$103bn Contracted, $100m a Quarter Billed: Nscale's Pre-IPO Backlog (September 2026) — and Why the Same Word Means Three Different Things

A two-year-old company is telling prospective investors it has roughly $103 billion of contracted revenue — a figure within a percentage point of the backlog CoreWeave reported to the SEC in August, and about a sixth of Oracle's. Its most recently estimated quarterly revenue is around $100 million. Both numbers can be true at once, and the reason they can is the single most useful thing a reader can take from the AI infrastructure buildout: a backlog is a delivery schedule, not a bank balance, and the gap between them has to be financed by someone before a dollar of it arrives.

What was reported, and by whom

Three separate pieces of reporting stack into one story, and it is worth keeping them apart because their evidentiary weight differs.

On 26 August 2026, CNBC confirmed that Anthropic had struck a roughly $45 billion cloud agreement with Nscale, renting around 460 megawatts of capacity at a West Virginia development running Nvidia's Vera Rubin generation, with the facility expected to come online at the end of 2027. Bloomberg reported it first. The West Virginia governor's office acknowledged the reported agreement the following day, citing the same 460MW figure.

On 2 September, investor materials reviewed by The Information put Nscale's total contracted revenue at approximately $103 billion, up from about $51 billion before the Anthropic contract and further AI-native commitments. The same materials indicated second-quarter revenue estimated at more than $100 million, up from roughly $37 million in the first quarter — and that second-quarter figure carries no Anthropic contribution at all, because the capacity does not exist yet.

On 4 September, Reuters reported that Nscale is in talks to raise about $3.5 billion in pre-IPO funding: as much as $1.5 billion of convertible notes with Third Point set to lead, approximately $2 billion sought from Nvidia, and Goldman Sachs working the process. The notes carry a double-digit discount to the IPO price, adjusted up to a $30 billion valuation. Nscale was valued at $14.6 billion in March 2026 after a $2 billion Series C.

Three documents, three meanings of the word "backlog"

This is the part that generalises well beyond one company. Backlog is not a defined accounting term with a single meaning; what it means depends entirely on the document it appears in.

None of these is dishonest. They are three different points on a disclosure ladder, and a reader who treats them as interchangeable will draw conclusions the evidence does not support. Oracle's figure can be tested against a filed timetable — the mechanics of which we worked through in the Q1 FY2027 preview. CoreWeave's is furnished under the securities laws with its own boundary conditions stated in the same sentence. Nscale's is a marketing number shown privately to people deciding whether to buy shares, and the person describing it said as much.

The four gates between a signature and an invoice

Contracted revenue has to pass through four distinct gates, each of which can slip independently of the others, before it becomes cash.

Gate one is the slow one, and it is why the Anthropic contract contributes nothing to 2026 revenue: the campus is expected online at the end of 2027. Gate two has a public reference point — CoreWeave's second-quarter filing records the industry's first bring-up and validation of Nvidia's Vera Rubin NVL72, the same generation reported for West Virginia, which tells you the hardware exists but is only now being stood up at scale anywhere.

Gate three is where the money is, and it is measurable. Gate four is the one nobody can verify from outside, because it depends on a customer's own funding several years into the future.

What the gap costs, in a filed number

CoreWeave is the useful comparison for one reason: same business model, later stage, disclosure that can be read. Its second-quarter 2026 results, furnished on 11 August 2026, put a price on gate three.

Read the interest line first. CoreWeave paid $640 million of net interest in three months — about 25 cents of every revenue dollar, and more than the $626 million net loss the quarter produced. That is what carrying the gap between a signed backlog and a built data centre costs at scale, and the company funded it visibly: a $3.1 billion syndicated delayed-draw term loan, a $1 billion strategic investment from Jane Street, and more than $10 billion of unsecured debt and convertible bonds in the quarter, including an inaugural Eurobond.

