800v Architecture: The Quiet Revenge of Direct Current

A single AI server rack, running at the industry's old 54-volt standard, needs something like 450 pounds of copper busbar to deliver a megawatt of power. That's roughly a motorcycle's worth of metal, wedged into a cabinet that also has to leave room for liquid cooling lines, network cabling, and airflow. Engineers have a name for this problem: the copper wall. Push enough current through a low-voltage system and you eventually run out of physical space to put the copper.

The fix, as it turns out, is to stop using AC and go back to DC — the current Thomas Edison spent the 1880s insisting was the future, and lost. Today it's showing up again at both extremes of the electrical grid: in transmission lines carrying wind power a thousand kilometers across the Gobi Desert, and in the last few inches of wire between a server's power supply and an Nvidia GPU. Edison, it turns out, wasn't wrong about DC. He just didn't have transistors.

The reason DC lost the War of the Currents in the first place had nothing to do with Edison's engineering being bad. Direct current is genuinely well-suited to short distances — steady, low-loss, easy to store in a battery. What DC lacked was a way to change voltage cheaply. George Westinghouse and Nikola Tesla had transformers, simple iron-and-copper devices that could step AC voltage up for long-haul transmission and back down for household use. Edison had nothing equivalent, and by 1892 the company bearing his name had dropped it and gone all-in on AC; it became General Electric. He lost the technical argument and, not coincidentally, waged a fairly ugly PR campaign along the way, including some infamous, gratuitous demonstrations designed to paint AC as lethal. History mostly remembers him for being on the wrong side.

What's changed is that wide-bandgap semiconductors — silicon carbide, gallium nitride — now do cheaply, at kilohertz switching speeds, what only AC transformers could once do: change DC voltage without massive losses. That single technical unlock is why UHVDC transmission lines are spreading across China and the American West, and why the inside of a data center rack looks less like 1995 and more like 1895, if 1895 had had NVIDIA in it.

Which brings us to the part of this story that's happening right now, and that nobody's fully written the ending to. There are, at the moment, two competing 800v architectures inside the AI industry, both marketed under the same name, and they don't much resemble each other. Nvidia's reference design rectifies power at the row level and hands a locked, unipolar 800v bus to a vertically integrated rack — buy the Kyber system, take Nvidia's power architecture, no substitutions. The Open Compute Project's rival spec, co-authored by Google, Meta, and Microsoft, does something more interesting: it rectifies power in a sidecar rack next to the compute, defaults to a bipolar ±400-volt configuration, and accepts any accelerator that meets the spec — Nvidia's chips, but also Amazon's Trainium, Google's TPU, Microsoft's Maia.

The tempting story is that this is physics reasserting itself — the same equation, P equals V times I, showing up at wildly different scales and demanding the same answer. But the actual reason the hyperscalers picked ±400 volts wasn't elegance. It was borrowed infrastructure. The electric vehicle industry spent the past decade building a planet-scale supply chain of 400-volt-class components — inverters, capacitors, contactors — at a cost curve nobody entering fresh could match. Google, Meta, and Microsoft didn't choose 400 volts because it's the "right" number. They chose it because component makers already produce millions of 400-volt parts a year for cars, and inheriting that supply chain is cheaper than building a new one at 800 volts, where yields are still climbing and even the safety certifications are unfinished business.

That distinction matters more for investors than the voltage number itself, because it tells you where the durable demand actually sits. The fight isn't really Nvidia versus the hyperscalers — both camps need essentially the same physical layer underneath them, just wired differently. Whoever supplies the medium-voltage rectification, the wide-bandgap silicon, and the copper-free interconnects gets paid regardless of which architecture wins, which is a more comfortable position than picking a side in a standards war. Eaton and Schneider Electric are the clearest examples: both are shipping 800v-class power shelves into hyperscale builds today, and neither has any reason to bet its business on one topology over the other — they're building for both. Infineon and STMicroelectronics sit a layer further down, supplying the silicon carbide and gallium nitride devices that make high-frequency rectification possible in the first place; their addressable market grows with total 800v deployment, not with who wins the spec fight. The same logic extends outward from the rack to the grid — Siemens Energy and Prysmian are building the HVDC converter stations and subsea cabling for the transmission side of this story, and GE Vernova's grid software is increasingly the layer that decides which projects get connected at all, given multi-year interconnection queues. None of these companies need Mt. Diablo or Nvidia's reference design to prevail; they need the industry to keep building at 800v, full stop, and every signal points to that continuing regardless of which camp's logo ends up on the sidecar.

The more interesting investment question isn't who wins the architecture fight — it's what happens to component pricing as volume compounds. The 400-volt class has a decade-long head start riding the EV industry's cost curve; the 800v class is earlier and currently carries something like a 15 to 25 percent cost premium at the component level. That gap is the thing to watch, because it will close as annual unit volumes climb, and the pace of that closing is a reasonable proxy for how quickly capital shifts from legacy AC infrastructure into this new stack. Vertiv, sitting at the intersection of power and thermal management inside the data center itself, is arguably the single name most directly levered to that volume ramp, since it profits from total watts deployed at high density almost independent of which voltage topology carries them.

None of this resolves neatly, and it shouldn't. Infrastructure decisions like this tend to calcify long after the original justification stops mattering — American homes still run on 120 volts and 60 hertz for reasons that made sense around 1900 and haven't made sense in decades, and nobody rips out a national grid once switching costs harden around it. The more useful question for anyone allocating capital into this space isn't which spec wins by 2030. It's whether you're positioned in the layer of the stack that gets paid either way — the silicon, the rectifiers, the copper alternatives — rather than betting on a single reference design that might look, in retrospect, like 2026's equivalent of Betamax.