Nscale’s $103B IPO: 2-Customer Concentration Risk
Nscale lost $1.02 billion on $140.6 million in six-month revenue. Now it wants public investors to believe $103 billion in contracts makes that sensible.
Nscale lost $1.02 billion on $140.6 million in revenue in the first half of 2026. Now it is heading toward an IPO waving around more than $103 billion in contracted value.
That is either the shape of a generational AI infrastructure company being built at speed, or a very expensive lesson in what happens when two customers hold most of your future in their hands. Probably a bit of both.
The $103 billion number is real. So is the catch.
Nscale filed for an IPO on Friday, September 18, after being spun out of Australian crypto-mining business Arkon Energy only two years ago. The London-based company sells AI computing capacity: the data centres, power, GPUs and cloud plumbing that AI labs need when running a few servers simply will not cut it anymore.
The headline number in its filing is more than $103 billion in total contracted value. That gets attention for obvious reasons. It is a big number even in a market that has become numb to ridiculous numbers.
But here is the bit founders and investors should not skim past: roughly 85% of that contract value is tied to just two customers.
Microsoft accounts for a $43.8 billion agreement running through 2033. Anthropic accounts for another $44.6 billion agreement. The Anthropic deal is conditional on Nscale securing financing and meeting milestones its own filing describes as stringent. Anthropic also retains rights to walk away or cancel in certain circumstances. ([techcrunch.com](https://techcrunch.com/2026/09/22/nscales-ipo-will-test-wall-streets-appetite-for-concentrated-ai-bets-once-again/))
That is not diversified revenue. It is concentrated exposure wearing a very expensive suit.
To be clear, big contracts are better than no contracts. I would rather have Microsoft and Anthropic on the customer list than a thousand tyre-kickers saying they are “exploring AI”. But a signed contract is not the same as cash in the bank, and future contracted value is not the same as a resilient business.
Public-market investors are about to decide whether they care about that distinction.
Nscale is selling the full-stack AI dream
Nscale is not trying to be merely another GPU landlord. Its pitch is vertical integration: power, data centres, compute and the software layer to run AI workloads.
That explains its July agreement to acquire Anyscale, the company commercialising Ray, the open-source framework created to distribute AI and Python workloads across large numbers of machines. Nscale said Anyscale’s approximately 200-person team would join it, while Anyscale would continue serving customers under its own brand. The deal was expected to close in the second half of 2026; Nscale did not disclose financial terms. ([nscale.com](https://www.nscale.com/press-releases/nscale-acquires-anyscale))
This is smart strategy, not corporate theatre.
Compute is a brutal business when all you own is compute. You borrow money, buy chips that depreciate faster than most founders admit, pay for power, build facilities, and pray demand stays ahead of supply. The more interchangeable your GPU capacity is, the less pricing power you have when the market turns.
Software changes the conversation. If Nscale can own the layer where customers develop, train, fine-tune, deploy and manage AI workloads, it becomes harder to replace. It has a better chance of earning more from each customer, smoothing out lumpy infrastructure economics, and keeping workloads inside its ecosystem.
That is why the Anyscale move matters more than a shiny acquisition announcement. Nscale is trying to escape the commodity trap before Wall Street decides AI cloud is just a leveraged version of renting out forklifts.
Growth is impressive. The financing appetite is the real test.
Nscale’s first-half figures are not subtle. Revenue rose to $140.6 million for the six months ended June 30, 2026, from $10.4 million in the comparable period a year earlier. Its net loss rose to $1.02 billion from $369 million.
Yes, revenue growing that fast is impressive. No, it does not make the loss irrelevant.
A business building AI infrastructure has to spend heavily before much of the revenue arrives. Land, power access, data-centre build-outs, GPU fleets, networking and financing costs do not wait politely for the customer to start consuming capacity. That is the nature of the beast.
But this model only works while capital remains available and confidence remains high. The moment either gets shaky, the company with the biggest future promises and the largest funding requirement is often the one that gets punched first.
