AI Rack Power Heading Toward Megawatt Scale: Wide-Bandgap Semiconductors See Supply Chain Opportunities

0xBroomberg
Published todayAbout 8 min read

Nvidia AI rack power is surging from 140 kW today to over 1 MW by 2028–2030. The shift to 800 V high-voltage DC is the linchpin, opening a supply-chain entry point for silicon-carbide and gallium-nitride chipmakers.

01

How fast is rack power climbing?

Current Hopper and Blackwell racks draw roughly 140 kW each.
Rubin Ultra will push past 600 kW; by 2028–2030 a single rack will deterministically exceed 1 MW (1,000 kW).
This means → in just a few years, one rack goes from powering "a small building" to powering "a factory floor." Legacy power distribution simply cannot keep up.
02

Why is power demand rising this sharply?

Generative AI and large language models press on compute density, memory bandwidth, interconnect capacity, cooling, and power delivery simultaneously — five dimensions all drawing more watts at once.
In plain terms = no single component got hungrier; every component did, and stacked together the growth is exponential.
Digitimes Research says the traditional data-center power-distribution architecture has hit a physical ceiling.
03

Why is 800 V high-voltage DC the answer?

Data-center power architecture is moving to 800 V high-voltage direct current (HVDC) — converting AC to 800 V DC and feeding racks directly.
Compared with legacy setups, 800 V HVDC offers clear advantages in transmission efficiency, cable losses, and system integration.
This means → higher voltage lets the same power flow at lower current, so cables are thinner, heat is lower, and losses shrink — the key step toward megawatt-class racks.
04

Silicon carbide or gallium nitride — who gets the ticket?

Nvidia's 800 V HVDC supply chain is opening a new market entry for silicon-carbide (SiC) and gallium-nitride (GaN) power-device makers. SiC — power chips made from silicon carbide, strong at high voltage and high temperature — and GaN — power chips made from gallium nitride, faster switching — each target different niches.
SiC leads in voltage rating and thermal management; GaN leads in switching frequency. Both compete differently from legacy silicon-based devices.
In plain terms = it is not about one replacing the other. Different spots in the power chain call for different materials — SiC on the high-voltage backbone, GaN where fast switching matters.
05

How big is the opportunity, and when does it land?

The report offers a quantified market outlook for SiC and GaN in AI data centers but does not disclose specific figures.
On the supply-chain side, SiC and GaN vendors are gradually entering 800 V HVDC procurement systems.
This reflects a deeper test: the real commercialization hurdle is not the technology itself but whether these vendors can lock in share when Rubin Ultra and later generations deploy at scale — the window is 2028–2030.

Content is for reference only, not financial advice.

AI Rack Power Heading Toward Megawatt Scale: Wide-Bandgap Semiconductors See Supply Chain Opportunities · nashnova