When an AI data center is announced, everyone cheers. News that tens of thousands of NVIDIA GPUs will go in, that tens of MW of power has been secured, and that a world-class big tech company has signed on as a customer looks like a symbol of the AI boom.
But in the UK, a crack has appeared in one of these glamorous plans. Reports emerged that Microsoft has withdrawn from its contract on the Loughton project in Essex, run by UK data center company Nscale.
The key wasn't a shortage of GPUs. According to reports, the grid connection needed there could be delayed until the early-to-mid 2030s at the earliest.
If a facility targeted to run in 2027 can't get electricity on time, no amount of top GPUs piled up will let it run AI services. This incident poses a rather uncomfortable question for investors worried about an AI bubble.
Even if AI demand is real, can the infrastructure that turns that demand into money be finished on time?
What Happened in the UK?
✦ The Loughton Data Center Plan
Per Nscale's announcement
Nscale announced its Loughton AI campus plan with Microsoft in September 2025. The facility was to start at 50MW of initial AI compute capacity, expandable to 90MW, with about 23,040 NVIDIA GB300 GPUs deployed for use in Microsoft's Azure services.
Nscale's announcement explained that the GPUs were due to be supplied in the first quarter of 2027.
This is no small server room. 50MW of power is the foundation needed to keep running large-scale AI compute, and from a customer's view, when you can actually start using it matters as much as the number of GPUs.
Nscale's earlier announcement gave a fourth-quarter 2026 operating target, but the follow-up plan presented the GPU supply timing as the first quarter of 2027. The schedule had already been slipping.
✦ Timeline So Far
As of October 19, 2026
Then in September 2026, reports emerged that power supply problems could delay the facility's operation significantly. Then on October 18, news spread that Microsoft had pulled out of the contract for this site.
According to Guardian and Reuters reports, the point at which the grid operator could supply the power this site needs could be pushed back to the early-to-mid 2030s.
There's one thing to distinguish here. Microsoft hasn't publicly stated an official reason for the withdrawal. Still, by the reported schedule, a site expected to run in 2027 would have to wait years more for electricity.
Among the circumstances known so far, the grid schedule looks like the most likely background to the contract withdrawal.
Why Does Late Power Tangle an AI Business?
A data center doesn't end with putting up a building and installing GPUs. Power supply, cooling equipment, networking, and operating software all have to be in place before you can sell compute services to customers.
If you spent money on equipment and land first but can't get electricity, investment costs pile up while revenue doesn't arrive as planned.
If Azure customers want Microsoft's AI services, they need compute capacity they can use right now. If the contracted facility only comes online years later, Microsoft may have to find data centers in other regions or other providers.
It's fortunate if that gets resolved quickly. But an AI data center isn't something you finish by switching on servers in the cloud like a software business. It's tied to the physical realities of the grid, land, and permits.
The UK government requiring AI Growth Zone candidates to demonstrate access to at least 500MW of power by 2030 also shows that national regulation around AI data centers is no small matter.
Nscale is also in a tough spot. With Microsoft dropping off this project's customer list, it has to sell the empty capacity to other customers or solve the power problem.
Nscale said it is continuing with the project and its UK business and looking for alternative power sources, and explained that its other business relationships with Microsoft remain fine. But it's clear that the homework Nscale has to solve just got bigger.
Worries About an AI Bubble
One project's contract withdrawal isn't evidence that the entire AI industry is a bubble. But it does show where the risks lie that investors worried about an AI bubble should watch.
When investing in AI-related companies, you shouldn't consider only whether there are customers who want to use that AI.
You should consider whether the related infrastructure comes online on time, whether usage revenue can cover the investment and interest, and whether there's cash to hold out if expected income is delayed.
✦ AI Infrastructure Spending and Payback Risk
Gartner's estimated scenario
Gartner estimated that AI infrastructure spending will reach about $1 trillion in 2026, and that an additional $6 trillion could be invested in 2027 to 2030.
It also projected that if the current pace of investment continues, over $3.5 trillion of infrastructure investment could be exposed to impairment risk by the end of 2030.
By Gartner's calculation, for data center owners to cover their investment, financing costs, and a 12% return on invested capital, they would need to generate a cumulative $8.5 trillion in revenue by 2030.
These numbers are Gartner's estimated scenario. Still, the question is clear. Is money coming in through AI services as fast as data centers are multiplying? And can the related infrastructure get power on time and come online?
The Bond Market Is Already Asking About the Cost
✦ CoreWeave Bond Yields
Per Reuters, for specific bonds
✦ CoreWeave (CRWV, Nasdaq) Closing Prices
Sep 28 to Oct 16, US Eastern Time close (USD)
CoreWeave is a so-called 'neocloud' company that provides GPU compute to AI developers. Its stock and bonds were shaken badly in mid-October.
According to Reuters, some bonds that yielded 9.675% at their June issuance saw their secondary-market yield surge to 15% in the week of October 16. It means investors have begun demanding higher rewards when lending to such infrastructure companies.
Reflecting that worry, the stock also plunged over the same period. CoreWeave's close fell about 27.8%, from $85.07 on September 28 to $61.39 on October 16. On October 14 in particular it dropped 15.46% in a single day, and it fell another 11.91% on the 16th.
In a July analysis, Reuters projected that the combined capital expenditures of Microsoft, Alphabet, Amazon, Meta, and Oracle could exceed free cash flow in 2027.
The point is that AI generating revenue and the money invested in AI being recovered as enough cash are different matters.
That Doesn't Mean AI Demand Is Fake
✦ Microsoft's Latest Fiscal-Year Filing
Cloud revenue growth and cost increase
In Microsoft's latest fiscal-year filing, Azure and other cloud services revenue grew 41% from the previous year. At the same time, as AI infrastructure investment grew, the Intelligent Cloud segment's cost of revenue also rose 44%.
Real demand exists, and at the same time cost and investment burdens are growing. What this Loughton contract withdrawal tells us is closer to "not every data center plan works out in time and cost, even when AI demand exists" than to "nobody uses AI."
In this case, Microsoft didn't cling to a facility whose power would arrive late. There was no official announcement, but it probably had other options for securing compute capacity elsewhere.
But investors are now raising, more and more strongly, the questions "When does the electricity come in, who pays the costs, and when does the cash come back?" rather than just the number of GPUs bought.
This isn't investment advice. You know that, right? Let's each protect our own hard-earned money.
That's all for this post. See you next time.
Related posts
- CoreWeave's $104.2B Backlog, So Why Is the Stock Falling?
- Bitcoin Firm Strategy (MSTR) at Risk of MSCI Index Removal
- Samsung & SK hynix Fall Together Despite the HBM Boom
🌐 Korean version: 마이크로소프트가 데이터센터 계약에서 조용히 발을 뺐다...? 이유가 뭔지 알아보자
🌐 Japanese version: マイクロソフト(Microsoft)、データセンター契約から静かに撤退?理由を探る

No comments:
No comments yet. Want to leave the first one?
Post a Comment