Data centers are running into a thermal wall. AI training clusters consume enough electricity to require dedicated substations, and a significant fraction of that power does not make it to the processors. It dissipates as heat in the power delivery circuitry itself. Every voltage conversion step between the grid and the chip is a heat source. The solution has always been to move the conversion closer to the processor, but silicon-based power devices have physical limits: they become unreliable above roughly 150 degrees Celsius, which constrains how much power you can push through them without accumulating dangerous heat.
Intel Foundry believes gallium nitride, a wide-bandgap semiconductor that can operate at higher temperatures and voltages than silicon while switching faster, might finally break through that ceiling. On April 9, 2026, the company announced what it calls the world's thinnest GaN chiplet, with a base silicon layer just 19 micrometers thick, roughly one-fifth the width of a human hair, built on 300mm GaN-on-silicon wafers, the production-standard wafer size that makes volume manufacturing theoretically feasible. The company presented the work at the IEEE International Electron Devices Meeting in December 2025. Semiconductor Engineering TrendForce
The thinness is a manufacturing achievement, not the story. The story is what the integration enables: for the first time, Intel has built full digital control circuits directly onto the same die as GaN, using a single manufacturing process that combines GaN N-MOSHEMT transistors with silicon CMOS. Most GaN chips in use today are analog or discrete power devices that still require a separate silicon chip to handle digital control functions. Eliminating that separate companion chip removes an intermediate conversion step, which means less energy lost as heat and faster response to load changes.
The performance numbers from Intel's paper give a sense of what is physically possible. Inverters on the chiplet switched in 33 picoseconds across the full 300mm wafer. The GaN transistors blocked up to 78 volts, which is the voltage class used internally by AI accelerator power delivery networks, and achieved radio-frequency cutoff frequencies above 300 GHz, putting them in range for both 5G and 6G wireless infrastructure alongside power applications.
The 300mm wafer size is what makes this worth watching beyond a research curiosity. Growing GaN on wafers that size is harder than on smaller substrates, which is why most GaN research historically happened elsewhere. Intel's demonstration shows it can be done and done consistently across a full production-scale wafer. If this moves from conference paper to fab line, GaN power management chips could theoretically be manufactured on existing 300mm equipment without requiring new tooling. GlobalFoundries announced a separate GaN manufacturing partnership with Navitas Semiconductor in November 2025 for its Burlington, Vermont facility, with production expected later in 2026, evidence that the industry is already moving toward 300mm GaN production regardless of Intel's specific outcome. Semiconductor Engineering
The reliability data from Intel's paper is encouraging but limited. The GaN passed time-dependent dielectric breakdown, positive bias temperature instability, high-temperature reverse bias, and hot-carrier injection stress tests, the standard qualification suite for automotive and infrastructure applications. "Met required reliability metrics" is engineering language for "it held up during the test." Whether it survives a decade of thermal cycling in a data center rack is an open question. No product has shipped. No customer has been named. Intel announced the technology on April 9, 2026 via Intel Foundry Services, but the announcement is a research result, not a product roadmap. TrendForce
The 200GHz figure circulating in coverage needs context. Intel's paper demonstrates RF cutoff frequencies above 300 GHz. Cutoff frequency is where a transistor stops amplifying, not where it computes. A transistor that cuts off at 300 GHz is not a 300 GHz processor. The extrapolation to 200GHz AI processors is not something Intel has shown in a shipped product, and treating the paper as evidence for that claim misreads what it actually proves.
What the paper does support is the thermal argument for GaN in data center power delivery. Higher switching frequencies, higher voltage tolerance, operation at temperatures that would degrade silicon, and a path to integration on production-scale wafers are a combination that changes the economics of where power conversion can live in a rack. Whether Intel is the company that commercializes it, and on what timeline, remains to be seen.
