Anthropic's $45 billion 'cloud' deal is really a bet on 460 megawatts. Do the division.

Strip the word 'cloud' out of the announcement and the deal gets easier to read. Anthropic has reportedly agreed to pay Nscale about $45 billion over six years, and what it is buying, underneath the software language, is 460 megawatts of electricity turned into computation at a single campus in West Virginia. That is the number that matters: 460 megawatts. Everything else in the sentence — the chips, the racks, the cooling, the six-year term — is downstream of whether that much power can be delivered, continuously, to one place. The models are the headline. The megawatts are the deal.

The terms, as reported: roughly $45 billion, six years, about 460 megawatts drawn from an Nscale facility in West Virginia, running on Nvidia's forthcoming Vera Rubin chips. The figures come from CNBC, Bloomberg and TechCrunch, each citing people familiar with the talks, so treat them as reported rather than filed. The site is expected to come online around the end of 2027. Anthropic has been signing agreements like this at a pace that has stopped being remarkable — Google TPUs, Amazon, and now a British company most people have never heard of — all of it ahead of a planned IPO. I want to do the one thing these announcements are built to discourage, which is arithmetic.

Price it per kilowatt-hour

Take the deal at face value and divide. 460 megawatts, running flat out, is about 24 million megawatt-hours of electricity over six years. Put $45 billion over that and you land near $1,850 per megawatt-hour — call it $1.85 a kilowatt-hour. Now hold that number next to what electricity actually costs. A large industrial buyer in the United States pays on the order of $60 to $90 a megawatt-hour; wholesale power in a cheap market runs lower still. The compute contract, priced as if it were an electricity bill, comes out more than twenty times the price of the electricity itself. That gap is the whole story, and it points in a direction most of the coverage walks straight past.

The $45 billion is not a power bill. Overwhelmingly it is silicon and the shed that holds it — Vera Rubin systems, which stack six chips into a single package and sit at the top of what is currently manufacturable, plus the cooling, the networking, the construction and Nscale's margin. Priced honestly, the raw electricity to run 460 megawatts for six years is on the order of a couple of billion dollars: low single digits as a share of the contract. The electrons are the cheapest line item in the deal. And yet the electrons are the binding constraint, because you can always order more Rubin racks and you cannot always find 460 megawatts of firm power with somewhere to plug it in. That inversion — cheapest input, scarcest input — is the fact this industry keeps rediscovering.

The electricity is the cheapest line item in the deal, and the only one you cannot buy more of on demand. That is why it decides everything.

Nscale is selling a grid connection

Nscale is worth understanding, because what it is tells you what Anthropic is actually paying for. It is a London-based company founded in 2024, spun out of a Bitcoin miner called Arkon Energy by a former miner, Josh Payne. Its most recent round valued it around $14.6 billion, it is backed by Nvidia, and it is preparing an IPO. The crypto-mining lineage is not a piece of trivia here — it is the qualification. Bitcoin miners spent a decade becoming expert at exactly one thing: finding cheap, abundant power and standing computers on top of it, quickly, in places no one else wanted to build. That is the identical problem an AI data center solves. The neoclouds that matter are, disproportionately, miners who swapped the workload and kept the skill.

Nscale's West Virginia campus — it calls the site Monarch — is a 2,250-acre parcel it acquired in early 2026, with a claimed path to as much as 8 gigawatts over time and a first phase measured in the low gigawatts. Notice where it is. West Virginia is coal country, with a grid built for heavy industry that has spent forty years leaving, which means spare interconnection capacity and a political system extremely willing to permit new load. Anthropic is not really buying a building in Appalachia. It is buying a place on the grid — a firm, permitted, high-voltage connection — and the guarantee that someone else will absorb the years of construction and utility negotiation between now and 2027. The scarce asset is the interconnection queue, and Nscale is selling its place in line.

What 460 megawatts is

460 megawatts is enough electricity to power roughly 345,000 American homes at once. One company, one campus, drawing continuously what a mid-sized city draws, to serve a single AI lab's models. And 460 is the contracted figure for the computers; the real load on the local grid is higher, because every watt pushed into a chip needs a fraction of a watt more to carry the heat back out. This is the arithmetic that keeps surfacing as moratoria and rate fights from Virginia to the grids I spend most of my time reporting on across the global South: the load is real, it is large, and it lands on a system other people share. What is different in West Virginia is that the state has decided it wants the load. Whether the households on the same wires end up quietly underwriting the connection is the question that tends to arrive about eighteen months after the ribbon-cutting.

Here is the part that decides whether deals like this keep happening. The industry's bet is that firm power at this scale is available if you are willing to go where it is and lock it up six years at a time. That bet works in West Virginia, and in a handful of other places with stranded grid capacity and willing regulators. It does not obviously scale to a hundred more campuses, because there are not a hundred more Appalachias sitting on 460 spare megawatts with an interconnection ready to energize. The frontier labs are, in effect, racing to claim the last easy megawatts before the rest of the market prices them in. After that, the constraint stops being money and becomes physics and permitting — the two things a balance sheet cannot hurry.

For Anthropic, the six-year term is the point. Committing $45 billion of future compute ahead of an IPO does two jobs at once: it guarantees the capacity the company needs to keep training, and it tells prospective investors the capacity is already secured, which for an AI lab is a large share of the story it is selling. The risk sits on the other side of the table, with Nscale, which has to build the thing, and with whoever finances that construction against a single anchor tenant's promise to pay. That is a familiar shape from other capital-heavy build-outs — fiber, pipelines, merchant power — and it usually holds right up until the anchor tenant's demand forecast turns out to be wrong.

The number that decides it

So price the deal the way I would price any of them. Not by the $45 billion, which is mostly silicon and will be quoted breathlessly and then quietly restructured. Not by the Vera Rubin chips, which will be superseded before the concrete is dry. Price it by the 460 megawatts — the cheapest input and the only one nobody can conjure on demand. If Nscale can deliver firm power at that scale, on time, off a repurposed coal grid, the deal is sound and there will be twenty more like it. If it cannot, the chips sit in crates and the $45 billion becomes an expensive lesson in the difference between ordering computers and powering them. Everything Anthropic is buying runs on the one input the contract prices most cheaply and can least afford to get wrong. Watch the megawatts. The rest is commentary.