Nscale’s $35B IPO: $1.02B Lost, $103.4B TCV
Nscale lost $1.02 billion on $140.6 million of half-year revenue. At a reported $35 billion IPO valuation, investors are being asked to bet its $103.4 billion TCV can be delivered.
Nscale lost $1.02 billion in the first half of 2026 on $140.6 million in revenue. At a reported $35 billion IPO valuation, investors are being asked to bet that its $103.4 billion in total contract value can actually be delivered.
That is either the sort of audacious move that creates a category leader, or an extremely expensive bet on execution.
The number Wall Street will want to believe
London-based Nscale filed publicly for a New York IPO on September 18, seeking to list on the NYSE under the ticker NSCL. Its pitch is the AI infrastructure dream in its purest form: own the power, own the data centres, own the GPU capacity, own the cloud software, and get paid every time somebody wants to train or run an AI model.
The headline number is $103.4 billion in active and contracted total contract value as of August 31. Nscale says that figure supports roughly 461,000 active or contracted GPUs. Its revenue rose from $10.4 million in the first half of 2025 to $140.6 million in the first half of 2026 — growth of 1,252%.
That is a proper ramp. No argument.
But this is where people need to put the champagne back in the fridge. Contracted total contract value is not the same thing as cash in the bank, revenue this year, or even revenue next year. It is the projected value of contracts over time, provided Nscale can build the facilities, energise them, install the kit, finance the kit, keep the customers, and deliver the computing capacity it promised.
That is not a minor list of chores. That is the business.
Nscale’s own filing shows the current gap clearly: $140.6 million of six-month revenue and a $1.02 billion net loss. Bloomberg reported the same figures from the filing. The company is not selling a mature, cash-spitting cloud utility to public investors. It is selling a claim on what a fully built AI utility could become. ([bloomberg.com](https://bloomberg.com/news/articles/2026-09-18/nvidia-backed-data-center-firm-nscale-files-publicly-for-us-ipo?utm_source=openai))
Nscale is not really selling GPUs — it is selling certainty
Every AI company says compute is scarce. Fair enough. The more interesting question is: scarce for whom, at what price, and for how long?
Nscale’s answer is vertical integration. Rather than merely leasing a warehouse, filling it with Nvidia hardware and hoping customers turn up, it is trying to control the whole chain: power, land, data centres, GPU financing, cloud infrastructure and the software layer that makes massive AI workloads usable.
That last bit matters more than most investors will appreciate.
In July, Nscale agreed to buy Anyscale for a reported $1.65 billion. Anyscale was founded by the people behind Ray, an open-source framework used to distribute AI and Python workloads across large computing clusters. Nscale gets more than another logo for the investor deck. It gets a software layer that can help customers process data, train models, run inference and manage increasingly messy AI workloads across its infrastructure.
That is the sensible bit of the strategy. The bloke who owns the industrial estate makes money. The bloke who owns the software that determines whether the factory actually runs properly often makes even more.
Nscale says Anyscale’s roughly 200 staff will join the company, while Anyscale continues to operate under its own brand. The official terms were not disclosed; the $1.65 billion price was reported separately. ([nscale.com](https://www.nscale.com/press-releases/nscale-acquires-anyscale?utm_source=openai))
This acquisition tells you Nscale knows the ugly truth of AI infrastructure: hardware is a commodity faster than people think. A GPU is scarce until the next generation arrives, a competitor signs a supply agreement, or a customer decides its model no longer needs quite so much compute. Software, workflow lock-in and operational reliability are the bits that make a customer harder to dislodge.
The overlooked risk is not demand. It is concentration.
The sales story is enormous, but it is not broadly diversified in the way the words “$103 billion backlog” make people imagine.
Fortune reported that ByteDance accounted for 73% of Nscale’s 2025 revenue. Separate reporting on the IPO filing said Microsoft and Anthropic represented 85% of Nscale’s contract backlog. That does not make Nscale a bad business. It makes it a business with a handful of very large, very sophisticated counterparties holding an extraordinary amount of leverage.
