Construction AI: the Guardian finds nobody visited Stargate UK's flagship site, and Microsoft's HQ makes way for a gas-powered data centre

The Guardian reported on Friday 4 July that OpenAI apparently never visited Cobalt Park, the site near Newcastle that was meant to host the primary Stargate UK build, before the project was announced. Stargate UK, the partnership between OpenAI, Nvidia and British cloud provider Nscale, was unveiled in September 2025 during Donald Trump's state visit, promising up to 8,000 Nvidia GPUs in the North East by Q1 2026, scalable to 31,000, inside a £31bn package the government presented as proof of Britain's AI competitiveness. A source close to the project described the touted plans, worth around £20bn on paper, as a PR stunt. No planning applications were lodged at Cobalt Park. No construction began.

This lands on top of what the Guardian's Aisha Down found back in March, when she visited Nscale's other supposed supercomputer site in Loughton, Essex, and found a working scaffolding yard, still registered to a different owner despite Nscale saying publicly it had bought the land. The same investigation found the government had press-released a £1.9bn investment contract with Nscale that had never been signed, and the Department for Science, Innovation and Technology admitted it was "not playing an active role in auditing these commitments." OpenAI paused the whole thing in April, citing UK industrial electricity at roughly four times the price of the US and the Nordics. Nscale, meanwhile, put €695m into Portugal to supply 66,000 GPUs to a Microsoft campus at Sines. The money went where the power was cheap.

So here's the bit for anyone whose business plan leans on the data-centre boom. I've been calling this pipeline a major UK workload all year, and I still think that's right, the Savills and Barbour ABI numbers on actual starts are real. But this story is the clearest evidence yet that the announced pipeline and the buildable pipeline are different documents. The government counts pledges at the press-release stage and verifies nothing. A £20bn scheme sat in the national AI numbers for ten months without a single planning form. If a client did that to you on a framework bid, you'd have a word for it.

The procurement filter: Before any data-centre scheme goes in your pipeline forecast, check three things you can verify independently: a lodged planning application, a completed land transaction, and a dated grid connection. No paperwork, no pipeline.

The same week the paper pipeline unravelled, a real one advanced. On 2 July, developers announced a proposal to redevelop Microsoft's UK headquarters campus at Thames Valley Park, near Reading, as a 72MW data centre, with a public consultation running 3pm to 7pm on Tuesday 7 July at Pearson Hall in Sonning. Microsoft plans to move its staff into the existing HERE building on the same park by August 2027 to free the site. The developers say the scheme responds to falling demand for office space, which is quite the epitaph for a headquarters campus, and that construction would support about 250 jobs with around 115 permanent roles once operational. Those are developer figures, so treat them as a prospectus rather than a payroll.

What makes this one worth your attention is the power strategy. Rather than joining the grid connection queue that has stalled schemes across the South East, the proposal runs on natural gas fuel cells "while additional grid capacity is brought forward," with the kit described as future-fuel ready for biomethane or hydrogen as the network greens. I'm not sure "future-fuel ready" survives contact with a planning committee in 2026, and the gas angle will draw objections, rightly. But the direction is honest: the developer has looked at the single biggest constraint on UK data-centre delivery and designed around it rather than waiting for it.

Put Friday's Stargate story next to this one and you get a working definition of a real scheme. Named site, named occupier strategy, a power answer on day one, and a consultation you could walk into on Tuesday afternoon. For the contractors and M&E firms circling this market, that contrast is the whole game. The schemes that solve power up front are the ones that will actually appoint people. That's what it's about.

Worth doing: If data centres are a target sector, get someone to the consultation events. The pre-planning stage is where you learn a scheme's power strategy, programme and procurement route months before the OJEU-watchers do.

A dated update to Thursday's brief, where Anthropic had accused Qwen-linked operators of running nearly 29 million exchanges through Claude to copy it. The reply has arrived. From 10 July, Alibaba is banning employees from using Claude Code, classifying it as high-risk software and pointing staff at its own Qoder tool instead, as reported by TechCrunch and the South China Morning Post on 4 July. The trigger: security researchers found code in the tool that read system time zones and proxy configurations, scanned for keywords associated with Alibaba, ByteDance and Baidu, and sent environmental data back to Anthropic using steganographic techniques, which is a fancy way of saying it hid what it was doing.

Anthropic's Thariq Shihipar says this was an experiment launched in March to prevent account abuse by unauthorised resellers and to protect against distillation, which, given what Anthropic told the Senate about those 25,000 fraudulent accounts, is at least a coherent story. But hold both halves of last week together. The provenance question I raised on Thursday was whether you could trust a Chinese open-weights model with your project data. What we've learned since is that a flagship American coding tool was quietly fingerprinting its users for four months and nobody outside noticed until researchers went looking. The trust question doesn't have a nationality. It has an audit trail, or it doesn't.

For a UK construction business the exposure is smaller but the same shape. Your teams are adopting AI coding assistants, document tools and copilots faster than anyone is reading the network traffic. Most of it is benign telemetry. The point is that someone in your business should be able to say what these tools send home, because the person who finds out from a journalist is the person explaining it to a client under an NDA.

For your board pack: Add one line to your AI tooling register: what does this tool transmit, and who checked? If nobody can answer, that's the finding.

Three stories, one habit. A £20bn national flagship turned out to have no planning application behind it; a quieter 72MW scheme turned up with a consultation date and a power strategy; and a coding tool used by thousands of developers turned out to be checking where they lived. In each case the difference between the headline and the truth was a document somebody could have checked: a planning register, a land title, a network log. Nobody checked until a journalist or a researcher did.

Construction people are actually well set up for this. You wouldn't start on site without a signed contract, and you wouldn't accept a subcontractor's insurance on their say-so. Apply the same instinct to AI announcements, whether it's a government pledging megawatts or a vendor pledging privacy. The evidence standard you already use for money is the right one for models.

A practical step: Pick your top three assumed data-centre prospects this week and pull their planning status from the local authority portal. It's an hour's work, and it turns a hunch into a forecast.