Microsoft Walked Away From 23,040 GPUs Because the Grid Had No Power

Daily Insights · Saturday, October 24, 2026

The obstacle is simple: the world is building AI factories faster than the electrical grid can connect them. On October 18–19, 2026, Reuters confirmed what The Telegraph first reported: Microsoft terminated its agreement to anchor Nscale's planned data center in Loughton, Essex, because UK grid officials told Nscale that adequate power may not be available until the early-to-mid 2030s. The facility was supposed to house 23,040 Nvidia GB300 GPUs and deliver 50 MW of AI capacity to Microsoft Azure in 2027.

What Actually Happened

Loughton was not just a data center. It was a promise: 50 MW of capacity, expandable to 90 MW, 23,040 high-end accelerators, and a named hyperscaler anchor. Microsoft had already committed a portion of its reported $30 billion UK investment to the site. The turn came when the distribution network operator — UK Power Networks — indicated the site could not receive sufficient electricity until the 2030s because it is waiting on national transmission network upgrades it does not control.

• Microsoft: "This decision relates to a single site." It says it remains committed to the broader Nscale partnership.
• Nscale: "We remain committed to both the Loughton project and the UK," and is exploring alternative power sources plus other UK expansion sites.
• The facility was originally scheduled to begin operating in early 2027.

Why the Grid Is the Real Bottleneck

The Loughton failure is a symptom of a structural mismatch. UK data center demand now far exceeds grid connection capacity. According to The Stack, 315 data center projects are queuing for connection, with roughly 73 GW of data-center demand versus about 45 GW of peak UK demand in 2025/26. The UK's own transmission and distribution upgrades cannot keep pace with the build-out being announced by hyperscalers.

This is not a UK-only problem. Every frontier AI market — London, the US, the Middle East, India — is discovering the same constraint: compute capacity is now limited by interconnection, not chips or capital.

How This Changes the Underwriting Lens

For smaller investors, the lesson is that announced capacity is not contracted capacity. Three diligence questions should now be standard for any AI infrastructure exposure:

1. What is the grid connection status? Has the DNO/TSO issued a firm connection offer, and what is the queue position?
2. Who controls the generation? Is the project relying on the grid, on-site generation, or a hybrid? On-site generation changes the capex profile and the timeline.
3. What is the anchor contract duration? A single anchor leaving, as at Loughton, can strand a large share of the asset's economics.

Nscale itself is moving. It remains committed to Loughton, is exploring on-site generation, and is reportedly in talks for a second ~$45 billion, 460 MW Anthropic facility at its Monarch campus in West Virginia — a site with a dedicated electrical grid and on-site generation. That contrast is telling: the same company is pivoting its largest new commitment toward a location where power is a design feature, not a queue item.

Why It Matters Now

AI compute is now the bottleneck for many frontier models. Power is the bottleneck for AI compute. The Loughton termination is the clearest public signal yet that interconnection risk can break even an anchor-tenant, multi-tenant AI campus before a single GPU is installed. It also raises the bar for any pre-IPO or private-market infrastructure exposure that depends on unconnected grid capacity.

The downside is direct: stranded assets, delayed revenue, and potential dilution in pre-IPO rounds where investors priced in near-term utilization. The upside is structural: every dollar spent on grid-hardened sites, on-site generation, and long-duration storage becomes a defensible moat. The market will increasingly distinguish between "we have chips" and "we have power."

Practical Takeaways

• Track grid-connection news as a leading indicator for AI infrastructure valuations, not just an operational detail.
• Prefer companies with dedicated transmission or on-site generation over those waiting on utility upgrades.
• Watch for secondary discounts in pre-connection AI assets; the Loughton re-tenanting market may offer entry points for operators who can wait.
• Pair every AI infrastructure thesis with a named, dated power-source risk and a mitigant.

How Ad Valorem Supports This Work

Ad Valorem focuses on overlooked, mispriced opportunities in private markets and technology investing. As grid constraints reshape AI infrastructure, we look for secondaries, restructured assets, and special situations where the market has priced in the bottleneck but not the solution. Our work is research, analysis, and deal sourcing — not investment advice. For informational purposes only.

Support Ad Valorem

Disclosure: This content is AI-generated. For informational purposes only. Not investment advice.