Nscale is raising $3.5B before an IPO, with about $2B from Nvidia

This is vendor financing, which is old and lawful and common in capital-intensive industries. It is also, at this scale and repeated across this many counterparties, a genuine analytical problem.

Within a single fortnight Nvidia agreed to buy Hugging Face for $12.9 billion, discussed putting $2.5 billion into Thinking Machines Lab, backed Lambda as it signed $35 billion with Anthropic, took the lease on a Texas data centre from Hut 8, and committed $2 billion here. Jensen Huang then forecast roughly 70 percent revenue growth for fiscal 2028, well above analyst estimates.

Some portion of the commitments underpinning that forecast come from customers Nvidia has capitalised. Nobody outside the company can say what portion, because the disclosures do not separate them.

What an IPO would expose

A public listing is where this becomes concrete, because Nscale would have to publish what its contracts actually contain.

A $45 billion contract with Anthropic is a headline number until you know its duration, whether it carries minimum commitments, how it is priced against market rates, and what happens if Anthropic's growth disappoints. Anthropic's own compute obligations now total roughly $90 billion across Nscale, Lambda and Volta, against a run rate reported above $65 billion — and Anthropic's S-1 is expected within weeks.

Two companies with overlapping disclosure timelines, each of whose forward revenue depends on the other's forward spending, will be describing the same contracts to public investors at roughly the same time. That is the first real test of whether these numbers survive contact with securities disclosure.

The Figure deal is the more interesting one

The $3.5 billion agreement with Figure AI targets 100,000 GPUs for humanoid robot training, deploying in the second half of 2027.

Robotics has been the sector where compute demand was theoretical — real-world data collection, not compute, has been the constraint. A contract of this size says Figure believes that has changed, and that it can generate enough training data to keep 100,000 accelerators usefully occupied.

It is also a two-year forward commitment on hardware that does not exist yet, from a company whose product is not shipping at volume, sold by a provider raising money to build the capacity. Every leg of it is a forecast.

Nscale has not disclosed revenue, utilisation, or the terms of any of its contracts, and has not confirmed the raise.