Nscale, the AI infrastructure company, filed a 192-page S-1 on 18 September 2026 for a US listing on the New York Stock Exchange, reportedly seeking a valuation of up to $35 billion. The filing shows $33 million of revenue in 2025, $140.6 million in the first half of 2026, a net loss of $1.02 billion for those six months, and total contract value of $103.4 billion at 31 August. It does not name its largest customer. The Financial Times worked that out from an exhibit to an earlier filing. This article explains how a customer can be identified without being named, why a contract backlog is not revenue, and what a reported construction delay would change. Figures come from the Fortune, The Next Web and Outpost reports of the filing; we did not read the S-1 itself.

What the filing shows

The headline numbers are consistent across the outlets we read. Revenue was $33.0 million in 2025 and $140.6 million in the first half of 2026. The net loss for the first half of 2026 was $1.02 billion, about 7.3 times the revenue in the same period. The valuation target of up to $35 billion is reported by the Financial Times, and the S-1 leaves the offering size and price blank. For context, Crypto Briefing reports the company's Series C in March 2026 was at $14.6 billion, so the target is roughly 2.4 times that valuation six months later.

The backlog is larger than anything in the income statement. Total contract value was $103.4 billion at 31 August, which Crypto Briefing summarises as over $100 billion. Outpost reports it includes a deal with Anthropic worth $44 billion and a $43.8 billion contract with Microsoft running through December 2033; Fortune puts the Anthropic figure at $44.6 billion. Taking the two reported figures together, those two customers account for roughly $88 billion, about 85% of the total, by our arithmetic. That concentration in the backlog is a separate fact from the concentration in revenue, which we turn to next.

An unnamed customer who can still be identified

The S-1 does not name the company's biggest 2025 customer. The Next Web reports that neither the prospectus nor the loan exhibit mentions ByteDance, and that the prospectus does not name the top customer. The identification came from a different document. The Financial Times found a $105 million loan agreement, with Macquarie Bank as agent, filed as an exhibit to Nscale's draft registration statement in September 2025. The loan, first signed on 12 June 2025 and restated on 22 September 2025, finances 2,304 Nvidia B200 chips at Nscale's data centre in Glomfjord, Norway. It names a cloud services contract with Spring (SG) Pte. Ltd., which the FT identifies as a Singapore subsidiary of ByteDance; Fortune cites a Forbes Australia link for the subsidiary relationship.

That is the mechanism worth understanding. A prospectus can be required to disclose customer concentration without naming the customer, but a lender that is financing hardware against one customer's contract often needs the contract described, and loan exhibits are filed with the registration statement. The two documents were written for different readers and with different disclosure rules, and a careful reader can join them. Nscale declined to comment on the relationship, according to Outpost.

The share is reported two ways. Spring accounted for $24 million of $33 million in 2025, which is about 73%; the FT's headline, as relayed by The Next Web, says nearly 75%. Crypto Briefing's headline says 75% and its body 73%. We use 73%, which is the figure the arithmetic supports. Separately, the S-1's first-half 2026 figures show the largest customer, unnamed, providing 52% of revenue, according to Fortune. Nothing we read says whether that customer is ByteDance, so we do not assume it is. Outpost mentions one reading of the prospectus that gives the same 52% figure.

According to Tom's Hardware, as relayed by The Next Web, the lender required Nscale to monitor how Spring uses the chips and to report compute anomalies or suspicious configurations, and Nscale had a third party carry out due diligence on ByteDance and Spring. We did not read Tom's Hardware directly. If accurate, it shows the customer-specific conditions the financing carried, which is the kind of detail that rarely appears in the main prospectus text.

Backlog is not revenue: two measures of the same contracts

The $103.4 billion figure is total contract value, which counts everything customers have committed to pay over the life of their agreements. Fortune also reports a different measure from the filing: $56.4 billion in remaining performance obligations, an accounting term for contracted revenue not yet recognised. The sources we read do not explain the $47 billion difference between the two. Our reading of the terms is that remaining performance obligations are a narrower, accounting-defined subset, but we could not verify what the gap consists of. A reader sizing the company should be aware that two backlog numbers exist and that the larger one is the one in the headlines.

Neither is revenue. For a data-centre operator, contract value turns into revenue only when capacity is built, powered and handed over to the customer. A contract signed for a data centre that has not been energised pays nothing. The arithmetic shows how large the conversion gap is. The $103.4 billion total contract value is about 735 times the $140.6 million of first-half revenue. The $56.4 billion of remaining performance obligations is about 400 times it. These are our calculations, and they make a point of scale, not a prediction: they do not say how fast conversion will occur, only how much of it is ahead.

The valuation arithmetic

A $35 billion valuation is about 1,060 times 2025 revenue of $33 million. That multiple is not a useful pricing guide, because the company is in the early part of a build-out and 2025 is a base year. A fairer reference is the first half of 2026 annualised: $140.6 million doubled is $281 million, and $35 billion divided by that is about 124 times. That is still a very high multiple, and the figure assumes the second half merely matches the first, which, for a company in a build-out phase, understates the likely revenue and overstates the multiple. We do not benchmark this against other infrastructure companies, because we did not source peer valuations.

The bull case is that contracted revenue is the point. If $56.4 billion of obligations convert over the next several years, as the Microsoft contract's 2033 horizon suggests, the revenue base is vastly larger than today's. Investors are buying the conversion. The cost of that conversion is also in the numbers: a $1.02 billion loss in six months on $140.6 million of revenue shows how much is being spent ahead of revenue.

The reported Norway delay

A summary published by aiweekly on 16 October relays a Bloomberg report that Nscale's Norway data centre is about six months behind schedule, with 6.4 megawatts of a planned 30 running, and that the delay brings penalties to Microsoft and ByteDance. We could not find the Bloomberg report itself and the Fortune article we read mentions no delay, so we treat this as an unconfirmed secondhand report. It matters because of the conversion mechanism above: if capacity is late, revenue is late, and customer contracts may carry penalty terms that reduce what Nscale earns. The reported figure of 6.4 of 30 megawatts would mean roughly a fifth of the planned capacity was operating. If the report is accurate, it is the clearest illustration of why backlog and revenue differ. If it is wrong or out of date, it should not be relied on.

The counter-argument

A defender of the offering would say that customer concentration is normal for a company at this stage and is falling. The two biggest contracts are with Anthropic and Microsoft, customers with large budgets and public commitments to AI capacity, and the share of revenue from the single largest customer has dropped from about 73% in 2025 to 52% in the first half of 2026. On that view, the filing describes a company moving from one early customer to a diversified base of very large ones.

The sceptical response is that the backlog is concentrated in the same way. About 85% of contract value sits with two customers, so the revenue base would be exposed to those two contracts even as the ByteDance share shrinks, and the customers' own plans, financing and willingness to enforce terms matter as much as Nscale's execution. A spokesperson's comment on ByteDance would not settle that, and Nscale declined to comment.

What would change the picture

  • A priced offering, which would replace the reported valuation target with a real one.
  • Quarterly disclosures of capacity energised against capacity contracted.
  • Disclosure of penalty and delay terms in the Microsoft and ByteDance contracts.
  • An explanation of the gap between total contract value and remaining performance obligations.
  • Confirmation or denial of the Bloomberg-reported Norway delay.