A data center under construction beside power transmission towers, with a timeline marking progress from 2024 to 2033.
Microsoft's exit from one planned UK AI data center is a useful case study in who actually carries the risk when AI infrastructure gets built. The Motley Fool's piece, republished on AOL, uses the exit to sort AI suppliers into two groups. One group sells the hardware. The other builds and runs the data centers and waits for customers to pay.

The framing is useful but too tidy. The filings and reporting below show that the exposure shifts with each deal's structure. Nothing here is a buy or sell call.

A data center under construction beside power transmission towers, with a timeline marking progress from 2024 to 2033. What actually happened at Loughton​

Reuters reported that Microsoft pulled out of a deal to anchor a flagship Nscale data center in Loughton, Essex. UK Power Networks reportedly told Nscale the Essex site would not get sufficient power until the early to mid-2030s. Microsoft's own statement was narrow. A spokesperson said the decision covers a single site.

The scope of the exit matters:

  • Reuters said the site was due to start operating early next year. Its plan was 23,040 Nvidia GB300 GPUs providing 50 megawatts of computing capacity for Microsoft's Azure services in Britain.
  • Nscale says the "London-wide grid issue has no impact on our other contracts." It also says it is exploring alternative power sources for the site.
  • Microsoft and Nscale are also working together on projects in Norway, Portugal and the US. IT Pro adds that the Norway facility is believed to be behind schedule too.
  • One outlet noted that the specific reason for Microsoft's withdrawal was not publicly stated. The power delay is the reported backdrop, not a Microsoft-confirmed cause.

So the evidence supports a site-specific power and schedule failure. It does not show that Microsoft has lost faith in AI demand or abandoned UK investment. Data Center Knowledge draws the broader lesson: customer demand and financing commitments do not guarantee that electricity will be supplied in time to meet commercial schedules.

The hardware sellers: paid on delivery, not risk-free​

Nvidia, Broadcom and AMD sit on the equipment side. The article's core point is that they are paid for products they supply rather than for operating data centers. That is broadly right, but each carries its own exposure.

  • Nvidia: The article notes it generally recognizes revenue on delivery. It also says Nvidia has invested in more than 200 companies, which ties its fortunes to some customers.
  • Broadcom: The article says the risk is customer concentration. Custom accelerators built for one buyer can't easily be sold to another.
  • AMD: AMD's October 2025 announcement with OpenAI shows a hybrid deal. The companies described a 6-gigawatt, multi-year agreement. AMD's own release says it issued OpenAI a warrant for up to 160 million AMD shares that vests as milestones are met. The first tranche vests with the initial 1 GW deployment, with more tranches as purchases scale to 6 GW. Vesting is also tied to AMD share-price targets.

The AMD warrant is not unconditional, and the 6 GW plan is not already delivered. Still, the structure shows a chip vendor sharing some of the customer's success risk.

The operators: contracts are not cash​

Oracle, CoreWeave and Nebius build or rent capacity and recover the money over time. The Loughton episode shows how a perfectly good customer contract can stall on something as mundane as a grid connection.

Oracle​

Oracle's quarterly filing for the period ended August 31, 2026 shows the scale. Remaining performance obligations were $664 billion, up from $455 billion a year earlier. Cash spent on capital expenditures was $28.5 billion in the quarter, against $8.5 billion a year earlier. Oracle attributes the rise mainly to data-center expansion.

RPO is contracted work not yet recognized as revenue. It is not cash collected, and it says nothing by itself about margin or timing.

There is also a qualification to the "operator funds everything" story. In its fiscal 2026 third-quarter release, Oracle said most of the RPO increase that quarter related to large AI contracts where it did not expect to raise incremental funds. In those cases the equipment was funded by customer prepayments, or customers bought the GPUs and supplied them. That describes those contracts, not every Oracle deal.

CoreWeave​

Under a September 2025 order form with an initial value of $6.3 billion, Nvidia agreed to buy CoreWeave's residual unsold capacity through April 2032. This is subject to termination provisions and delivery and service-availability requirements. It is a conditional backstop, not a guarantee. It also means Nvidia is simultaneously supplier, investor and backstop customer. I could not retrieve the filing itself in this session, so that detail rests on the research brief and the article's own mention of purchases through 2032.

Nebius​

Nebius's 2025 annual report describes a Microsoft agreement effective September 7, 2025. It covers dedicated GPU capacity at its Vineland, New Jersey data center over five years. The filing also shows $982.5 million of customer advances under strategic agreements, which ease upfront funding. It still has to deliver the capacity.

Nebius also warns that AI demand may prove unsustainable, and that reduced customer spending could materially harm it. It lists excess industry capacity, pricing pressure, renegotiations and impairments among the risks. Those are company-disclosed risks, not predictions.

Alphabet: both sides, with deep pockets​

Alphabet sells cloud capacity to outsiders, including Nvidia GPUs and its own TPUs. It also uses the same infrastructure internally. Its 10-K reports $91.4 billion of 2025 capital expenditures and expects 2026 technical-infrastructure investment to rise significantly. It also notes that data-center projects can take years and that assets under construction may sit idle until they are ready. Established profit engines make that easier to carry than it is for a pure-play neocloud.

Why neoclouds feel delays hardest​

Channel Insider makes a sharp point about the economics. Neoclouds can be particularly exposed to delays because they often have concentrated customer bases, higher capital costs, and equipment plans tied to specific deployments. It adds that a site that opens years later than planned has less time to earn returns on expensive accelerators before newer generations change pricing and performance expectations.

Power is a wider problem than one site. Reuters reports grid queues delaying data-center projects by as long as 13 years. IT Pro cites Savills data showing only 850 MW of power capacity delivered across EMEA since the start of 2025, down 11% from the prior comparable period.

A practical checklist for reading any AI capacity announcement​

IT teams relying on cloud AI capacity, and anyone reading vendor press releases, can use four tests:

  1. Order versus delivery. Has hardware actually shipped, and has service gone live?
  2. Backlog versus cash. RPO and "contracted value" are not revenue or collections.
  3. Who funds the equipment. Check for customer prepayments or customer-supplied GPUs before assuming the operator carries the full capex.
  4. Planned versus energized. Treat megawatt figures as real only when the grid connection has firm dates.

Teams planning Azure or other cloud AI deployments in a specific region should also avoid hard-wiring site-specific capacity into delivery plans until power and timelines are firm. One commentator makes a similar point about Loughton.

Bottom line​

The Motley Fool's main idea holds up: announced AI deals say little about who has been paid and who is still waiting. The caveat is that structure matters more than category. Hardware vendors face concentration and credit risk, and some use equity incentives to secure orders. Operators may be cushioned by prepayments, customer-supplied equipment or backstops. And none of it works without electricity. Loughton shows that even a blue-chip anchor tenant can leave when the power isn't coming.

 

References

  1. As AI Labs Burn Cash, the Real Question Is Which Suppliers Get Paid - AOL.com AOL.com 2026-10-25T04:20:00+00:00
  2. Nscale Loses Microsoft Deal as Power Delays Threaten AI datacenterknowledge.com
  3. Microsoft Pulls Out of Nscale’s UK AI Data Center Project Over Grid Delays channelinsider.com