Nscale’s $3.5B Pre-IPO Raise Is a $103B AI Bet
A two-year-old AI cloud company is seeking $3.5 billion before an IPO while waving around $103 billion in contracts. That is either world-class execution or a very expensive lesson in the difference between demand and cash.
A two-year-old AI cloud company is seeking $3.5 billion before an IPO while waving around $103 billion in contracts. That is either world-class execution or a very expensive lesson in the difference between demand and cash.
Nscale is the story worth watching because it exposes where the AI boom has moved: away from clever chatbot demos and into the brutally capital-intensive business of power, data centres, chips, debt and contracts that may run for years before they produce anything resembling free cash flow.
The London-based company is reportedly in talks to raise up to $3.5 billion ahead of a potential US listing as soon as September 2026. The proposed package includes up to $1.5 billion in convertible notes and about $2 billion in financing from Nvidia. That is not a normal startup raise. It is a giant wager that AI demand will remain strong enough, long enough, to justify building the digital equivalent of ports, railways and power stations at warp speed.
The core story: $3.5 billion before the float
Nscale is trying to raise fresh capital at precisely the point most founders would be polishing their IPO deck and pretending the hard work is done. Good on them for not pretending.
According to reporting on the talks, Nscale wants up to $3.5 billion in pre-IPO financing. Nvidia is central to the proposed package, which matters enormously. Nvidia is not merely selling the picks and shovels in this gold rush; it is increasingly helping finance the companies buying huge quantities of those picks and shovels.
Nscale has already shown it can raise money at industrial scale. In March 2026, it raised $2 billion at a $14.6 billion valuation, with backing that included Nvidia, Dell Technologies and Nokia. Before that, it raised a $1.1 billion Series B in September 2025.
Then came the operating ambition. Nscale agreed in July to acquire Anyscale, reportedly for $1.65 billion. Anyscale is the company behind the commercial platform built around Ray, an open-source distributed-computing framework used to run AI workloads across large fleets of machines. The deal is important because it shifts Nscale from being principally a seller of capacity toward becoming a seller of a more complete AI operating environment.
That may sound like consultant jargon. It is not. If you own the data centre, secure the GPUs, schedule the workloads and provide the software layer customers use to train and run models, you get more of the customer’s spend and make yourself harder to replace.
Nscale is telling prospective investors it has roughly $103 billion in total contracted revenue. The headline figure jumped after a reported $45 billion computing agreement with Anthropic, alongside other customer commitments. But before people start spraying champagne around, read the fine print hiding in plain sight: contracted revenue is not revenue in the bank. It is an estimate of future income based on signed lease agreements, and the company’s materials reportedly frame it as illustrative rather than formal revenue guidance.
That distinction is not boring. It is the entire ball game.
AI infrastructure is now an execution business, not a story business
A model company can hire brilliant researchers, burn cash and release a shiny benchmark chart. An infrastructure company has to obtain land, grid connections, transformers, cooling, permits, chips, network equipment, customers and financing—then make every piece show up on time.
That is why Nscale’s move is more interesting than another model launch. It is trying to build a full-stack AI cloud at a pace normally reserved for wartime logistics or a resources boom.
The company has announced major deployments and commitments across the United States and Europe. Its existing Microsoft agreement covers approximately 200,000 Nvidia GB300 GPUs across four countries. A Texas site alone was slated to receive around 104,000 GPUs, while deployments were also planned in Portugal, the UK and Norway. More recently, Nscale announced a multi-year arrangement with humanoid robotics company Figure, with an initial $3.5 billion compute commitment and the potential to scale beyond $6 billion.
These are eye-watering numbers because the underlying economics are eye-watering. Nscale has noted that a 20-megawatt cluster with 10,000 GPUs can cost just under $2 billion to stand up. That tells you why companies such as Nscale cannot fund their growth with a cheerful Series A and a few LinkedIn posts.
The physical constraints are also real. A GPU is useless if it is sitting in a warehouse because the power connection is delayed. A finished data centre does not print money if a customer delays workloads. And a signed multi-year agreement is less comforting if the customer’s own business model changes or its funding dries up.
