Why AI Data Centers Are Trading Silicon for Gallium Nitride Power Semiconductors

AI data centers are trading silicon for Gallium Nitride (GaN) because the material conducts electrons with far less resistance, allowing power supplies to shrink in physical size while wasting less energy as heat. That physical advantage moved from theory to domestic scale this week, as Navitas Semiconductor and GlobalFoundries announced the first shipments of U.S.-manufactured fifth-generation GaN chips from a 200-millimeter line in Burlington, Vermont.[1]

The announcement targets a specific, escalating bottleneck in artificial intelligence infrastructure. While the industry fixates on the graphics processing units performing the calculations, the immediate constraint on building larger data centers is delivering stable, high-density power to those server racks without melting the surrounding hardware.[4]

The solution relies on the fundamental physics of wide-bandgap semiconductors. When a layer of Gallium Nitride meets a layer of Aluminum Gallium Nitride, a phenomenon called a two-dimensional electron gas (2DEG) forms at the interface. This creates a pool of highly mobile electrons that can carry current with minimal friction.[3]

That mobility allows GaN field-effect transistors to switch on and off at frequencies significantly higher than traditional silicon components. Faster switching means the passive components surrounding the chip—specifically the inductors and capacitors that store and filter energy—can be physically smaller, reducing the overall footprint of the power converter by 50% to 70%.[3]

In a server rack drawing tens of kilowatts, saving a few percentage points of conversion loss translates to massive reductions in the cooling load required to keep the facility operational. Less heat generated by the power supply means more thermal headroom for the processors themselves.[3][4]

However, the transition is not a simple one-to-one substitution. GaN transistors currently cost roughly twice as much to manufacture as their silicon counterparts. The economic justification relies entirely on system-level savings—smaller passive components, denser server racks, and lower electricity bills—offsetting the premium paid for the semiconductor die.[3]

The partnership between Navitas and GlobalFoundries attempts to address the supply side of that equation. By adapting Navitas's proprietary device architectures to GlobalFoundries' established 200-millimeter manufacturing platform, the companies aim to leverage the economies of scale that come with larger wafers and standardized foundry processes.[2]

"The first shipment from our U.S. manufacturing line demonstrates how GF and Navitas are turning advanced GaN innovation into a secure, scalable domestic supply," said Kannan Soundarapandian, senior vice president of GlobalFoundries' power business, in the September 1 release.[1]

The initial product family includes 650-volt GaN FETs with on-resistance values ranging from 11 milliohms to 150 milliohms. This wide specification range indicates an intent to address multiple stages of power conversion, from the main rack feed down to the point-of-load, rather than serving a single niche application.[2]

The companies stated that the first wafers are scheduled to ship in September 2026, with internal samples following in October and strategic customer samples by the end of the year. While the timeline represents a concrete manufacturing milestone, independent analysts note that details regarding manufacturing yield, process stability, and volume-production economics remain undisclosed.[1][2]

The domestic manufacturing location also carries strategic weight. Producing these critical power components in Vermont establishes a secure, U.S.-based supply chain for infrastructure that policymakers increasingly view as a matter of national security, reducing reliance on offshore fabrication for the physical layer of AI.[1]

The broader power semiconductor landscape is evolving into a segmented relay race rather than a winner-take-all contest. Silicon Carbide is increasingly favored for the highest voltage applications, while GaN dominates the high-frequency, high-density conversion stages. Traditional silicon still plays a role in delivering the final, stable voltage to the processor.[4]

As the industry pushes toward 800-volt direct-current distribution to support megawatt-scale AI racks, the efficiency of that relay becomes the defining metric of data center design. The success of the Navitas and GlobalFoundries venture will be measured by their ability to deliver consistent, high-yield components that justify the architectural shift away from silicon.[4]