7 min read

NVIDIA is shifting data-centre power delivery from AC to 800 V DC. More than two dozen suppliers-from Infineon and STMicroelectronics to Schneider Electric and Vertiv-have already pledged silicon, components and complete power systems. From 2027 onwards, racks rated at one megawatt and above will come online. For operators in the DACH region, this isn’t a distant vision; it’s a procurement decision that needs to be made today.

Related:Data-centre energy balance and the EnEfG  /  Computex: Huang, TSMC and AMD’s 10-billion-packaging bet

Why AC distribution is hitting its limits

An old-school rack in a corporate data center draws five to fifteen kilowatts. At that level, a 415 V or 480 V three-phase distribution system copes just fine. A fully loaded GPU rack for AI training today sits at 40 to 120 kilowatts. NVIDIA’s next-generation platform around Vera Rubin is targeting one megawatt per rack. That is no longer a jump-it is a different order of magnitude.

The reason is not academic; it is etched into the current path. Today the energy to each GPU passes through several stages: from the grid to the rack level, then from AC to DC, and finally stepped down in multiple steps to the few volts the chip needs. Every stage siphons off a few percent. In total, a double-digit share of the input power is lost as heat before any computation even begins. At one megawatt per rack, those losses are no rounding error; they become their own cooling load and their own line item on the electricity bill.

At these power densities, the AC path shows its weaknesses. Every conversion from AC to DC inside the servers consumes energy and eats up real estate. Busbars grow thicker, line losses rise with the square of the current. Raising the voltage lowers the current for the same power-and with it the losses. That is exactly what 800 V DC delivers: fewer conversion stages, thinner copper cross-sections, fewer PSUs in the rack.

NVIDIA puts it plainly: moving from 415 V or 480 V three-phase to an 800 V DC infrastructure delivers more scalability, better energy efficiency, less material, and higher power density. It sounds like marketing, yet it describes a real physical problem. Anyone who has ever tried to spec a 400 V panel for 100 kW racks knows the point at which the copper busbars cost more than the cabinets themselves.

Who’s on board with the 800-volt standard

An architecture is only as resilient as its supply chain. That’s exactly where things have shifted since the autumn. At the OCP Global Summit in October, NVIDIA unveiled the 800-volt concept alongside a roster of partners who have since announced concrete building blocks. On the silicon side, the line-up includes Analog Devices, Infineon, Innoscience, Monolithic Power Systems, Navitas, onsemi, Power Integrations, Renesas, ROHM, STMicroelectronics and Texas Instruments. Among the power-system vendors the names are familiar: ABB, Eaton, Hitachi Energy, Mitsubishi Electric, Schneider Electric, Siemens and Vertiv.

Texas Instruments delivered the clearest tangible proof in March. The manufacturer showcased a complete 800-volt DC reference architecture-not just individual chips but the entire path from mains connection to the voltage at the chip. That’s the difference between an intention and something a planner can actually work with. Meanwhile, Foxconn is building a 40-megawatt data centre in Kaohsiung from day one designed around 800 volts.

Technically, the shift is being driven by a new generation of semiconductors. Converters based on gallium-nitride and silicon-carbide switch faster and with lower losses than classic silicon. That’s why names like Navitas, Innoscience and onsemi appear on the partner list: they supply the power switches without which an 800-volt path at these efficiency levels simply wouldn’t work. Anyone who understands the converter stages can see that this isn’t just a voltage swap; the entire power-electronics stack is taking a step forward.

For the practitioner, that means the components aren’t coming from a single vendor you’re locked into. Multiple sources for converters, busbars and protection gear reduce the risk that a project stalls at a supply choke-point. It’s precisely this breadth that has given three-phase standards their decades-long stability. Seeing it now for DC is the real signal-not the headline with the partner count.

What DACH operators should take away from this

Most enterprise data centres in Germany, Austria and Switzerland today run comfortably on three-phase alternating current. Not a single one of them will need to be converted tomorrow. The point is a different one: anyone planning high-density AI workloads in earnest over the next two years makes the power decision at the start of the project, not at the end. An 800-volt busbar cannot be shoehorned into a space already wired for 400 volts.

Three questions help put the situation in perspective. First: what is the planned power density per rack? Below 50 kilowatts, three-phase AC remains economical; above that threshold the maths favours direct current. Second: own operation or colocation? Tenants leasing space are at the mercy of the provider’s roadmap and should flag 800 volts in the lease contract early. Third: how is the data centre cooled? One-megawatt racks exist only with liquid cooling. If you are still using air cooling, you already have a bigger project on your hands than the voltage.

Then there is the staffing question. 800-volt DC is more hazardous in a fault scenario than three-phase AC because a DC arc does not self-extinguish. Protection gear, isolators and safety procedures differ from the familiar AC regime. This is not a showstopper, but it is an item that belongs in every honest project budget. Teams that currently maintain AC kit will need additional training for 800 volts. Discovering that need only at the commissioning date is the classic mistake.

It pays to take a sober look at your own energy balance. The efficiency gain from 800 volts is real, yet it only materialises at high density. In a facility running a few dozen standard racks, the voltage is not the lever that moves the electricity bill. Germany’s Energy Efficiency Act already sets PUE targets that are easier to hit by optimising cooling and load management than by switching voltages.

800-volt DC is not a trend to chase. It is the consequence when a single rack draws so much power that the distribution gear costs more than the servers inside it.

When the switch pays off-and when it doesn’t

The bluntest answer is also the least comfortable: it depends on density and the time horizon. For a new build scheduled to house GPU clusters from 2027 onwards, 800 volts is the obvious choice. Components are available, supply chains are broad, and the efficiency edge kicks in immediately. Deciding today to wire for 400 volts means investing in an architecture that will already lag a generation by move-in day.

For existing sites the opposite holds true. A live data centre with moderate density is not torn apart for a voltage whose benefits only appear at loads you do not yet run. The costliest error in infrastructure is rarely a decision made too late. It is the one made too early, tying up capital before the need materialises. If you are not planning megawatt racks, you can wait without a guilty conscience.

What everyone shares: the direction is set. The next two years decide less about whether to switch than about when. This clarity did not exist before the OCP announcement and the TI reference design. That clarity is the real planning upside.

Frequently Asked Questions

Do I need to retrofit my data center to 800 volts now?

No. An existing operation with racks under 50 kilowatts remains economically viable with three-phase power. 800 volts only becomes relevant at very high power density-think AI training or inference clusters targeting megawatt-scale racks.

What is the concrete advantage over 400-volt AC?

Higher voltage lowers current for the same power, reducing line losses. It also eliminates conversion stages and rack-level power supplies, saving energy, copper, and floor space once each rack’s load reaches several hundred kilowatts.

Are the components actually available?

Yes. Over 20 manufacturers have announced silicon, components, and power systems, including Infineon, STMicroelectronics, Schneider Electric, Siemens, ABB, and Vertiv. Texas Instruments demonstrated a complete reference architecture in March.

Does 800 volts matter for colocation customers?

Indirectly. If you lease space, you depend on your provider’s roadmap. If you need high GPU density from 2027 onward, make sure 800-volt capability is written into your lease and factored into site selection early.

Can 800 volts work without liquid cooling?

Not in practice. The power densities that justify 800 volts cannot be dissipated with air cooling. NVIDIA’s Vera Rubin systems are designed for full liquid cooling.

Title image: AI-generated (May 2026), C2PA certificate embedded