates automotive customer share. STMicroelectronics is the second European supplier and held the Tesla design win for years. ON Semiconductor has the third-largest automotive SiC share, partly via the GTAT acquisition that gave it captive wafer supply. Rohm in Japan is fourth. Several Chinese suppliers — Sanan IC, Yangzhou Yangjie, BYD's internal capability — are scaling but starting from a small base.

GaN supply is more concentrated and earlier-stage. Power Integrations has a dominant share in fast-charger applications. Infineon, GaN Systems (acquired by Infineon in 2023), Navitas, EPC, and Texas Instruments are the main device suppliers. TSMC and several other foundries offer GaN-on-silicon process services; the materials are typically grown on 6-inch or 8-inch silicon substrates, which is cheaper than the 6-inch SiC substrates required for SiC.

The supply constraint is real. Wolfspeed reported in 2024 that its capacity was sold out through 2026. Infineon's SiC wafer revenue grew over 50% year-over-year. Lead times on SiC modules for EV inverters were 12-18 months for most of 2023-24, easing only modestly in 2025. The shortage has cost premium EV makers reportedly $1-2 billion of foregone revenue over the period.

Source: Wolfspeed, Infineon, STMicro, ON Semiconductor FY2023-24 earnings disclosures; Yole Développement SiC market reports.

AI-data-center implications

An AI data center pulls power through a stack of conversion stages — utility 230 kV down to 35 kV down to 480 V AC down to 48 V DC down to 0.8 V at the GPU pin. Each conversion stage has a power-semiconductor device at its core (this is described in detail in electrons to tokens). The total semiconductor content for power conversion in a 100 MW data center is in the high tens of millions of dollars — comparable to the analog content, smaller than the GPU content but still strategic.

The current architectural transition is from 48 V DC bus distribution to 400 V or 800 V DC bus distribution. The motivation is the same as in EVs: at higher voltage, the wires are smaller, the conversion losses are lower, and the equipment fits more density into the same rack volume. The technical bottleneck is the device supply at these voltages, which is exactly where SiC and GaN are scaling. Whoever wins the AI data-center DC-bus standardization (currently a fight between Open Compute Project specs, NVIDIA's preferred topology, and individual hyperscaler customs) will lock in a multi-decade device-supply preference.

Strategic read

Power semiconductors are the part of the AI buildout that is hardest to substitute. You cannot run a frontier AI cluster without converting hundreds of megawatts from grid to chip, and that conversion has to go through SiC or GaN somewhere in the stack. The supplier list is short. The capacity ramps are visible and largely fixed for the next 3-5 years. China is investing aggressively to close the gap, but the wafer-growth equipment for SiC (particularly the long-time-constant boules that single-crystal SiC requires) is itself a chokepoint, and the leading equipment suppliers (II-VI now Coherent, others) are concentrated in the US and Europe.

For an investor, the AI-infrastructure exposure here is more durable than the GPU exposure. NVIDIA can be substituted by AMD, by hyperscaler custom silicon, by future Chinese GPUs — the substitution risk is real even if it is slow. SiC suppliers have no equivalent substitution path. If you cannot get SiC modules, the data center does not energize. The combined SiC + GaN device market is on track to roughly triple over 2023-2028 to about $20 billion, and the named suppliers will capture most of that growth.