What is 48V Power Architecture — and Why Is the AI Industry Already Moving Past It?

What a power architecture is — and why voltage is the key variable. A power architecture defines how electricity moves from the grid to the chips that need it. Alternating current (AC) from the utility arrives at the building, is converted to direct current (DC), and is distributed at a defined voltage across the facility. That voltage determines the amount of current required to deliver any given amount of power — and current is what heats cables, causes resistive losses, and limits how much power can be pushed through copper of a practical size.

The physics is captured in two laws. Power equals voltage times current (P = VI), so for the same power, higher voltage means lower current. Resistive losses scale as current squared times resistance (I²R), so halving the current cuts losses by three-quarters. When Google and the Open Compute Project pioneered 48V DC distribution in the early 2010s, they applied this logic to replace 12V. Four times higher voltage meant one-quarter the current — and one-sixteenth the resistive losses. That improvement alone allowed racks to scale from roughly 10 kW to over 100 kW while keeping cabling practical.

Why AI has broken every prior assumption. The problem is that AI chips have scaled power consumption faster than any previous generation of semiconductor technology. In 2022, a single AI processor consumed approximately 0.4 kW. By 2024, that figure had crossed 2 kW and single racks were approaching 150 kW. In 2026, NVIDIA's GB200 NVL72 racks run at roughly 120–140 kW — right at the practical ceiling of 48V architecture. NVIDIA's Vera Rubin Ultra, arriving in 2027, is projected to push single-rack demand toward 600 kW. The Feynman architecture targeted for 2028 reaches 1 megawatt per rack.

Delivering 1 MW through a 48V architecture requires over 200 kilograms of copper cabling. The currents involved — more than 20,000 amperes — demand conductor cross-sections that make physical deployment impractical. Google's statement at OCP APAC is the clearest public signal yet: 48V has become the 12V of the AI era. It served its generation well. It is now the constraint on the next one.

NVIDIA's 800V answer — and why it changes data centre design. NVIDIA's solution, previewed at GTC 2025 with the Kyber NVL576 system and formally detailed in January 2026, is 800V high-voltage direct current (HVDC). In this architecture, AC from the grid is converted to 800V DC at the rack perimeter in a dedicated sidecar unit — not inside the compute rack. Inside the rack, DC-to-DC converters step voltage down to what GPU chips actually need (approximately 1–3V). The sidecar approach frees the compute rack entirely of bulk power electronics, replacing 200-kilogram copper busbars with lighter, more efficient conductors.

At 800V, delivering 1 MW requires 1,250 amperes of current — compared to 20,833 amperes at 48V. The copper weight falls from over 200 kg to approximately 12 kg for the same power delivery. The space freed inside the rack, the reduction in heat from resistive losses, and the elimination of AC-to-DC conversion stages inside the compute unit together explain why NVIDIA's partners — including Delta Electronics, Lite-On, Eaton, Schneider Electric, and Vertiv — are building 800V supply infrastructure for 2027 deployments.

The OCP alternative: Mount Diablo and ±400V DC. Google, Meta, and Microsoft have taken a parallel path. Through the Open Compute Project, the three hyperscalers are co-developing "Mount Diablo" — a bipolar ±400V DC architecture. Instead of a single 800V rail, Mount Diablo uses two opposing 400V rails with a centre ground point. Total potential across the system is 800V, but the bipolar design offers different grounding and safety characteristics than NVIDIA's unipolar approach. A sidecar power unit handles AC-to-DC conversion outside the compute rack, disaggregating power from compute in a way that hyperscalers with mixed-vendor compute fleets prefer.

Meta's Prometheus datacenter — a multi-gigawatt facility coming online in 2026 — and the planned Hyperion supercluster targeting 5 gigawatts are the scale at which these architectural choices become existential for infrastructure teams. The choice between NVIDIA's 800V track and the OCP's ±400V DC track is the most consequential unresolved variable in data centre design today. Until the industry standardises on shared connectors, grounding conventions, and safety codes, data centre operators building infrastructure for both ecosystems carry genuine integration risk.

Who builds the power semiconductors Asia's data centres need. The voltage transition is a substantial tailwind for Asia's power semiconductor industry. The GaN (gallium nitride) devices that dominate 48V power conversion — in the 40–65V breakdown class — are produced primarily by Navitas Semiconductor, EPC, GaN Systems, and Infineon, with GaN-on-Silicon manufacturing concentrated in Asia. For the 800V conversion stage, silicon carbide (SiC) MOSFETs from onsemi, STMicroelectronics, ROHM (Japan), and Infineon are the preferred switching devices. ROHM in particular has deep foundry relationships across Japan and Southeast Asia.

Taiwan's ODM power supply manufacturers are directly in the commercial path. Delta Electronics and Lite-On Technology are both explicitly named NVIDIA partners for 800V rack power systems. AcBel Polytech and Chicony Power Technology are positioned for sidecar PSU production. Japan's Murata and TDK supply the passive components — capacitors, inductors, magnetics — that every power conversion stage depends on. The data centre power transition is not a story about US chip companies alone. It is a story about Asia's power component and manufacturing ecosystem becoming critical infrastructure for AI.

Three predictions for 2026–27. First: NVIDIA's 800V DC architecture enters production deployments at neocloud operators and frontier AI labs by Q2 2027 — ahead of the hyperscaler mainstream, which waits for Mount Diablo interoperability standards to mature. Second: the standards fork between NVIDIA's 800V unipolar and OCP's ±400V bipolar is not resolved before end-2027; data centres built for one ecosystem will require active bridging equipment to support the other.

Third — and this is the prediction most likely to matter commercially — Taiwan's power supply ODMs capture disproportionate share of the 800V sidecar unit market by end-2027 relative to their current 48V PSU share. The sidecar is a new category with no installed base; Delta, Lite-On, and Chicony are better positioned than their European and North American competitors to execute new-design production at the volumes hyperscalers require. Watch the ODM earnings calls from Q1 2027 for the first evidence of 800V revenue materialising at meaningful scale.