The $51 Billion Question: What Nscale's IPO Really Tests

Nscale says it has $51 billion in contracted revenue. Last year, it earned $33 million.

That gap looks like a mistake. It is not. It is the defining feature of the company's business model, and whether you understand that feature determines whether this looks like the future of AI infrastructure or a capital trap.

Nscale is preparing for a U.S. IPO that could raise as much as $3 billion. The company was founded in May 2024 by Josh Payne, a 32-year-old Australian who previously ran a Bitcoin mining operation called Arkon Energy. On Wednesday, it announced that Fidji Simo — former head of product and business at OpenAI, and the CEO who took Instacart public — has joined its board. She follows Sheryl Sandberg, Susan Decker, and Nick Clegg, all added earlier this year.

The board reads like a credibility campaign. That is probably not a coincidence. A two-year-old infrastructure company with a $14.6 billion private valuation needs to convince public-market investors that its business can survive outside the startup funding ecosystem. Simo's IPO experience is relevant. What's less clear is whether the business underneath the board deserves that kind of introduction.

So what does Nscale actually do?

It is a neocloud. That means it buys GPU servers, puts them in data centers, and rents out computing power by the hour. The economics are the spread between what a customer pays per GPU-hour and what it costs Nscale to own and operate that GPU — depreciation, data center rent or build cost, power, cooling, and interest on the debt used to buy the hardware in the first place.

The business is fundamentally capital intensive. Nscale has taken on roughly $3 billion in identified financing — including a $1.4 billion loan secured against GPUs, a $790 million facility in Norway, and a $900 million credit line signed in July. The assets backing that debt are chips that lose value the moment newer, faster ones arrive. Nvidia releases new GPU generations roughly every year. If a facility is built around today's Vera Rubin chips, the next generation makes them cheaper overnight — for everyone except the person who already bought them.

That depreciation problem is the hidden center of the neocloud model. It works only if you can keep utilization high enough, for long enough, to earn back the hardware cost before it becomes obsolete. Long-term contracts help. They also create risk if the customer's plans change.

And this is where the $51 billion comes from. It is not revenue. It is the sum of multi-year commitments that will be recognized as revenue over the life of the contracts — some of them stretching six years into the future. The Anthropic deal alone is roughly $45 billion across six years, with the West Virginia facility not expected to come online until the end of 2027. The Microsoft deal is reportedly $14 billion for 200,000 GPUs across Texas, Portugal, the UK, and Norway.

Annualized, Nscale's current revenue run rate is estimated at $400 million to $500 million. The company went from $33 million in all of 2025 to just over $100 million in the second quarter of 2026. That is growth, but it is growth from a tiny base. The $14.6 billion private valuation, set in March, is roughly 30 times that run rate. The IPO would raise $3 billion on top of that.

Most people think about neoclouds through the demand story. There are not enough GPUs. AI companies need compute now. Big Tech is spending $650 billion on AI infrastructure this year. Nscale exists because the hyperscalers cannot build fast enough.

That is part of the picture. The rest is that Big Tech is spending $650 billion building its own infrastructure. Microsoft, Google, and Amazon are not Nscale's only customers — they are also Nscale's competitors. If Microsoft decides it can build capacity faster in-house, or if it renegotiates terms when its own data centers come online, Nscale loses a contract that was a centerpiece of its $51 billion headline.

The risk is not theoretical. OpenAI paused the Stargate UK project in April 2026, citing high UK energy costs and regulatory uncertainty. Nscale's Essex site, announced in September 2025, had not begun construction by March 2026 due to grid connection issues.

There is another layer. The neocloud model has been compared to a real estate play. Buy the building, collect the rent, appreciate the asset. But GPUs are not real estate. They are manufactured products with a planned obsolescence built into the supply chain. The "real estate" appreciation only works if the underlying hardware remains useful. It is not clear that it will, not five years from now.

None of this means Nscale is a bad business. The demand for AI compute is real. The company has secured major contracts with credible customers. The founding team has operational experience in deploying hardware at scale, adapted from Bitcoin mining where the economics are similarly about cheap power and dense compute. Modular, prefabricated data centers designed for liquid cooling and rapid deployment are a genuine improvement over the older model of general-purpose data centers retrofitted for GPU workloads.

But the question for a public-market investor is not whether the demand exists. It is whether the margins survive. The spread between what Nscale charges and what it costs needs to be wide enough and stable enough to service $3 billion in debt, replace aging hardware, and still leave room for the equity return that a $14.6 billion-plus valuation implies.

Here is what you can test. When Nscale goes public, look at three things. First, what percentage of the $51 billion in contracted revenue is from the two largest customers. If most of it is Anthropic and Microsoft, the concentration risk is the story. Second, look at the debt terms. Are the loans interest-only during the buildout, or does interest start accruing before the facilities produce revenue? Third, compare Nscale's projected gross margins to what it costs to buy the same GPUs on the open market and run them yourself. If the margin is thin, the company is taking all the execution risk for a return that doesn't justify it.

The board of ex-Meta executives and former OpenAI leadership is impressive. But boards don't keep GPU servers running. They don't negotiate power contracts or manage the depreciation curve. What does is the operating team, the capital structure, and whether the spread actually works.

The way to think about Nscale is not as an AI company. It is a hardware leasing business in the most capital-intensive version of that model, selling to customers who are simultaneously building their own version of the same thing. The $51 billion in contracts is real. The question is whether those contracts earn more than the cost of the debt, the depreciation, and the execution risk combined. If they do, the business works. If they don't, the contracts are just a promise that becomes a liability.