The power demands of artificial intelligence data centers have become a central headache for the tech industry. Training and running large models requires enormous amounts of electricity, and the infrastructure needed to cool and operate those servers is straining grids and budgets alike. Onsemi, a semiconductor company best known for power management and sensor chips, says it has a direct fix.

Hassane El-Khoury, Onsemi's president and chief executive, told the editorial desk that the company's silicon carbide chips are designed to solve what he calls the "power problem" facing AI data centers. According to El-Khoury, Onsemi's silicon carbide (SiC) technology can significantly improve the efficiency with which electricity is converted, distributed, and consumed inside massive server farms.

The claim is a notable one because data center power inefficiency has become a bottleneck for AI deployment. As more companies race to deploy generative AI, the energy required to keep GPUs and other accelerators running 24/7 has surged. Analysts have warned that without dramatic improvements in power electronics, the growth of AI could be limited by physical infrastructure constraints. Onsemi is betting that its SiC chips can loosen that constraint.

What makes silicon carbide different

Silicon carbide is a wide-bandgap semiconductor material that can handle higher voltages, temperatures, and switching frequencies than traditional silicon chips. In power conversion applications — the systems that take high-voltage alternating current from the grid and turn it into the low-voltage direct current that processors need — SiC chips waste less energy as heat. That means more of the electricity pulled from the grid actually reaches the computing hardware, and less needs to be removed by cooling systems.

For a data center, this efficiency gain translates directly into lower operational costs and reduced cooling load. A typical AI training cluster can draw tens of megawatts. If power conversion losses can be cut by several percentage points using SiC, the savings in electricity bills and cooling infrastructure become significant.

Onsemi has been manufacturing SiC chips for several years, targeting electric vehicles and industrial power systems. The company now sees AI data centers as a natural next market. El-Khoury's comment suggests Onsemi believes its SiC products are already ready to scale inside the data center power chain, from the uninterruptible power supplies to the voltage regulators right next to the processors.

The scale of the AI power problem

AI data centers are not just larger versions of traditional server rooms. They pack far more transistors per square meter, and those transistors draw far more current. The latest GPUs from companies like Nvidia can draw 700 watts or more per chip, and a single rack may hold dozens of them. All that power needs to be delivered with extreme precision and reliability, and every conversion step along the way loses some energy.

The industry has responded with new power architectures, including 48-volt server racks and direct liquid cooling, but the fundamental efficiency of the power semiconductors themselves has not kept pace. That is where Onsemi's SiC chips enter the picture. By using SiC MOSFETs and diodes, data center operators could potentially reduce power losses by 50 percent or more in some conversion stages, according to industry estimates — though Onsemi’s CEO did not provide specific numbers in the briefing.

El-Khoury framed the solution as a direct answer to a problem that is only getting worse. As models grow larger and inference workloads become more common, the total power appetite of AI could dwarf current projections. Onsemi wants to be the company that supplies the building blocks to keep the lights on — literally.

Onsemi's position in the market

Onsemi is not the only company making silicon carbide power chips. Competitors like Wolfspeed, STMicroelectronics, and Infineon have also invested heavily in SiC wafer production. But Onsemi has been expanding its own manufacturing capacity and has worked to integrate SiC into existing power system designs. The company’s earlier work in automotive power gave it experience with high-reliability, high-volume production, which is essential for the data center market.

The CEO’s statement did not include specific product names, pricing, or customer announcements. But the claim that Onsemi “solves” the power problem implies that the company sees its technology as a comprehensive fix, not just a marginal improvement. Whether that holds up depends on how well the chips perform in real-world data center deployments, which are notoriously conservative about adopting new power components.

Adoption of SiC in data centers has been slower than in electric vehicles, partly because data center power design cycles are long and risk-averse. A failure in a power supply can take down thousands of servers. Operators need proven reliability before switching from silicon to SiC. Onsemi will need to demonstrate that its chips meet those standards, not just on paper but after years of operation.

What it means for the industry

If Onsemi’s silicon carbide chips can deliver on the promise of solving the data center power problem, the implications go beyond cost savings. More efficient power conversion would mean that existing grid connections could support more compute per watt. That could accelerate the build-out of AI infrastructure without requiring as many new power plants or transmission lines. It would also reduce the carbon footprint of AI, which has become a focus for regulators and environmental groups.

The CEO’s statement is a clear signal that Onsemi sees a huge opportunity and is positioning itself as a leader in that space. Whether the company can capture that opportunity depends on execution, manufacturing scale, and customer trust.

For now, the message from El-Khoury is simple: the tools to fix the AI power problem already exist, and Onsemi has them. The rest of the industry will be watching closely to see if that claim holds under the heat.