From SiC and GaN to AI Server Power Supplies: Mapping the Supply Chain and Taiwan’s Supplier Roles

AI Server Performance Also Depends on How Power Reaches the Chips

Discussion of AI servers usually centers on GPUs, HBM, and high-speed networking. Yet stable operation of all these components shares another prerequisite: a power system that converts electricity efficiently and remains stable as loads change rapidly.

A power supply converts incoming electricity into the voltage and current the equipment needs. As the power rating of an individual unit rises, designers must handle more heat and higher currents within limited mechanical space. SiC, or silicon carbide, and GaN, or gallium nitride, are attracting attention because they offer new device choices for these problems.

Concrete products and designs already illustrate the shift. In 2025, Delta announced a 5.5kW server power supply certified to 80 PLUS Ruby, with stated peak conversion efficiency of 97.5%. Infineon’s 12kW reference design, introduced the same year, combines Si, SiC, and GaN.[1][2]

Both examples show that power-supply upgrades involve joint design of materials, circuits, and systems. Understanding the industry opportunities begins with separating the supply-chain layers.

1. How Do Si, SiC, and GaN Share the Work Within a Power Supply?

Si is silicon, while SiC and GaN are wide-bandgap semiconductors. Their wider bandgaps and other material properties can allow devices to withstand higher electric fields or reduce losses under particular operating conditions. Material selection still depends on operating voltage, switching frequency, conduction losses, thermal management, and cost.

These are common roles, with overlap in actual circuits. Infineon’s 12kW reference design, for example, uses SiC in the front-end PFC stage and GaN in the isolated LLC conversion stage. A separate TI reference design uses GaN in both PFC and LLC stages. SiC therefore cannot be rigidly classified as a “front-end material,” or GaN as a “downstream material.”[2][3]

Silicon also continues to serve high-end power supplies. Engineers choose suitable devices stage by stage, so using all three materials in one unit can be a rational system-design outcome.

2. From Materials to Chips: SiC and GaN Follow Different Paths

The power-semiconductor supply chain can be divided into five layers: substrates, epitaxy, device design, wafer fabrication, and packaging/testing. Some companies span several layers; others specialize in a particular process.

SiC: A Crystal and Substrate Come First, Followed by the Device’s Epitaxial Layer

SiC power devices typically begin with conductive SiC substrates. After crystal slicing and polishing, an epitaxial layer suited to the device design is grown before wafer fabrication. Defects, uniformity, and electrical properties in both the substrate and epitaxy affect the final device. Wolfspeed and Coherent product information similarly distinguishes conductive SiC for power applications from materials used in other applications.[4]

GaN: Power Devices Can Be Grown on Silicon Substrates

One important route is GaN-on-Si: growing a GaN-containing epitaxial structure on a silicon substrate. MOCVD processes can use precursors such as trimethylgallium and ammonia to form thin films in a controlled environment. The supply chain therefore also includes specialty chemicals, gases, deposition equipment, and metrology tools.[5]

Taiwan’s Episil-Precision publicly offers GaN-on-Si epitaxial products, providing a concrete example of this role. GaN-on-Si also demonstrates that a “GaN device” does not imply the entire substrate is gallium nitride. GaN-on-SiC, common in radio-frequency communications, involves different substrate choices and design requirements that cannot simply be carried over to server power supplies.[6][7]

Packaging involves much more than enclosing a chip. High-current connections, package parasitics, and heat-removal paths all affect whether a device can achieve its performance. ASE’s power-packaging portfolio includes solutions for SiC and GaN devices.[8]

3. What Is Inside a PSU Besides Power Chips?

A typical AC-input server PSU, or power supply unit, first handles input protection and electromagnetic interference. Rectification and power factor correction (PFC) establish a DC supply, followed by isolation, voltage conversion, rectification, and filtering. Controllers continually monitor voltage, current, and temperature while coordinating the stages.[9]

PFC primarily improves the AC input current waveform and power factor. Conversion efficiency measures how much input power becomes useful output power. Both matter, but they are different metrics.

Opening a PSU reveals several converging material and component supply chains:

The table describes common materials and functions; actual combinations depend on circuit and product requirements. Magnetics, capacitors, laminates, and thermal materials can be cross-checked against technical or product information from TDK, APAQ, Murata, Isola, and Henkel.[10][11][12][13][14]

GaN’s support for higher switching frequencies may create room to shrink magnetic components, but core losses, winding losses, temperature rise, and electromagnetic interference must be rebalanced. Higher power density often requires coordinated changes to chips, magnetics, PCB layout, and cooling.

4. Where Do Taiwanese Suppliers Fit?

Taiwan participates from upstream wafers through rack power delivery. The following 12 representative companies are classified using public product and company information. Their listed capabilities do not establish particular supply relationships among them.

These companies occupy different technical positions. Epitaxy suppliers manage material quality; foundries turn designs into devices; capacitor and inductor suppliers support energy storage and filtering; power-supply makers handle complete-unit design, integration, and qualification. Identifying which product layer a company delivers is the first step to understanding how technology upgrades affect it.

5. From PSUs to Complete Rack Power, Product Boundaries Are Expanding

Power-related coverage often discusses different product levels together, but an individual PSU’s rating cannot be directly compared with an entire power shelf’s output.

AcBel’s publicly listed 33kW power shelf combines six 5.5kW PSUs. LITEON’s 110kW power shelf displayed at COMPUTEX 2026 is a higher-power system-level product.[21][22] Comparisons also need input and output voltages, cooling conditions, and power available to the load after allowing for redundancy.

This also clarifies the different roles of TAI-TECH and BizLink. TAI-TECH’s relevant inductors can serve conversion near GPUs and CPUs, while BizLink’s busbars and connectors transmit power within the rack. Both participate in server power delivery, but their products and qualification requirements differ.[18][19]

6. Why Are Power Architectures Moving Toward 800V?

As distributed power rises, current becomes a challenge. Using power = voltage × current and ignoring losses, delivering 100kW requires approximately 2,000A at 50V, but only about 125A at 800V. This compares distribution at equal power; it does not mean an actual rack routes all its current through one connector.

Higher distribution voltage reduces current, easing demands on conductors, contacts, and resistive losses. Once power enters the equipment, further conversion is still required to produce the low voltages needed by GPUs and CPUs.

In an official August 2026 article, NVIDIA described a gradual move toward high-voltage DC using 800VDC power racks compatible with existing AC infrastructure, listing the second half of 2026 for their introduction. This is a product and architecture roadmap that can be tracked, not evidence that data centers have universally adopted 800V.[23]

For the supply chain, the shift affects high-voltage conversion devices, isolation, protection, connectors, busbars, and thermal design together. Changing supply voltage also increases the importance of system integration.

Looking downstream from SiC and GaN reveals a chain from materials to complete systems. Looking upstream from GPUs reveals the links among board-level regulation, rack distribution, and data-center power. Taiwanese suppliers participate at several intersections of these paths. Understanding each company’s role and the evidence for its product progress provides a clearer view of changes in the power industry.