Onsemi's GaN window makes a power semiconductor the bottleneck behind the meter

Antoine Jalabert runs the gallium nitride division at Onsemi and has spent years making a case the data center industry has not yet acted on, placing GaN's server moment at 2026/27, when he expects the material to ramp into mass production, and framing the timing as necessity rather than upgrade. “When you look at Nvidia's roadmap,” he told Data Center Dynamics, “there's no other way than to use GaN that will allow you to reach those kinds of power densities.”

GaN is a wide-bandgap compound of gallium and nitrogen that tolerates higher voltages and temperatures and whose electrons move faster; in power conversion the practical effect is heat that never gets made, because GaN loses far less energy to waste than silicon or silicon carbide when it switches, letting a device run at higher frequency. Jalabert's point is that a switch-mode power supply running faster can be built smaller, and for a rack designed against a power-density target, the conversion stage is the part that has to give.

The material itself is old news. Scientists have had GaN in hand since the early 1930s, though the first GaN transistors did not arrive until the mid-1990s. What followed was not the server ramp the industry forecast but chargers — phone adapters and electric-vehicle adapters — which Jalabert describes as the one place nobody was looking. “The world has been expecting GaN to ramp into the server and telecom space since we started talking about this technology around 10 to 15 years ago,” he said. “Why? Because that is where you really need power density. And did it? No.”

Onsemi, founded in 1999 as a spinoff of Motorola's Semiconductor Components Group and based in Scottsdale, Arizona, is one of a modest number of companies currently offering GaN-based power semiconductors, and in October 2025 it unveiled a vertical gallium nitride offering. Vertical GaN differs from the lateral variety, which is grown on silicon or sapphire substrates, by conducting electricity vertically through a 3D structure rather than across a plane.

Nvidia's own behavior is the better test of Jalabert's claim. The chipmaker has been buying on the generation side of the meter — a $3 billion SB Energy investment under consideration behind an OpenAI campus in Ohio, a stake in Cloverleaf — because the buildout's constraint has moved off land and onto electrons. The firm investing at that layer is also the firm whose roadmap decides how many electrons a rack needs, and the argument from the supply side is that at the densities Nvidia is targeting, the conversion stage has to change material.

The AI buildout now has two power bottlenecks, and the industry has spent two years pricing only one of them. The grid side is well understood: interconnection queues, load-class rules, and generation contracts set the schedule, and grid permission is the asset that prices before electrons do. The other bottleneck sits inside the enclosure, on the far side of the meter, where a rack's ceiling is a materials question, and Jalabert is describing it as the constraint Nvidia's roadmap runs into.

That also locates GaN inside the delivery constraint this buildout now organizes itself around, which runs through supplier slots and raw materials rather than through labor alone. A power semiconductor is one of those slots: a capacity commitment made years before the first rack draws current, with a qualification cycle on one side and a fab on the other. Naming a 2026/27 mass-production window is a capacity claim, and the distance between a claim and the capacity itself is where a buildout's schedule lives.

The parallel worth drawing is with Schneider Electric's factory-built power train, which compressed the most controllable part of the data center timeline and left the interconnection queue exactly where it was, because the queue and not the switchgear sets the pace. GaN is a smaller version of the same move in a different place, relieving a constraint behind the meter, where engineering can still buy time, and doing nothing about the one in front of it, which means the rack can get denser without the queue moving at all.

On Jalabert's account, switching losses already settle the efficiency question; the open question is whether the 2026/27 window holds. Onsemi's vertical offering is less than a year old. If the ramp arrives on schedule, the power semiconductor would join the grid contract and the generation slot as an input a hyperscaler has to secure years ahead of energization, and the firms that own the roadmap would get a pricing conversation this buildout has not yet handed out at the component layer. If it slips, the density targets slip with it, and there is no secondary market for a materials ramp the way there is for a queue position.