Microsoft Walks Away From the UK's Would-Be Largest AI Data Centre

The Telegraph reported on October 18 that Microsoft has axed its deal to anchor Nscale’s Loughton, Essex data centre, the site announced 13 months ago as the future home of the UK’s largest AI supercomputer. Microsoft was the facility’s only announced customer, committed to renting more than 23,000 Nvidia chips there, and its withdrawal leaves the 2 billion pound project in apparent limbo. The proximate cause, per the report and months of prior coverage: the site never secured its power connection, and UK grid queues stretch for years.

Thirteen months from state-visit showpiece to cancellation

The arc is worth replaying because it captures the whole infrastructure cycle. In September 2025, the Loughton supercomputer was announced with Jensen Huang and Microsoft leadership on stage in the UK during President Trump’s state visit, part of the 31 billion pound US-UK Prosperity Deal that promised up to 60,000 Nvidia Grace Blackwell Ultra chips to Nscale across its sites. The BBC reported the Essex site would create up to 250 permanent jobs and open in 2026. By September 2026, the Guardian reported the supercomputer’s launch had slipped, potentially into the mid-2030s, with grid connection the binding constraint. Now the only announced customer is gone. Nscale is not a small company failing; it also signed a separate Microsoft deal for roughly 200,000 GB300 GPUs across its US and European sites. The Essex project died specifically because a 2 billion pound building without electricity is just a building.

The pattern behind one cancelled site

One caveat before reading this as a demand collapse: Loughton’s problem was grid connection queues, an infrastructure-planning failure, not a demand signal. But the cancellation lands in a market showing both kinds of stress at once. In the US, the Data Center Watch Report documents 64 billion dollars in projects blocked or delayed by local opposition, 18 billion blocked outright, with at least 142 activist groups organized across 24 states and opposition that is genuinely bipartisan (55% Republican, 45% Democratic among officials taking public stances against 50 MW+ projects). The largest delayed project, Virginia’s 24.7 billion dollar PW Digital Gateway, has been tied up in lawsuits for years. Separately, the New York Times has reported the debt side of the squeeze: CoreWeave’s bond yields above 15% and some Oracle debt trading near 80 to 90 cents on the dollar, signs that lenders are repricing AI infrastructure risk even as equity markets wobble. Anthropic’s Q3 revenue beat and IPO repricing told the same story from the demand side last week: revenue is real, growth is decelerating, and every number downstream of hypergrowth assumptions gets rewritten.

What to watch

Three threads are worth tracking from here. First, Nscale’s other Microsoft commitment: the 200,000-GPU agreement across European and US sites is the company’s actual backbone, and whether Microsoft honors it while walking from Essex tells you whether this is one bad site or a portfolio-level repricing. Second, the UK’s AI superpower program, which leaned on exactly this kind of US-hyperscaler anchor deal; a replacement tenant for Loughton, or its conversion to another use, will indicate whether the 2025 deal pipeline survives contact with grid reality anywhere in Britain. Third, watch whether cancelled-but-funded sites like this get acquired by the compute brokers, because a 2 billion pound shell with grid interconnect progress is exactly what capacity-starved buyers might pick up cheap. The broader lesson for anyone planning AI capacity in 2026: the scarce resource is no longer chips or even capital, it is energized land, and the projects that survive are the ones that secured power before they announced customers.