The same quarter added roughly 500 megawatts of active power, taking the total to 1.5 gigawatts. That is the honest order of magnitude for what standing up 460MW in West Virginia will involve — not a figure anyone outside the company can compute precisely, but not a mystery either.

Nvidia on both sides of the same transaction

The loop is real and should be named plainly: a supplier taking equity in a customer whose principal use of that capital is to buy the supplier's product. Nvidia participated in Nscale's Series C and, per Reuters, is being asked for roughly $2 billion more.

The discipline is in the scaling. Nvidia's second-quarter fiscal 2027 results, filed 26 August 2026, show revenue of $96,221 million, data centre revenue of $89.0 billion, gross margin of 75.0% and third-quarter guidance of approximately $108 billion. A $2 billion investment is about 2% of one quarter's revenue. It is not what is holding the top line up.

What matters is the broader version of the same question — how much of the buildout's demand rests on counterparties whose own funding is not yet closed — and Nvidia has already had to number part of it. Its own filing put guarantee exposure at $108.5bn and future commitments at $366bn, which we took apart in the residual-value guaranty piece. Vendor equity in a private neocloud is a smaller instance of the same structure, disclosed later and in less detail because the recipient is private.

Power is the binding constraint, and it is legislated

The physical bottleneck is not silicon. It is megawatts, land and the legal certainty to connect them, and that is settled by state law rather than by markets.

West Virginia's microgrid framework, referenced by the governor's office alongside the reported agreement, provides siting and regulatory certainty for large-load developments and includes provisions designed to keep the associated infrastructure costs from being shifted onto existing residential and business ratepayers; under the law, 50% of the revenue generated through the framework is directed toward reducing the state income tax, with the remainder going to host counties, communities and infrastructure. The market-relevant fact is narrow: arrangements like it shorten the time from signature to energised megawatt, and that timeline is gate one.

Where this touches something you can actually trade

Nscale is private, so there is nothing to trade in it directly. The exposure is indirect, and it runs through three channels of decreasing strength.

The first is index composition. CoreWeave's Q2 release notes its selection for the Nasdaq-100. Anyone holding NAS100/NQ therefore already owns a business with exactly the economics described above — enormous backlog, heavy leverage, interest expense larger than the loss line. A second such listing would change the index's character at the margin rather than transform it, but the direction of travel is one-way for now.

The second is credit. More than $10 billion of unsecured debt and convertible issuance from a single neocloud in a single quarter is a supply event in its own right, and the terms at which the next one prices are a cleaner read on how the market is underwriting these contracts than any equity move.

The third, and the weakest, is the currency channel. This capex is dollar-denominated and dollar-financed, which means its cost is set at the long end of the Treasury curve rather than by the policy rate — the mechanism we traced through the August term-premium selloff. A higher term premium raises the cost of every year of a five-year buildout at once. That reaches the dollar only at one remove, and it would be overreaching to turn a data-centre lease into an FX view. Understanding the channel is the point; the meter scores eight currencies on five factors, and this is one input into one of them.

What would change the picture

Four things, in order of how much they would tell you.

A filed prospectus. The moment Nscale registers publicly, the $103 billion becomes a disclosed figure with a definition, audited financials beside it and, in all likelihood, a conversion schedule. That single document would move the number one full rung up the ladder in the first table above.

Customer concentration. If roughly $45 billion of $103 billion sits with one counterparty, the backlog's risk profile is not the average of its contracts — it is dominated by one of them. A prospectus would have to address it.

The financing terms. Whether the pre-IPO raise closes at $3.5 billion, and at what discount, is a live read on how much appetite exists for this structure at this point in the cycle.

Delivery dates. End of 2027 is the stated target for the West Virginia campus. Every quarter that slips is a quarter of contracted revenue that does not arrive while the interest on the money that built it still does.

None of that is a forecast about the listing, the company or any share price. It is the list of things that would convert a private number into a checkable one — and until they do, the honest description of $103 billion is that it is a plan with signatures on it.

Educational macro context only — not investment advice.