Microsoft’s AI expansion just hit a physical limit

Microsoft (MSFT) can buy more artificial intelligence chips. However, it cannot build an electricity grid overnight.

That problem is becoming harder to ignore.

Microsoft has withdrawn from a planned agreement for capacity at Nscale’s AI data center in Loughton, England, Reuters reported, citing the Telegraph. The project has suffered repeated delays, while grid officials have told Nscale that power for the site may not be available until the 2030s.

Microsoft declined to comment, and Reuters reported that it has not disclosed the reason for its withdrawal, so grid delays are not the confirmed cause. Nscale says that other contracts and its relationship with Microsoft remain unaffected.

However, the incident highlights a growing restriction on the explosion of AI.

Investors were concerned for years about Microsoft’s ability to acquire enough Nvidia (NVDA) chips to satisfy demand for Azure and AI services. Physical infrastructure, such as grid connections, transmission, and power production, is now becoming just as crucial.

It’s a significant hurdle because Microsoft expects roughly $175 billion in calendar 2026 capital expenditures as it races to bring more computing capacity online. The company has repeatedly said demand remains greater than available capacity.

Microsoft’s abandoned AI project was supposed to be enormous

The Loughton location was no small experiment. In September 2025, Nscale and Microsoft declared the project to be the largest AI supercomputer in the United Kingdom.

The project initially anticipated 23,040 Nvidia GB300 graphics processing units in the first quarter of 2027, according to Nscale. The planned campus was built for 50 megawatts of initial AI capacity, scalable to 90 megawatts.

Additionally, the initiative was a perfect match for Microsoft’s larger U.K. investment plan.

Microsoft says it is investing $30 billion, or about £23 billion, in the U.K., including $15 billion of capital spending on cloud and artificial intelligence infrastructure. Its description of that investment specifically included the Loughton supercomputer alongside new data centers in West London and South Wales.

Nscale has not abandoned Loughton. The company told Reuters it remains committed to the site and is exploring alternative power sources. Its other Microsoft projects are also continuing.

Microsoft and Nscale have international deals. Nscale confirmed contracts for 200,000 Nvidia GB300 GPUs in Europe, the U.S. (Texas), and Portugal in October 2025. The companies now partner in Norway and Portugal.

This evidence doesn’t indicate that Microsoft is abandoning AI infrastructure or Nscale. Instead, it is proof that location is becoming more important.

Even with clients, processors, and funding arranged, a data center still faces a challenge that no software business can solve: getting enough energy to the right location at the right time.

“Power used to be something the system simply plugged into. Now, it is something we design around,” Microsoft said.

Microsoft made that observation in September while explaining how AI infrastructure is forcing the company to redesign everything from chips to electricity distribution.

Microsoft’s next AI problem is hiding outside the data center.

Bloomberg / Getty Images

Power is becoming a bigger part of Microsoft’s AI equation

The issue facing Microsoft is not a lack of demand.

Microsoft said earlier this year that it expected to remain capacity constrained through at least 2026 even while rapidly bringing graphics processing units, central processing units, and storage online. Management said stronger customer demand and product usage continued to justify its infrastructure spending.

More Microsoft:

Microsoft’s $665 target hinges on a new AI advantage

Meta and Microsoft just sent employees a memo about Anthropic

Microsoft just sent a strong message to Anthropic

What is changing is the nature of the constraint.

According to Microsoft, entire campuses can function at the gigawatt scale, and AI data-center racks that previously needed tens of kilowatts may now require hundreds of kilowatts. Therefore, power distribution and availability must be taken into account prior to building a facility, not after servers are installed.

The same problem is showing up in the U.S., as AI and data-center demand put increasing strain on the PJM Interconnection, the electricity market covering 13 states and Washington, D.C.

PJM expects data-center demand to nearly triple to 29 gigawatts by 2030, while regulators and utilities are debating how quickly new power generation and transmission can be built, Reuters reported.

Transmission upgrades are already delaying 37.7 gigawatts of planned new generation in the area, according to another Reuters analysis.

For hyperscalers, this presents a difficult issue.

AI demand can increase much faster than utilities can build transmission lines, substations, and generating capacity. Data centers can be constructed in years. Major grid projects can take much longer.

That does not necessarily imply that AI spending is failing for Microsoft shareholders. However, it changes what investors should focus on.

The debate has centered heavily on whether Microsoft is spending too much on chips and data centers. Increasingly, another question matters: How quickly can those billions of dollars actually become usable computing capacity?

Microsoft’s AI returns may increasingly depend on the grid

Microsoft’s investment argument is unlikely to change as a result of the Loughton loss.

Microsoft has a vast worldwide network of data centers, many infrastructure partners, and substantial financial resources. To move capacity to areas where electricity is simpler to obtain, Nscale itself has projects assisting Microsoft in the United States, Norway, and Portugal.

However, this freedom has its drawbacks.

Relocating an AI task abroad may raise concerns around connection, latency, data sovereignty, and legislation. It also takes time to build someplace else. Additionally, other hyperscalers are pursuing many of the same power-rich areas.

That’s why there’s more to this tale than just one Essex data center.

Microsoft’s AI strategy increasingly relies on converting previously unheard-of capital expenditures into Azure capacity that generates income. Investors are already closely examining whether the industry’s massive expenditures will justify the profits from AI.

That conversion may be slowed by electricity.

Nscale is itself an example of how much money is flowing into the infrastructure race. The Nvidia-backed company raised $2 billion at a $14.6 billion valuation in March and has since filed for a U.S. IPO.

Reuters said Nscale now has more than $103 billion in contracted value, with Microsoft and Anthropic among its key customers.

Power availability is more than just a technical issue because of the magnitude of such obligations.

I find it noteworthy that AI’s limited resources are always evolving. Graphics processing units were the first need. Data-center capacity came next. Electricity and the infrastructure required to provide it are growing increasingly important.

For Nvidia chips, Microsoft is able to pay a bigger check. It is able to finance another data center complex or execute another contract.

A separate issue is obtaining a grid connection, which could not be available until the 2030s.

For Microsoft investors, this implies that power production, transmission lines, and substations may be just as important to the next stage of the AI boom as the next generation of processors.

Related: UBS resets Microsoft stock price target for investors

#Microsofts #expansion #hit #physical #limit