The Power Block

Now let's talk about what we actually build. We call it the Power Block. Think of the Power Block as a complete, self-contained power plant in a standardized package. Instead of designing every site from scratch — which is slow and expensive — we engineered one repeatable building block that we can stamp out and deploy again and again. It contains the fuel cells, the electrical equipment, the controls, the cooling, and the structure, all integrated and pre-engineered to work together. This is the same insight that made shipping containers transform global trade. Before containers, every cargo load was a custom puzzle. After containers, everything became standardized, stackable, and fast. The Power Block does that for on-site power. Standardize the unit, and suddenly you can build faster, cheaper, and more predictably — and you can scale across many sites without reinventing the wheel each time. So let me actually walk you through one, component by component, in the order the power flows. Follow the electricity from where it's made all the way to the chip, and you'll understand the whole machine.

Step one: making the power. It starts with the fuel cells. In one block we have fifty of these solid-oxide fuel cells, and each one produces about 325 kilowatts of electricity. Group ten of them together and you get a unit we call a block, producing about 3.25 megawatts. We run four of these blocks actively, plus one extra sitting in reserve as a hot spare — so if anything ever needs servicing, the spare instantly carries the load and the customer never notices. That gives us roughly 13 megawatts of solid, dependable power per Power Block, with no dependence on the grid at all. The fuel cells run on natural gas, and remember: they produce direct current, DC, right from the start. That fact shapes everything downstream.

Step three: the backbone. All that power gathers onto what we call the backbone — a single, common electrical highway running at 400 volts DC. This is the spine of the entire Power Block, and here's the key thing to hold onto: the backbone never changes. It's the same in every configuration, for every customer. It's the expensive, permanent, disruptive infrastructure — the part you absolutely don't want to be rebuilding. Everything I've described so far — the fuel cells, the per-cell breakers, the block breakers, the backbone — is identical no matter what kind of customer we're serving. That sameness is the secret to how we move fast.

Step five: the busway — distributing it across the room. From the cassette, power travels down a busway. Picture an overhead power rail running the length of the room, like a track that equipment can tap into anywhere along its length. This lets us add, move, or reconfigure rows of computers by simply tapping into the rail, instead of rewiring the building each time. One clever detail: our DC busway is dual-rated, meaning the same physical rail can carry either 400-volt or 800-volt DC. So upgrading a room from one DC voltage to the other doesn't require ripping out and replacing the rail.

Step seven: keeping it cool. All that power becomes heat, and heat is the enemy of computers. So woven right into the Power Block is a cooling system that runs liquid directly to cold plates sitting on the chips themselves — pulling heat away right at the source, which is far more effective than just blowing cold air around a room. And here's a point I love: our cooling is engineered to use almost no water. We measure this with something the industry calls water-use effectiveness, and ours is so low we describe the system as drought-immune. In a world fighting over water, that's a genuine differentiator, and it's built right into the block rather than bolted on beside it. So that's the whole journey: fuel cells make DC power, nested breakers protect it, a common 400-volt backbone carries it, a swappable cassette shapes it, a busway distributes it, the White Hall consumes it, and an integrated water-light cooling system keeps it all from overheating. Same backbone every time; only the cassette and the gear past it change. We hold patents on the architecture of this whole assembly, because the specific way we've integrated it is genuinely novel.

Now here's the piece that makes the whole platform flexible, and it's my favorite to explain. Different customers, and even different parts of the same campus, need power in different forms. Some equipment wants traditional A.C. Some wants 400-volt DC. Our newest, most efficient setups want 800-volt DC. In a normal world, that means three completely different designs, three different supply chains, three different headaches. We solved it with what we call cassettes. Picture the cassette as the one module that sits between the unchanging backbone and the customer's racks, and determines what kind of power comes out. The big expensive structure — the fuel cells, the protection, the backbone — stays exactly the same. Only the cassette, and the gear right past it, changes. It's like a high-end camera with interchangeable lenses. The camera body is the expensive, permanent investment. The lens determines what kind of picture you get, and you change it based on what you're shooting. Our cassettes do that for electricity. We have three of them, and they're worth understanding because each serves a different kind of customer.

The second is the pass-through cassette, and it's the simplest of all. Here the customer wants power at the same 400-volt DC that the backbone already produces. So there's nothing to convert — this cassette is essentially a smart, monitored connection point. It just adds protection and metering and passes the power straight through. It's the lightest, most efficient option, almost like a clean junction box, and it's a natural fit for a family of modern data center equipment built around that voltage.

Now here's the elegant part, the design idea I'm proudest of. Those three cassettes aren't three completely different boxes. They're one common chassis, populated three different ways. Same enclosure, same input stage, same mounting, same connections — and then you populate it with an inverter module to get A.C., or a simple link to get 400-volt DC, or a step-up module to get 800-volt DC. One part on the floor, three configurations. That dramatically cuts the number of distinct components we have to build, stock, and service. I want to be precise about one thing, because it's easy to oversell. Converting a pod from one type to another isn't flipping a switch while everything's running. It's a planned maintenance task: you swap the module inside the cassette, and you match the busway and the rack connection to go with it, and that pod is briefly offline while you do it. Going between the two DC voltages — 400 and 800 — is the lightest change, because that dual-rated busway I mentioned stays right where it is. Going between A.C. and DC is heavier, because A.C. needs its own kind of busway. But none of it — not one path — ever requires rebuilding the facility. That's the whole point.

The Power Block needs something to live in and on, and we engineered that too, rather than buying generic steel and hoping it fits. We designed modular structural systems — we call them the DG Building and DG Rail — that the Power Block components mount into. Think of it as a purpose-built skeleton, designed from the start to hold our equipment, route our power, and let us slide cassettes in and out cleanly. Because it's modular and standardized, it ships and assembles faster than custom construction, and it's another piece we hold patents on.

Let me connect this back to the AI customers, because the way modern AI chips work makes our approach even more valuable. The most advanced AI systems don't use one chip in isolation. They lash together huge numbers of chips that have to talk to each other constantly, at blinding speed, working as one giant brain. Nvidia's technology for wiring these chips together at high speed is a big part of why they're so dominant. These tightly-coupled clusters of chips are incredibly demanding — they need massive, steady, clean, uninterrupted power delivered exactly where the chips are, with no hiccups. That is precisely what our architecture is built to deliver. Power generated on-site, kept as efficient DC, delivered directly to dense clusters of chips, with fault protection that keeps it rock-steady. We didn't design a generic power plant and hope AI could use it. We designed the power platform around the way these AI chip clusters actually behave.