Third Point led a $3.36 billion convertible note into Nscale on 25 September. Nvidia is contributing $1 billion of it. The world's most valuable chipmaker is now a creditor of a two-year-old London company whose entire product is rented Nvidia chips.
It filed to list on the New York Stock Exchange a week earlier, under the ticker NSCL. The registration statement claims $103.4 billion of total contracted value (TCV). Two customers account for nearly all of it.
The case for owning the whole stack
The model does not resell someone else's cloud. It buys the land, signs for the power, builds the liquid-cooled halls, installs the racks, and writes the orchestration layer on top. A customer signs one contract and receives a working AI factory.
That reach is the pitch. Lock in power before the grid runs out of interconnection capacity. Lock in processors before allocation tightens further. Capture a margin at every layer, from the substation to the scheduler that decides which model runs first.
The supply chain backs the story. Dell, Nokia, and Nvidia have all invested. Nvidia's chief executive, Jensen Huang, has called it a national champion for AI infrastructure. In a market where allocation decides who ships, that endorsement carries more weight than most term sheets. It is also one of the few builders putting steel in the ground at gigawatt scale in Europe, a position that carries a scarcity value pure resellers lack.
Total contracted value claimed in the S-1
Up from about $51B reported on 6 August, mostly on a single Anthropic agreement. · S-1, Bloomberg 2026
The contract book supports it. Microsoft signed for roughly 200,000 Nvidia GB300 processors across sites in Texas and Portugal. Anthropic committed in late August to an agreement worth about $45 billion. Together they lifted total contracted value from about $51 billion in early August to $103.4 billion by the end of the month.
The build is real, not paper. A 240-megawatt campus in Texas is leased from Ionic Digital and designed to scale to 1.2 gigawatts over time. A second site at Sines in Portugal anchors the European side of the same contract. Much of it runs through a joint venture with Aker ASA, the Norwegian industrial group, which brings hard assets and engineering discipline to a model that otherwise rests on contracts alone.
Europe needs the capacity. The continent has the model labs and the policy ambition but not the power or the halls. That scarcity is why the order book filled so quickly, and it is why the company can claim a national-champion role without much competition pushing back.
Nscale is a national champion for AI infrastructure.— Jensen Huang, CEO, Nvidia
The case against: one supplier, two tenants
A neocloud, a specialist builder that rents out AI compute, converts capital into machinery and rents it back at a spread. The spread holds only while the tenants stay and the hardware keeps its value. Both assumptions carry weight here, and neither sits fully in the company's hands.
Bloomberg reported on 21 September that Anthropic and Microsoft dominate the $103.4 billion of contracts. Tenants of that size do not accept list prices. They negotiate, and one renegotiation moves the revenue line further than any new customer can repair. A supplier with two buyers is a supplier with two points of leverage against it.
How much of the $103.4B is actually guaranteed?
The disclosure gap matters more than the headline. The registration statement records $2.6 billion of active contracted value as of 31 August. The rest is signed but not yet billing. A reader cannot trace the $103.4 billion to an unconditional minimum payment floor, contract by contract. For anyone underwriting the book, that is the first question the filing does not answer.
Then comes the structure of the new money. The notes convert at the IPO price minus a double-digit discount and stop accruing value above a $30 billion mark. Buyers are protected on the downside and capped on the upside. The structure is a lender's trade wearing an equity label, and it tells you how the lead investor is pricing the risk.
The balance sheet has already absorbed about $6.8 billion across equity and debt. The stack includes a $1.4 billion loan backed by processors, a $900 million revolving facility, $790 million arranged in Norway, and the $1.65 billion purchase of Anyscale in July. The last public accounts, covering seven months to December 2024, show a $24 million loss on a business that now employs more than 700 people.
None of this is unusual. CoreWeave, Nebius, and Crusoe run comparable balance sheets, and public markets have rewarded them. The open question is what happens when the debt matures before the hardware depreciates. Power contracts last twenty years. Processors do not.
The question is not whether AI demand exists. It is who owns the processors when the demand curve stops bending upward.
The convertible structure tells you which side the lead investor thinks it is on.
Why the capital keeps arriving
Nothing here is specific to Nscale. Inference is projected to consume the large majority of AI energy, and inference demand scales with usage rather than with training runs. Someone has to own the buildings. The private market has decided that someone should be a specialist rather than a hyperscaler, because a specialist moves faster on land, power, and cooling.
That logic pulled billions into a handful of builders in under two years. It also pulled capital into dozens of rivals. When every credible operator can raise, the advantage shifts from access to capital toward access to power. Management says the pipeline exceeds 10 gigawatts. A pipeline is not a permit, and a permit is not an energized line. City AM reported in July that grid delays had already pushed one of its sites to look for alternative generation.
The timing adds its own pressure. The company wants to price a listing into a market that has grown nervous about AI capital expenditure. In the same week it announced the convertible, a wave of commentary asked whether data center spending has outrun demand. A listing is a one-time window. If it closes, the debt stays.
What $3.36 billion actually buys
Strip the framing and the raise funds three things: land and power, processors, and the gap between them. Only the first is genuinely scarce, and it is the slowest to build.
| Metric | Bull reading | Bear reading |
|---|---|---|
| Contracted value | ✔ $103.4B total | ✗ $2.6B active |
| Customer base | ◐ Anthropic and Microsoft | ✗ Two tenants |
| Supplier | ✔ Partner and investor | ✗ Creditor and supplier |
| New capital | ✔ $3.36B raised | ✗ Converts at a discount |
S-1; Bloomberg; company filings, September 2026.
The valuation is the tell. At a reported $35 billion, the equity sits at roughly a third of contracted value, against a book that is mostly unbuilt and concentrated in two names. Crusoe's private round in September priced a similar energy-first model at $30 billion, with more operating history behind it. The listing asks the public market to pay the premium first.
The November tranche carries its own message. Nvidia's $1 billion lands only after the round closes. If the listing slips, the financing has a cliff built into it. The supplier becoming the lender of last resort is a strange position for the most valuable company in the world.
As we wrote in September, Crusoe's $3B round put a $30B price on energy-first AI. Nscale is now asking for a higher price on a younger book. The comparison is not fatal, but it is the most useful one available.
None of this settles the argument. Nscale may build the most important AI infrastructure in Europe, and the demand behind it is not imagined. The bear case is narrower than it first sounds. The company is priced as though nothing goes wrong between now and 2028, in a sector where every participant finances itself on the assumption that the curve keeps bending upward.