Nscale Filed to List at $35B on a $1.02B Loss — the 5 Numbers That Made Me Rewrite Our Singapore Engineering Budget

I run delivery teams, not a fund, and I would normally skip an IPO filing. I did not skip this one, because an S-1 is the only document in this industry where a compute supplier has to show you its actual economics, and because half the Singapore engineering leaders I work with are about to sign multi-year AI infrastructure commitments in their 2027 budgets. TechCrunch’s piece on 22 September put the word that matters in the headline: concentrated. Here are the five numbers I pulled out, and the three lines in our own budget they changed.

What Was Filed, and Why an S-1 Is Worth Your Time

Nscale is a London-based AI infrastructure developer. On Friday 18 September 2026 it filed a Form S-1 with the US Securities and Exchange Commission, planning to list on the New York Stock Exchange under the ticker NSCL, reportedly seeking around three billion dollars at a valuation near thirty-five billion, with Goldman Sachs, JP Morgan and Morgan Stanley leading the book. CNBC reported the filing the same day; TechCrunch’s analysis followed on 22 September.

The reason to read it has nothing to do with whether you would buy the stock. A registration statement is the one moment when a company in this layer of the stack must disclose, under liability, what its revenue actually is, what it is losing, who its customers actually are and how concentrated that book is. Your cloud provider’s sales engineer will not tell you those things. For anyone in Singapore whose 2027 roadmap assumes a certain price and availability of accelerated compute, that is a free and unusually honest look at the foundation.

The 5 Numbers

1. A $35 billion target valuation on $140.6 million of half-year revenue

For the six months ending 30 June 2026, Nscale reported revenue of 140.6 million dollars. The valuation being sought is around thirty-five billion. Whatever else that multiple is, it is a bet on the contracted future rather than the trading present — which is fine, and normal in infrastructure, and exactly why the next three numbers matter more than this one.

2. A $1.02 billion loss — against $368.9 million a year earlier

The same six months produced a net loss of 1.02 billion dollars. In the comparable period of 2025 the loss was 368.9 million on revenue of 10.4 million. So revenue grew roughly thirteen-fold year on year, and the loss grew roughly three-fold. The growth is real and the burn is real, and both are what building data centres looks like. For a customer, the operative question is simpler than for an investor: how long is this funded, and what happens to my capacity if the financing window closes?

3. A $103 billion contract book

Nscale has amassed over 103 billion dollars in total contract value. Read against 140.6 million dollars of half-year revenue, that is the entire thesis in one ratio: almost all of the value is contracted future delivery, not delivered service. The order book is the asset.

4. About 85 percent of that book is two customers

This is the number that changed my budget. Roughly 85 percent of the contract value comes from two agreements: about 43.8 billion dollars of compute supplied to Microsoft through 2033, and a further 44.6 billion dollar supply agreement with Anthropic. That is the concentration TechCrunch named. It cuts in two directions at once, and both matter to a buyer.

5. 25,000 active GPUs against 461,000 active and contracted

As of 31 August 2026 the portfolio included approximately 25,000 active GPUs, with 461,000 active and contracted, across five active and twelve contracted data centre sites. The gap between what is running and what is promised is roughly eighteen-fold. Every capacity conversation you have with any provider in this market over the next eighteen months sits somewhere inside that gap.

3 Expert Takes on What a Singapore Engineering Leader Should Do

Take 1 — Concentration is not automatically bad news for you. It is information about your queue position

The instinct when you read “85 percent of the book is two customers” is to treat it as a warning. It is more useful as a map. A supplier whose economics rest on two enormous anchor agreements is a supplier whose engineering attention, capacity allocation and roadmap will follow those two customers. If you are a Singapore enterprise buying a few hundred GPUs, you are in the roughly fifteen percent band, and you should plan on the service levels, the support responsiveness and the capacity guarantees that band actually gets — not the ones in the sales deck.

The flip side is real too: anchor customers of that size de-risk the build-out that gives you capacity at all, and they impose engineering discipline that smaller customers benefit from. I am not arguing against buying from concentrated suppliers. I am arguing that “where do I sit in this book” is now a question you ask out loud in procurement, and that the answer belongs in your risk register rather than in a footnote.

Take 2 — Portability is a hiring line, not an architecture slide

Everyone says they want workload portability. Almost nobody funds it, because it has no owner. The S-1 makes the cost of not funding it concrete: if the compute layer is this capital-intensive, this concentrated and this early in delivering against its promises, then the probability that you will want to move a workload within three years is not small.

Moving is an engineering capability, and capabilities are people. It means someone who keeps the training and inference stack close to portable interfaces rather than idiomatic to one provider, who maintains a tested path to a second environment, and who can tell you in engineer-weeks what a migration costs before you need the number. In Singapore that profile overlaps heavily with the platform and DevOps engineers we place for accelerated workloads — our guide to hiring DevOps engineers for GPU cloud covers the screening, and the broader shape is in how to hire AI infrastructure engineers in Singapore. The budget change I actually made was here: one role moved from “nice to have in H2” to funded in Q1.

Take 3 — The scarcest person in this market reads a contract and a dashboard

The eighteen-fold gap between active and contracted GPUs tells you the whole sector is selling against capacity that is not yet running. That is not a scandal; it is how infrastructure is financed. But it means the commitments you sign in the 2027 cycle will contain assumptions about delivery timing, and the person who can evaluate those assumptions needs to be fluent in two languages at once: what the contract says, and what your utilisation dashboard says.

Almost every Singapore engineering org I work with is short of this person. They have excellent infrastructure engineers who do not read commercial terms, and excellent finance people who cannot tell whether a committed-use discount is being consumed by real training runs or by an idle reservation someone forgot. The role sits between them, it is usually created by promoting a senior platform engineer, and it pays for itself the first time it catches an over-commitment. If you fund one thing out of this filing, fund that. The same argument in the Gulf, with the regional cost figures, is in our colleagues’ UAE AI engineer cost guide, and they reached it from the opposite direction when Crusoe’s valuation round raised the same questions in Dubai.

The 3 Budget Lines I Actually Changed

To be concrete, and to avoid the newsletter habit of drawing enormous conclusions from one filing, here is what moved and what did not.

1. Portability moved from aspiration to a funded role. One platform engineer in Q1 whose explicit remit includes keeping a tested second path for our heaviest workloads. Previously this was everyone’s job, which meant it was nobody’s.
2. FinOps literacy became a requirement, not a bonus. We added committed-spend modelling to the job description for our next senior platform hire, and a screening question that asks a candidate to talk through a reservation that was being under-consumed.
3. One question was added to every vendor conversation: what share of your contracted capacity is currently active, and where does our workload sit in your delivery sequence? We have asked it four times since Monday. Two suppliers answered precisely. That itself was useful information.

What did not change: headcount, roadmap, or our choice of provider. One filing from one company is not a reason to re-plan a year of engineering, and anyone telling you otherwise this week is selling something. It is a reason to make three lines of an existing plan sharper, which is roughly the right size of response to a single piece of news.

The broader point for Singapore is that the compute layer has become financially legible in a way it was not eighteen months ago. Filings like this one, and the disclosure that comes with a public listing, mean engineering leaders here can finally reason about their suppliers with numbers rather than vibes. That is worth an afternoon of your planning cycle, even if you never buy a share.

FAQ — The Nscale Filing and Singapore Engineering Budgets