800VDC Architecture Accelerates Adoption, Reshaping Power Chip Demand for AI Data Centers

Nashnova编辑部
Published todayAbout 12 min read

AI data centers are racing toward 800V DC power distribution, set to reshape demand for GaN, SiC, and vertical power delivery chips. DIGITIMES analyst Yao Jiayang sees the shift turning markedly sharper once Nvidia's Rubin Ultra GPU enters mass production in 2027.

01

Why are data centers switching from AC to 800V DC?

AI rack power is climbing toward the megawatt level. Every extra voltage conversion in a traditional AC distribution chain wastes energy and money.
The core idea behind 800VDC: In plain terms = cut out the middlemen. A solid-state transformer (SST) — a device that uses semiconductors instead of iron-core transformers — converts 13.8–35 kV medium-voltage AC directly into high-voltage DC for the rack, slashing conversion losses.
This means → better power usage effectiveness (PUE) doesn't just save electricity. It simultaneously shrinks cooling needs, floor space, and copper-cable costs — helping operators offset the heavy capex of AI infrastructure.
02

What roles do GaN and SiC each play?

The division of labor is clear: GaN devices handle DC-to-DC step-down; SiC devices serve both the Power Sidecar cabinet — the unit that feeds 800VDC to compute racks — and the SST.
Industry estimates put the SiC requirement for a single 1 GW data center at roughly 200,000 SiC modules, mostly inside SSTs.
This reflects a direct link: SiC demand growth is tethered to the pace of 800VDC deployment. Whoever scales SSTs first captures the first wave of SiC orders.
03

Why does vertical power delivery (VPD) keep coming up?

VPD's core idea: move the power-conversion module right next to the processor — beneath the substrate, on the PCB backside, or even inside the substrate. Put simply = the shorter the electrical path, the less energy is lost in transit, and the smaller the module.
The benefits are concrete: shorter path → lower parasitic resistance and inductance → reduced power-distribution-network impedance.
But deeper integration is still stuck on thermal management. Silicon, copper, and packaging materials have mismatched coefficients of thermal expansion — the closer they sit, the higher the risk of mechanical stress. An engineering breakthrough is still needed.
04

How far along are Western and Japanese vendors on GaN expansion?

Texas Instruments: operates one 8-inch GaN power-wafer fab in Dallas and another in Aizu-Wakamatsu, Japan.
Infineon: already capable of 12-inch GaN mass production; customer sampling begins in Q4 2025. This means → Infineon leads on wafer size, producing more chips per wafer and building a cost advantage that will compound over time.
Innoscience (China's GaN leader): 8-inch monthly capacity is expected to reach roughly 20,000 wafers by end-2025, with plans to scale to 70,000 within five years and to begin 12-inch production by 2030.
05

What do the Japanese players and cross-border alliances look like?

Renesas Electronics: mass-produces 6-inch GaN wafers at the former Transphorm fab in Aizu-Wakamatsu and 8-inch GaN wafers through Polar Semiconductor.
Rohm Semiconductor: licensed TSMC technology and plans to begin production at its Hamamatsu fab by the end of 2027.
The partnership patterns are telling: onsemi teams with GlobalFoundries and Innoscience to chase China market share; STMicroelectronics partners with Innoscience while also running its own 8-inch fab in Tours, France. This reflects a reality — GaN expansion is not a solo game. Cross-border alliances and capacity-sharing are becoming the norm.
06

Is 2027 the make-or-break checkpoint for this cycle?

DIGITIMES analyst Yao Jiayang judges that the 800VDC shift will become markedly more pronounced once Nvidia's Rubin Ultra GPU enters mass production in 2027.
This means → the GaN and SiC capacity that Western and Japanese vendors are building right now is a bet on demand crystallizing around 2027.
In plain terms = if Rubin Ultra ships on schedule, the expansion investments pay off; if it slips or demand disappoints, overcapacity risk looms. 2027 is the watershed for this power-technology cycle.

Content is for reference only, not financial advice.