If you are supplying a local cafe, you can put the prices up, change the menu and have a crack at better marketing. If a tiny collection of global technology giants represents the lion’s share of your revenue and contracted work, you live inside their capital-expenditure decisions.
One delayed data-centre site. One shift in chip availability. One customer revising its model roadmap. One decision to build more capacity internally. Suddenly, a spectacular contract-value number starts behaving like a theoretical number.
That is why the “AI boom” label is too lazy. There are two very different businesses hiding under it.
The first is the business that sells picks and shovels into genuine demand and earns a sensible return on capital.
The second is the business that spends an absurd amount today because it expects demand, capital markets and customer budgets to remain generous at the same time for years.
Nscale may become the first. Its filing does not yet prove it has escaped the risks of the second.
The contrarian view: the loss is not the biggest red flag
Plenty of commentators will point at the $1.02 billion loss and say, “See? Bubble.” That is too easy, and probably wrong.
Building data centres, securing power, buying or leasing equipment and standing up AI capacity costs a bomb before revenue catches up. Early losses are not automatically evidence of stupidity. They can be the price of building a hard asset base before everyone else wakes up.
The sharper question is whether the company is building infrastructure at returns that will still look attractive after competition does what competition always does: arrives, borrows money and cuts price.
Nscale has real momentum. It raised more than $3.3 billion through financing rounds, has significant external financing commitments, and has secured blue-chip relationships. Its revenue growth is real. Its backlog is real in the contractual sense.
But a backlog is only worth what it costs you to fulfil it.
That is the entire game here. Not the ticker. Not the hype. Not the number of GPUs in a press release.
For founders, this is a useful lesson. Investors love growth until they discover it was purchased with economics that cannot survive a normal market. You do not get points for winning revenue at any cost. You get paid for building a machine that turns capital into more capital.
Nscale is taking the opposite bet: spend heavily now, achieve scale fast, and become essential before the market turns commodity. It may be right. But it is a very binary strategy dressed up in cloud-computing language.
Why the Anyscale deal is the smartest part of the whole story
If I were running Nscale, I would be less interested in getting applause for the $103.4 billion contract figure than in making sure the Anyscale acquisition works.
A data centre customer can move servers. It is painful, but they can do it.
A customer whose engineering teams rely on your workflow tooling, orchestration, observability and production systems is much harder to rip out. That is where margin and durability come from.
The Anyscale purchase also gives Nscale a chance to sell beyond raw compute. A founder may begin with a small AI workload, use Anyscale’s software to scale it, then consume Nscale cloud capacity as the project grows. That is a more defensible commercial loop than simply racing rivals to rent out the next GPU.
The risk, naturally, is execution. Software companies and infrastructure companies think differently. One prizes developer adoption and product velocity. The other lives in long project timelines, financing agreements, construction schedules and utility constraints. Combining them can create a killer stack. It can also create an organisation that is too slow for software and too distracted for infrastructure.
No corporate waffle here: buying the right company does not create a moat. Integrating it properly does.
What this means for you
If you are an investor, do not value Nscale by dividing $103.4 billion by some arbitrary multiple and calling it a day. Start with four boring questions:
1. How much of contracted value converts into recognised revenue each year? Total contract value is a long-term promise, not this year’s sales. 2. What must Nscale spend to deliver it? Revenue without a clear view of construction, power, hardware and financing costs is just a seductive number. 3. How concentrated are customers and suppliers? A company can have enormous contracts and still be fragile if two or three partners hold all the cards. 4. Does software increase the return on infrastructure? This is why Anyscale matters. If it improves utilisation, pricing power and customer stickiness, the deal could be worth far more than its purchase price.
If you are a founder or operator, pinch the underlying lesson. Build the bit of your business that makes customers harder to replace. Distribution matters. Assets matter. But the real money sits where a customer’s workflow, data and habit become entangled with your product.
And if you are tempted to worship a giant backlog, don’t. I have seen enough businesses confuse signed paper with money. A contract is not cash. A forecast is not a result. And a $35 billion IPO valuation is not proof that the economics work.
It is merely Wall Street being asked to fund the proof.