NVIDIA's next-generation rack will pull a megawatt of power - the draw of a small town. There's no way to feed that at today's voltages. Everyone is watching the GPU; the real story for the next three years is the wire that feeds it - and the entire industry is being forced to do something it has avoided for decades: go to 800 volts.
Start with the demand, because demand is what makes this inevitable. NVIDIA doesn’t really sell you a chip anymore. It sells you a rack - a whole AI computer. And those racks are getting violently denser. Today’s pull maybe 50 to 120 kilowatts. The 2027 “Kyber” / Rubin Ultra rack is designed for ~600 kilowatts, on the way to a megawatt-plus. A megawatt. Into a single cabinet.
So before anything else, answer the question everyone skips: why is NVIDIA cramming more and more power into a rack in the first place? Because that’s where the performance is now - and understanding that is the whole ballgame.
For decades you got a faster computer by buying a faster chip. Moore’s Law is crawling; you can’t buy your way to much more performance one chip at a time. So performance moved up a level. The unit of compute stopped being the chip. It became the rack - and then the entire data center.
Here’s the key idea, and it’s simple. A training cluster wants to behave like one enormous brain - thousands of GPUs acting as a single machine, talking to each other constantly at staggering speed. And the enemy of that is distance. Every inch a bit has to travel costs you time, bandwidth, and power - and it costs more the farther it goes. A signal that crosses a rack is fast and cheap. A signal that crosses a room is slow and expensive. Spread your GPUs out and your “one giant brain” turns into a thousand chips waiting on each other.
So the entire trick is to cram as many GPUs as physically possible into the smallest possible space, wired together by the shortest, fattest links you can build. NVIDIA’s NVL72 put 72 GPUs into one rack and made them act like a single chip. Rubin packs in more. Rubin Ultra, more still. NVIDIA calls this “scale-up,” and it has one unavoidable consequence: more GPUs in the box means more power in the box. The kilowatts go up because performance goes up when you pack tighter. Density isn’t a side effect of the roadmap. Density is the roadmap.
And it’s NVIDIA’s roadmap specifically because NVIDIA owns the rack. It sets the architecture, it picks the cadence - a new, denser generation every single year - and it drags the supply chain along behind it. When the company that defines the system says “next year’s rack is 600 kilowatts,” the power industry’s only job is to figure out how to feed it. Which brings us straight to the wire.
Now, why feeding it forces 800V - and this is what makes it different from almost everything else you’ll read about in AI. Most of the buildout is about doing things better. This is about doing the only thing physics allows.
Here’s the problem, and it’s just physics - one equation’s worth.
Power = Voltage × Current (P = V × I).
The power a megawatt rack demands is fixed. So you get a choice: deliver it as high voltage and low current, or low voltage and high current. Today the industry runs at low voltage - which means brutal current. And current is the enemy. Push a megawatt at today’s voltage and the copper you’d need is the size of a fire hose, it runs hot, and you waste a shocking amount of the power just moving it. It doesn’t get expensive. It becomes impossible.
Picture the copper wire as a single narrow highway, and the power you need to move as cargo that has to get down it. You can ship that cargo two ways. Send a thousand little motorcycles - that’s high current, low voltage. The road jams, they slam into each other and into the guardrails, and most of the energy burns off as friction and pile-ups. That friction is the heat - the heat that wastes your power and threatens to melt the cable. Or send two giant trucks - high voltage, low current. Each one hauls enormous power, but with only two vehicles on the road they glide straight through: no friction, no traffic, the highway stays quiet and cool. There’s a one-line equation behind the whole thing, and it’s the ballgame: the power lost as heat in a wire is P(loss) = I² × R - current squared, times resistance. The square is everything. So a megawatt sent as motorcycles isn’t a thick wire. It’s a wire cooking itself.
The fix isn’t a better wire. It’s to stop sending motorcycles and start sending trucks - stop fighting the current and raise the pressure. So you raise the voltage to 800 volts of DC, delivered straight to the rack. Watch what the square does for you: going from 48V to 800V is about 16× the voltage, so about one-sixteenth the current - and because heat is I²R, the heat doesn’t fall 16×, it falls by roughly 16², more than 250×. Same power, a quiet, cool highway. And note the DC part, too - not just AC at a bigger number. Going direct-current end to end lets you stop flipping power back and forth between AC and DC, which is the other half of the win (more on that in a second). (Purists will note the industry is split on the exact number - NVIDIA pushes 800V, the big hyperscalers are also standardizing a ±400V flavor - but the direction is identical. The number is a fight for the standards bodies; the migration is not in question.)
So what does it unlock? Three things.
Density. It’s the only way to physically feed a megawatt rack - which, per everything above, is the only way to keep scaling performance. No 800V, no next-generation AI factory. Full stop.
Efficiency, and it comes from two places. First, the wire: because loss is I²R, carrying the same power at higher voltage (lower current) crushes the heat - NVIDIA says the copper itself can shrink by up to ~45%. Second, the conversions: going DC end to end lets you stop rectifying and inverting power over and over. Today’s path flips between AC and DC several times - utility AC, a double-conversion UPS, the power supply, the board - and every flip leaks. 800V HVDC rectifies to DC essentially once at the front and distributes DC the rest of the way, collapsing the chain from about four conversion stages to two and lifting end-to-end efficiency from roughly 83% to 92%+ - a gain of over 5%. Five percent sounds small. At a gigawatt campus, 5% is a power plant. And when the binding constraint on AI is how many megawatts you can even get to a site, every watt you don’t waste is another watt of compute you get to sell.
Better AI economics. More of every watt reaching the chip means more tokens per watt, which means cheaper AI over time. Power efficiency and inference cost are the same conversation now.
That’s the chain, start to finish: performance now comes from packing GPUs tighter → tighter means more power per rack → more power per rack breaks low-voltage physics → so the power has to go to 800V. The GPU got the glory, but the wire became the bottleneck, and rewriting how power reaches the chip is now mandatory, not optional.
Which raises the obvious question: if you’re converting messy grid power into 800 volts of clean DC, what will you use for that? It’s definitely not the 130-year-old transformer made of copper and iron that buzzes at 50 hertz in every substation today. That system is too large, too slow, and far too outdated to handle a load that can shift from idle to full speed in microseconds. The answer is the solid-state transformer (SST): a transformer rebuilt out of semiconductors, switching hundreds of thousands of times a second - small enough to fit, fast enough to keep up, smart enough to run in software. And the only reason it works is two materials most investors have never traded: silicon carbide (SiC) and gallium nitride (GaN) - the wide-bandgap chips that shrug off the voltage and heat ordinary silicon can’t.
That’s Part 2: the box that makes 800V real, and the supply chain racing to build it. Part 3 is where we’ll talk about where the value is actually created - who captures it, and a question I’ll try to answer: what really happens to copper, and what (if anything) this has to do with photonics.
The chip was never the whole story - the power always was.