Nvidia is paying its own customer — what Nscale's $103 billion "backlog" actually is

A two-year-old London company that booked about $100 million of revenue in its latest quarter is in talks to raise $3.5 billion ahead of a New York IPO, at a valuation investors are treating as heading toward $30 billion. The number doing the work in that pitch is not the revenue. It is the $103 billion of "contracted" business Nscale claims to have lined up — roughly double what it said a month earlier. That gap between the reported present and the promised future is the whole story of the AI cloud company, and the largest name on both ends of it is Nvidia.

This is not the first time that pattern has shown up. It is just the clearest version of it yet.

Where the money comes from

The structure, reported by Bloomberg and confirmed by several outlets, is simple on its face. Nscale wants up to $3.5 billion: about $2 billion from Nvidia itself, and up to $1.5 billion in convertible notes led by Daniel Loeb's hedge fund Third Point, with Goldman Sachs advising. Nvidia will not reap that return from holding equity in a pre-IPO cloud company. Nvidia will reap it because Nscale has contracted for roughly 194,000 of Nvidia's next-generation Vera Rubin GPUs.

This is the circular-financing loop that now funds most of the AI compute trade. Nvidia takes a relatively small equity stake in a neocloud — it put $2 billion into both CoreWeave and Nebius — and that capital flows straight back into Nvidia's own revenue as the customer spends it on chips. A few billion of equity buys Nvidia tens of billions of locked-in GPU orders and a customer that, unlike a hyperscaler, has no accelerator strategy of its own. The young cloud company gets the money it needs to buy hardware; Nvidia gets the orders. The loop does not require the cloud company to be profitable, or even to have much revenue yet.

What the $103 billion actually is

The pitch that makes the raise possible is the backlog, and its size comes from a single deal. In late August, Anthropic agreed to pay Nscale $45 billion over six years to rent about 460 megawatts of compute capacity at Nscale's West Virginia campus — a facility built to run on Vera Rubin chips that begin coming online only late next year. That one contract explains most of the jump in Nscale's stated backlog from $51 billion to $103 billion in the span of a month.

Here is the part worth slowing down on. That $103 billion is not revenue Nscale has earned, or even will earn this year. It is the face value of long-term contracts for capacity that is mostly not built and does not come online until late 2027. Anthropic, notably, was not the first tenant at that campus. Microsoft signed a letter of intent in March to take the space, then walked away over the summer; Google passed after reviewing its own data-center spending. The deal Nscale now anchors its flotation on is one the two buyers most able to build it gave up on.

Against that backdrop, the contrast between the headline and the actuals matters. Nscale booked roughly $33 million in revenue across all of 2025, then passed $100 million in a single recent quarter. On top of the backlog, it is telling investors it could one day produce annual revenue of about $18 billion and adjusted earnings of roughly $13.6 billion — a projection the reports describe as illustrative rather than formal guidance.

Who the structure protects

The financing terms are where the incentives become visible. Third Point's convertibles are priced at a double-digit discount to the IPO price, with that discount adjusting upward until valuation stops at a $30 billion cap. That cap is meaningful on its own: Nscale was valued at $14.6 billion in a Series C in March. A discount-convertible investor converts at a price below whatever the market pays at the IPO, with protection against how richly the market prices a company whose signature asset starts delivering months after it lists.

So take stock of the three parties. Nvidia gets paid for 194,000 chips on day one of the deal, whether or not Nscale ever converts its backlog into profit — it is the one party that wins regardless of how the delivery story ends. Third Point and the other note holders buy in at a discount to whatever the market decides, with a cap that anchors the valuation. The party left holding the delivery risk is the one that buys at the IPO price with neither the discount, the chip revenue, nor the history — a retail buyer of a two-year-old, unprofitable operator whose $103 billion of contracted demand must be turned into built, powered, and utilized capacity before it is worth anything on an income statement.

This is the lens for every neocloud that lists. Backlog is collateral and future contracted value, not earnings; the spread between what a GPU is charged per hour and what it costs to own, power, and depreciate is the real margin; and the money in the loop flows back to Nvidia first. The question is never whether demand is real — it is whether this operator converts a $103 billion promise into recognized revenue on schedule, and whether you want to hold the side of the trade that carries the schedule risk. That is what separates the investor who understands the loop from the one who just reads the backlog.