← Back to articles

The Rising Role of Power Semiconductors in Data Center AI Infrastructure

December 30, 2025

=======================================================================

Power semiconductors are poised to become as crucial as GPUs and CPUs in data centers, especially with the explosive growth of AI workloads demanding ever-increasing power capacity.

The AI Boom and Its Power Challenges

Data center operators have been densely packing servers with accelerators like GPUs, which can consume up to 700 W each. This trend significantly escalates energy requirements across racks and entire data centers, straining existing electrical infrastructure.

For example, Nvidia's latest integrated AI system, the DGX GB200 NVL72, draws 120 kW per rack, while giants like Google are planning for racks supporting up to 1 MW of IT load. Meanwhile, critical components such as transformers are experiencing long lead times of 28 weeks or more, delaying deployments.

Power Semiconductors: The Game Changer

Enter power semiconductors, which enable the development of solid-state transformers (SSTs). These devices are far smaller and more efficient than traditional transformers—potentially weighing just 500 kg instead of several tons—thanks to their ability to operate at very high switching frequencies.

"A solid-state transformer offers advantages in both supply chain and performance," explained Dr. Peter Wawer, division president at Infineon.
"Because SSTs operate at high frequency, they require silicon carbide components capable of switching at these speeds and blocking high voltages of up to 2 or 3 kV."

The Role of Silicon Carbide and Gallium Nitride

Silicon carbide (SiC) and gallium nitride (GaN) are key materials enabling these innovations due to their ability to handle high voltages and switching frequencies efficiently. Infineon, a leading power semiconductor supplier, is investing heavily in developing SSTs utilizing these materials.

The market for SSTs is expected to grow significantly, with projections reaching up to $1 billion by 2030, as they gradually replace traditional transformers in datacenter power infrastructure.

Enhanced Circuit Breakers and Distribution

Power semiconductors are also poised to upgrade other infrastructure components, such as circuit breakers. Silicon carbide-based solid-state circuit breakers can open in microseconds—much faster than electromechanical versions—providing enhanced safety and reliability in high-voltage environments.

In addition, these semiconductors are integral to reshaping power distribution within data centers, promoting more efficient and compact power supplies.

Higher Voltage and Efficiency

To meet the growing power demands, Nvidia advocates for 800 VDC power supplies, which reduce the number of voltage conversions, thereby lowering associated losses and complexities. Infineon supports this move, noting that higher voltages allow for increased efficiency by keeping current levels manageable.

"Fewer conversion steps mean less energy wasted," said Adam White, president at Infineon.
"This approach also simplifies the design and reduces inefficiencies."

The Path to Integration: Power Sidecars and Rack Upgrades

Before 800 VDC systems become mainstream, a transitional architecture—referred to as a "power sidecar"—may be employed. This involves an additional rack with power delivery and backup systems running alongside the IT rack, enabling power capacities of up to 600 kW, compared to today's typical 125 kW.

Both Google and other vendors are exploring such configurations for mega-racks supporting 1 MW or more, with dedicated AC-to-DC power racks providing support for main racks.

Scaling Power Supplies and Component Innovations

As power densities continue to increase, the internal power supply units (PSUs) within racks are also scaling up—moving from 8 kW units to 16 kW or higher—and transitioning to three-phase designs. Components made from SiC and GaN play an essential role here as well.

Enhanced Chip-Level Power Management

At the chip level, the rising power consumption of accelerators like GPUs and TPUs necessitates innovations in voltage regulation modules (VRMs). Traditionally discrete, VRMs are moving toward integrated or backside vertical power modules (BVMs), which sit beneath the chips, reducing parasitic losses and simplifying system design.

Looking ahead, industry hopes to embed VRMs directly into the chip substrate, a development expected beyond 2027.

Conclusion

As AI infrastructure demands escalate, the adoption of advanced power semiconductors, solid-state transformers, and innovative power management architectures will be pivotal to building efficient, scalable data centers capable of supporting the next generation of AI workloads.