Nscale has raised billions in equity and debt financing. Nvidia agreed earlier in September to provide $1 billion in convertible debt as part of a larger $3.1 billion financing package, according to TechCrunch. Reports on the IPO have varied: Axios cited an estimate that it could raise around $2 billion, while other reporting said Nscale was seeking $3 billion and targeting a valuation around $35 billion. The final terms will matter far more than the pre-IPO gossip. ([axios.com](https://www.axios.com/2026/09/21/nscale-ipo-nvidia-artificial-intelligence))
That is the key point: Nscale is not simply listing a business. It is listing a funding machine.
The overlooked risk is not AI demand. It is bargaining power.
Most people looking at this story will ask whether AI demand is real. That is the easy question. Demand plainly is real.
The harder question is: who gets to keep the economics?
Microsoft, Anthropic, Nvidia, cloud providers, data-centre operators, chip makers, power suppliers and financiers are all tied together in a merry-go-round of capital commitments, capacity agreements, investments and supply deals. Axios called Nscale’s filing a map of AI’s “circular economy,” with Nvidia a prominent player. ([axios.com](https://www.axios.com/2026/09/21/nscale-ipo-nvidia-artificial-intelligence))
This can work brilliantly while everybody needs more capacity yesterday. It gets less brilliant when one major buyer slows spending, delays deployment, changes its preferred chip architecture or decides it can get a better deal elsewhere.
Nscale’s own customer concentration makes that painfully clear. Microsoft and Anthropic are not small customers who will tolerate mediocre service because switching is annoying. They are sophisticated, powerful buyers with alternatives, internal expertise and the ability to negotiate hard.
Founders love saying they have landed a whale. Fine. Just do not confuse landing a whale with owning the ocean.
The company is competing with CoreWeave, Nebius, Lambda and Crusoe, among others. In a crowded race to finance and build AI capacity, scale is useful — but it can also become an addiction. Every new contract can create a bigger need for capital, more operational complexity and more exposure to execution risk.
The contrarian view: concentration is not automatically a deal-breaker
Here is the part the doom merchants miss: concentration is sometimes rational at the start of a massive infrastructure cycle.
When you are building power-hungry, capital-intensive assets, a few enormous anchor customers can make the project financeable. Their commitments may unlock debt, attract chip supply, justify construction and create the utilisation required to get an operation moving.
That is not a bug. It is often how infrastructure gets built.
The problem comes when management starts treating the anchor-customer phase as the finished business model. The right next move is not merely adding more megawatts. It is adding customer diversity, shortening the time between capex and cash generation, building software stickiness and ensuring a customer’s departure would hurt but not threaten the company’s existence.
Nscale’s Anyscale acquisition suggests it understands this. Owning a valuable developer and orchestration layer is a credible attempt to create more durable economics than raw GPU rental alone.
But strategy slides do not diversify customers. Revenue does.
What this means for you
If you are a founder, take the obvious lesson: a huge customer can validate your business while quietly becoming your business. Do not wait until the board asks uncomfortable questions to calculate what happens if your biggest account pauses, renegotiates or leaves.
Use this tomorrow:
1. Measure concentration honestly. Track the percentage of revenue, pipeline and contracted value tied to your top one, three and five customers. Do not bury it in a dashboard no one reads.
2. Separate contract value from cash. A five-year agreement is valuable, but it is not five years of collected money. Model cancellation rights, milestones, financing conditions, ramp schedules and delivery costs.
3. Build something customers cannot easily unbundle. Nscale is pursuing software alongside infrastructure for a reason. In your own business, ask where you are merely supplying capacity and where you are becoming part of the customer’s workflow.
4. Raise capital before you need to explain why you need it. Infrastructure businesses get punished when funding becomes urgent. The same is true for smaller companies with long implementation cycles or chunky working-capital needs.
5. Treat impressive growth as a question, not an answer. Revenue growing from $10.4 million to $140.6 million is extraordinary. A $1.02 billion loss means you still need to ask what that growth costs, how repeatable it is and who carries the downside if the assumptions shift.
Nscale may become one of the defining AI infrastructure companies of this cycle. It has serious customers, serious contracts and a sensible ambition to own more of the stack.
But its IPO is also a useful reminder that the hottest businesses can carry the oldest risk in capitalism: dependence on a few powerful buyers, financed by a market that assumes tomorrow will look even better than today.
That is not a reason to avoid the opportunity. It is a reason to price the risk properly.