Nscale’s $30B IPO Has a Microsoft Power Problem
A $30 billion IPO can be wrecked by one missing utility connection. Microsoft walking from Nscale’s UK site is a reminder that AI hype still runs on boring old electricity.
Microsoft did not just walk away from an Nscale data-centre deal. It reminded the entire AI industry that a GPU is an expensive paperweight without power.
That is the bit everyone conveniently forgets while throwing around trillion-dollar AI narratives. You can announce contracts, buy Nvidia chips, hire former Meta executives and talk about becoming the next hyperscaler. But if the grid cannot supply the electricity, you have not built infrastructure. You have built a very costly PowerPoint.
On October 18, 2026, Microsoft pulled out of an agreement to rent AI infrastructure at Nscale’s planned Loughton, Essex site. The project had been delayed, and the central problem was brutally physical: adequate grid power was not expected until the 2030s.
The timing could hardly be worse. Nscale is pursuing a US IPO at about a US$30 billion valuation. Investors are reportedly pushing for something closer to US$16 billion to US$19 billion. That is not a minor disagreement over pricing. It is the market asking a much more uncomfortable question: are these AI-infrastructure businesses selling reliable capacity, or selling a future that still needs electricity, debt and miracles to arrive on schedule?
Nscale’s problem is not that Microsoft left
Let’s be fair about what happened.
Nscale says its broader relationship with Microsoft remains intact, and the Loughton issue does not affect other Microsoft projects. This is not Microsoft cancelling every arrangement with the company. Nscale still has projects and relationships across Norway, Portugal, Texas and West Virginia.
But that defence misses the point.
For a company headed to public markets, the first question is not whether some contracts remain. It is whether the business can deliver the contracts investors are being asked to value today.
Loughton was meant to house 23,040 Nvidia GPUs by the first quarter of 2027, providing an initial 50 megawatts of AI capacity, with plans to expand to 90 megawatts. Instead, a basic input — power — turned into the bottleneck.
That is the real story. AI infrastructure is being marketed like software: scalable, inevitable and only limited by demand. It is not software. It is a capital-intensive, power-hungry, construction-heavy industrial business wearing an AI hoodie.
And industrial businesses get punished when they miss deadlines, blow budgets or depend on a single supplier, customer or regulator. They should be.
The $103 billion headline needs an asterisk the size of London
Nscale’s IPO filing showed more than US$103 billion in contracts. On the surface, that is eye-watering. It is exactly the kind of figure that gets people leaning forward at a dinner party and nodding as though they have found the next Nvidia.
Don’t do that.
About 85% of Nscale’s contract value comes from two customers: Microsoft and Anthropic. TechCrunch reported that the Microsoft agreement accounts for US$43.8 billion of compute through 2033, while Anthropic’s agreement is worth US$44.6 billion.
There is another catch: the Anthropic agreement depends on Nscale obtaining financing and meeting milestones the company itself describes as stringent. That does not make it fake. It makes it conditional — which is a very different thing from cash in the bank.
A contract backlog is valuable. A concentrated, conditional contract backlog tied to facilities that are still being built is valuable in a completely different way.
Founders regularly make this mistake because it is emotionally satisfying to count announced revenue as if it has already paid payroll. Investors make the same mistake because large numbers are soothing. They make uncertainty look like progress.
But a signed deal is not the same as an operating asset. A customer commitment is not the same as collected revenue. A GPU order is not the same as usable computing capacity. And an IPO valuation is definitely not the same as a business worth owning.
Nscale reported US$140.6 million in revenue for the six months to June 30, up sharply from US$10.4 million a year earlier. Good growth. But it also reported a net loss of US$1.02 billion, versus a US$369 million loss in the comparable period.
That tells you what this game costs. Growth is not arriving neatly. It is being bought with staggering up-front investment, while the company races to turn contracted demand into live, paid-for capacity.
The AI gold rush has found its railway problem
In every boom, the sexy story gets the headlines and the unsexy constraint gets the money.
During gold rushes, it was picks, shovels and transport. During the internet boom, it was fibre and data centres. In AI, everyone talks about models, agents and chips. The constraint is increasingly power: generation, grid connections, transformers, cooling, land approvals and the people capable of delivering all of it.