This is not software margins with a few servers tucked away in Oregon. It is heavy infrastructure wearing a software hoodie.
The $103 billion figure is impressive—and exactly why investors should be careful
I like big ambition. I have no issue with founders raising aggressively when they can see the market moving beneath their feet. In fact, waiting politely for certainty is how you lose a generational opportunity.
But operators and investors need to separate three things that are getting mashed into one giant AI number: bookings, revenue and cash.
Bookings tell you what customers have agreed to buy. Revenue tells you what has actually been delivered and recognised. Cash tells you whether you can keep paying suppliers, debt holders, staff and power bills while all the grand promises turn into infrastructure.
Nscale’s reported $103 billion of contracted revenue averages about $18 billion annually over contracts averaging 5.7 years, based on documents reviewed by The Information. Yet the same reporting said Nscale estimated more than $100 million in second-quarter 2026 revenue, following roughly $37 million in the first quarter. That growth is rapid. It also demonstrates the canyon between a future contract headline and present-day revenue.
That canyon is not necessarily a red flag. It is normal in infrastructure. But it is where businesses live or die.
The risk is timing mismatch. Nscale must spend now on chips, facilities and deployment, while much of the economic benefit arrives later. If demand stays ravenous, that operating leverage will look brilliant. If AI customers trim spend, delay buildouts or discover they overcommitted to capacity, the financing structure becomes very unforgiving very quickly.
Convertible notes are especially worth watching. They give lenders a path into equity, usually with downside protection and potential upside if the IPO performs. That can be a smart bridge to public markets. It can also be expensive capital disguised as confidence if the listing does not land as hoped.
The overlooked angle: Nvidia is building customers, not just selling chips
The easy take is that Nvidia is the obvious winner because everyone needs its hardware. True, but incomplete.
Nvidia’s involvement in Nscale is strategically powerful because it helps create more buyers capable of deploying Nvidia systems at astonishing scale. It is selling into the boom while helping finance the infrastructure required for the boom to continue.
That does not make it sinister. It makes it smart. But it does mean investors should understand the ecosystem rather than admire individual headlines. The chipmaker, the AI cloud provider, the model lab and the end customer can all be commercially dependent on one another. A large commitment at one layer can support a valuation at another, which enables financing for a buildout that buys more equipment from the original supplier.
That loop works beautifully while utilisation rises. It looks much less beautiful when utilisation falls.
Nscale’s acquisition of Anyscale is the best defence against becoming a commodity landlord for GPUs. Hardware capacity eventually gets compared on price, availability, latency and service. Software can create stickiness. Anyscale also says its platform will continue to run across major cloud providers after the deal closes, preserving multi-cloud flexibility for customers.
That is a sensible strategic move. Nscale is trying to own more than metal and megawatts. It wants to own the workflow.
What this means for you
If you are a founder, stop telling yourself that “AI strategy” means adding a chatbot to your product. The real question is whether AI changes your cost structure, delivery speed, pricing power or customer lock-in. If the answer is no, you have a feature, not a strategy.
If you are buying AI infrastructure or committing to long-term model spend, negotiate for flexibility. Demand clear capacity-delivery milestones, performance standards, exit clauses and transparency around what is reserved versus what is actually available. A big contract is not a win if your business is stuck paying for compute you cannot use.
If you are an investor, do not value infrastructure companies on contracted-revenue headlines alone. Ask four blunt questions: What has been delivered? What is being paid in cash today? How much capital must be spent before the next dollar of revenue arrives? And what happens if utilisation is 20% lower than the model assumes?
And if you are an operator, learn the Nscale lesson without needing a $3.5 billion cheque: own the bottleneck closest to your customer. Nscale is buying software because raw compute alone will eventually be compared like electricity. The winners do not just supply the input. They make the customer faster, more productive and harder to pry away.
Nscale may become one of the defining infrastructure businesses of the AI era. It may also become the case study everyone cites when they remember that contracts are promises, construction is hard, and capital is never as patient as founders hope.
Either way, the company is showing us where the AI race is now being fought: not on a stage with a demo, but in the unglamorous, expensive machinery underneath it.