800 VDC: How NVIDIA and Infineon Are Reinventing AI Datacenter Power

The announcement almost went unnoticed at GTC Berlin, buried under containerized AI factories: Infineon and NVIDIA are deepening their collaboration on 800-volt direct current architectures, with gallium nitride (GaN) at the heart of power conversion. It may be the week's most structural piece of news — it concerns how electricity enters the rack.

The problem is simple: AI rack power is climbing toward the megawatt, and conventional electrical architectures — AC distribution followed by successive conversions down to processor-level low voltages — are hitting their limits. Every conversion stage loses energy, generates heat and takes up space. Multiplied by hundreds of racks, these losses weigh directly on PUE and on the electricity bill.

The emerging answer: distribute 800 volts DC as close as possible to compute, then convert locally. Infineon is working with NVIDIA on the most critical stage — the high-voltage intermediate bus converter — where GaN switches at up to 1 MHz, with efficiency approaching 98% per stage according to the partners' projections. Fewer losses, smaller magnetics, more compact boards: density rises without heat following.

An Ecosystem Taking Shape

NVIDIA is not pushing alone: some thirty companies are converging on this common architecture — semiconductor makers such as Infineon, Texas Instruments and Navitas, and power specialists such as Delta, Eaton, Schneider Electric and Vertiv. Infineon also signed in September with Estonia's Skeleton Technologies on solid-state transformers, the link between the medium-voltage grid and 800 V DC.

For Europe, the signal is remarkable: the datacenter's electrical battle is partly being fought on technologies where the continent is strong — Infineon in Munich, STMicroelectronics, Schneider Electric. The watt value chain is joining the silicon one, and it has European roots.

What This Changes for Investors

Every efficiency point gained in conversion is a megawatt made available for compute at equal connected power. In a market where the megawatt is the scarce resource, distribution-chain efficiency is no longer an engineering topic: it is a direct driver of sellable capacity and margin.

The transition also creates a divide between assets: a hall designed for 10-20 kW racks will not reach the megawatt without a deep redesign of its electrical distribution. The line is being drawn between renovable assets and those that are not — this has become a selection criterion in its own right.

How We Apply This at FINXIA

In the TITAN DC AI strategy, we now integrate internal electrical architecture into our asset analysis: the distribution's ability to evolve toward high density, space available for new conversion stages, the quality of a site's electrical engineering. These parameters shape an asset's economic life as much as its grid connection.

Our brown-to-green reading is refined accordingly: transforming an existing datacenter also means rethinking its electrical backbone. Assets where this redesign is feasible at controlled cost will capture high-density rack demand; the others will depreciate faster than their accounting suggests.

Key Takeaways

800 V DC is becoming the emerging standard for AI factories: high-voltage DC distribution, GaN conversion close to compute, and an ecosystem of some thirty industrials aligned behind NVIDIA. The datacenter's electrical revolution is underway — and some of its key technologies are European.

Worth watching: the first 800 V DC racks in production with Vera Rubin platforms in early 2027, and dedicated product lines from European power players — Infineon, Schneider Electric, Vertiv.