That does not mean AI is a bubble. It means the market has stopped pretending that enthusiasm overrides physics.
Nscale is a useful case study because it has tried to own more of the stack. In July, it agreed to acquire Anyscale, the company behind the Ray distributed-computing platform, adding software to its existing pitch around power, data centres and GPUs. The stated ambition is sensible: do more than rent hardware and offer customers a full AI platform.
That may ultimately be a better business than a plain GPU landlord. Software can raise margins, reduce customer churn and give the company a reason to matter once chip capacity becomes less scarce.
But vertical integration cuts both ways. Owning more layers can create an edge. It can also mean owning more things that can go wrong at once: real estate, energy procurement, construction, financing, chip supply, cloud operations and software integration.
That is not a criticism of ambition. I like ambition. It is a criticism of lazy valuation logic.
You do not get a premium multiple merely by listing every hard problem in the supply chain and calling it a moat.
The overlooked angle: Microsoft’s exit may be rational, not catastrophic
Here is the contrarian view: Microsoft walking away from Loughton is not necessarily a vote of no confidence in AI demand or Nscale as a whole.
It may simply be a rational capital-allocation decision.
Microsoft has enormous AI demand, but it also has choices. If a site cannot get power for years, the sensible move is to put capital somewhere that can. There is no trophy for being loyal to a delayed data centre. There is only lost time.
That should worry Nscale investors more, not less.
The big cloud platforms have alternatives. They can build themselves, use another provider, shift workloads geographically or delay a project. A neocloud with a concentrated customer base has far fewer options when a major client changes course.
That is the asymmetry. Microsoft can reroute. Nscale has to replace demand, capital and credibility while still building the asset.
The same principle applies to every operator reading this. If one customer can materially alter your plan by changing theirs, you do not have scale yet. You have customer concentration with a nice haircut.
The market is already sniffing this out. Nscale raised US$3.36 billion in pre-IPO financing in September, with Third Point leading a group that included Nvidia, Apollo Global Management, Citadel, Hudson Bay Capital and Abu Dhabi Investment Council. Serious backers, obviously.
But serious investors do not remove execution risk. They merely prove that sophisticated people can see the opportunity — and are willing to accept the risk at a particular price.
Those are not the same thing.
Why the valuation fight matters more than the headline valuation
Nscale has not yet begun formal IPO marketing, despite filing publicly on September 18. The valuation discussion is therefore not settled.
That is good. Public markets are supposed to do this job. They are supposed to take a glossy growth story, identify the dependencies, price the risks and occasionally tell founders that their favourite number is ridiculous.
A US$30 billion valuation says the market should reward Nscale for future scale, strategic positioning and a giant book of AI demand. A US$16 billion-to-US$19 billion valuation says buyers want a larger discount for customer concentration, financing exposure, construction execution and power risk.
Neither number is holy. The gap is the message.
When there is a gulf between what sellers want and what buyers will pay, it usually means the story is ahead of the evidence. The solution is not better investor relations. The solution is more delivered capacity, more diversified customers, cleaner unit economics and fewer dependencies that can be exposed by one cancelled site.
What this means for you
Whether you are a founder, operator or investor, take three useful lessons from Nscale’s very public headache.
1. Separate demand from deliverability.
When you evaluate a business, ask two different questions: “Do customers want this?” and “Can this company deliver it profitably, on time?” Most people only ask the first because it is more exciting. The second is where businesses live or die.
2. Treat concentration as debt, even when it is not on the balance sheet.
If two customers represent most of your future revenue, you have a hidden liability. Price it, plan for it and actively reduce it. For operators, that means building a sales pipeline before you feel ready. For investors, it means refusing to value contracted revenue at full value when one client can wreck the plan.
3. Find the boring bottleneck before your competitors do.
In your business, it may not be power. It may be freight, compliance, staff training, payments, inventory, permits, data quality or plain old working capital. The glamorous part of the business gets copied. The boring constraint decides who actually wins.
AI will keep creating enormous businesses. I am not betting against that. But the winners will not be the companies with the loudest AI slogans or the biggest announced contracts.
They will be the ones that can turn electricity, capital and customer promises into dependable output — before the bloke next door does.