Nscale Is Going Public on $103 Billion of Contracts. Two Customers Are 85 Percent of It, and One Can Walk.

Nscale filed its S-1 on Friday, September 18, 2026. London-based AI data center developer, New York Stock Exchange, ticker NSCL, with Goldman Sachs, J.P. Morgan and Morgan Stanley at the top of a 23-underwriter syndicate. Share count and price range are both still blank, which means the number everyone wants gets set in the next amendment.

The number everyone is already quoting is $103 billion of contracted order book. Bloomberg and others spent the weekend pulling the same detail out of the filing: roughly 85 percent of that book belongs to two customers. A $43.8 billion Microsoft agreement running through 2033, and a $44.6 billion Anthropic agreement announced on August 26.

Every write-up I have read frames this as a customer concentration story. It is, but that is the least interesting risk in the document, and it is the one the market already knows how to price. The thing worth your attention is what those two contracts are conditioned on.

Start With the Arithmetic

Put the backlog next to the income statement and the scale mismatch stops being abstract.

Revenue grew 1,252 percent year over year, which is a genuinely enormous number and also what happens when you start from $10.4 million. Annualize the first half and you get roughly $281 million. The $103 billion book is about 366 times that. The $56.4 billion of remaining performance obligations, which is the more defensible figure because it is an accounting disclosure rather than a marketing one, is about 200 times it.

The loss line deserves its own sentence. Nscale burned $1.02 billion to produce $140.6 million of revenue, which is roughly $7.25 of loss per dollar booked, and the loss grew faster than the revenue did in absolute terms. That is not automatically damning for an infrastructure build, where the capital goes in years before the cash comes out. It does mean the company is going public specifically to fund the gap.

The Conditionality Runs the Wrong Direction

Here is the part of the filing that changed how I read the rest of it.

The Anthropic agreement is contingent on Nscale obtaining financing, and Anthropic retains the right to walk away or cancel if Nscale fails to hit milestones the prospectus itself categorizes as stringent. Read that next to the reason Nscale is listing. The backlog is the asset being used to raise capital. The capital is the condition on which the largest single piece of the backlog survives.

Normal customer concentration risk is one-directional. You have too few customers, one of them leaves, revenue drops. Investors price that with a discount and move on. What is in this document is a loop: the contract underwrites the financing, the financing sustains the contract, and a bad print on either side pressures the other. It is the same structural feature that has made the whole neocloud category hard to value, showing up here in unusually explicit language.

I want to be fair about this. Milestone-contingent offtake is completely normal in project finance, and a customer committing $44.6 billion through 2033 is entitled to ask that the facility actually get built. Anthropic is not being unreasonable. But an investor buying NSCL is buying a claim on a cash flow whose largest component is conditioned on an outcome that the investor's own purchase is meant to produce.

Microsoft Is Both Anchor and Warning

The West Virginia history is the detail that has been underplayed. Anthropic's 460 megawatts of Nvidia Vera Rubin capacity, expected to begin serving traffic in late 2027, sits at a development Microsoft exited over the summer during a review of its data center portfolio and its share of on-site power generation. Semafor reported that both Google and Microsoft were in talks for compute at the site before the deal went to Anthropic.

Microsoft is simultaneously the largest name in the order book, at $43.8 billion and 1.35 gigawatts, and the party that walked away from a specific project at this specific developer within the last few months. Those are not contradictory decisions. A hyperscaler reallocating a site while holding a separate multi-year commitment is ordinary portfolio management. It is still worth noticing that the counterparty whose name does the most work in the valuation has already demonstrated, at this developer, that it will step back from a project when its own numbers change.

Where This Sits Against the Comparables

Concentration is the defining feature of the category, not an Nscale defect. The useful question is whether NSCL is worse than the names already trading.

CoreWeave is the precedent everyone will reach for, and it cuts both ways. It went public with Microsoft at roughly two-thirds of revenue, the stock was volatile, and the concentration did in fact decline as OpenAI, Meta and others came on. Sell-side models now have Microsoft dropping toward the high thirties as a share of CoreWeave revenue. The bull case for NSCL is simply that this is the same movie: anchor tenants underwrite the build, then the customer list broadens.

The bear case is that Nscale is starting from a narrower base than CoreWeave did. The filing discloses a single buyer at 52 percent of revenue in the first half of 2026, 73 percent the year before, and effectively all of it the year before that. That is improvement, and it is improvement from a starting point of one.

Nebius is the counterexample worth holding in mind, because it has built a comparatively diffuse customer base, is carrying materially cheaper debt than CoreWeave, and has taken a different route on owning its own software layer. It also has a fraction of the contracted scale. Nobody in this category has yet shown you can have both.

Caveats Worth Naming

Three things cut against reading this too darkly. First, backlog conversion at 2.5 percent is what you would expect for contracts whose capacity does not come online until late 2027 and runs to 2033. Nobody should be alarmed that a facility not yet built is not yet billing. Second, the S-1 is a risk document by construction. Securities counsel is paid to write the concentration and contingency language in the most unflattering available form, and the presence of a scary sentence is not evidence that the scary thing is likely. Third, the underwriting syndicate is real. Three bulge-bracket leads and 23 books is not the shape of a deal nobody wants.

I will also note that terms are not set. Reported valuation talk has ranged from a $25 billion figure floated by bankers before the filing up toward $30 billion and above in press reports. None of that is in the prospectus, and the first amendment with a real range is the document that matters.

Our Take

The interesting fact is not that two customers hold 85 percent of the book. It is that the $103 billion headline and the $2.6 billion of active contract value are being reported as though they describe the same asset, and they do not. One is a schedule of intentions with conditions attached. The other is a business.

What Nscale is actually asking the market to underwrite is execution risk dressed as revenue visibility. The order book does not tell you whether the company can build 1.35 gigawatts for Microsoft and 460 megawatts for Anthropic on time, secure the power interconnects, finance the Vera Rubin deployments, and hit milestones stringent enough that the prospectus flags them. It tells you what happens if it does. Those are different questions and only one of them has a number next to it.

There is also a read here about the buy side of the compute market that has nothing to do with Nscale. Anthropic picked up a project Microsoft put down, at 460 megawatts, on a chip generation that does not ship at volume until next year, with a cancellation right attached. That is a lab buying optionality on capacity rather than capacity, and it is the third or fourth time this year a frontier lab has structured a compute agreement so that the developer carries the construction risk. The labs have learned something about this market that the neoclouds are still absorbing.

Practical read if you are budgeting inference capacity rather than trading the stock: nothing in this filing changes a price sheet this quarter. What it changes is your 2028 planning assumption. Two of the largest capacity commitments in the industry are attached to a company that has not yet built them and is raising the money to try, so treat announced gigawatts as a forecast rather than as supply, and discount any 2027 or 2028 capacity promise that depends on a facility currently existing as a financing milestone.

Three signposts for the next 60 days. First, whether the pricing amendment lands above or below the $25 billion bankers floated, because that spread is the market telling you what it thinks conditional backlog is worth. Second, whether Nscale discloses a third anchor customer before pricing, which is the single fastest way to reframe the concentration narrative. Third, whether any subsequent filing quantifies the Anthropic milestones, because right now the most important term in the largest contract in the book is described with an adjective.

We track provider capacity and pricing moves on the model pricing board and live availability on the status board. Neither will show you a gigawatt that does not exist yet, which is rather the point.