Construction AI: the grid, not the planners, is what's holding up Britain's biggest AI data centre

Here's the story that tells you where the real bottleneck sits. Nscale's data centre at Loughton in Essex was announced in September 2025 during the US state visit, billed as Britain's biggest AI supercomputer, with Microsoft signed up to take the capacity for Azure in the UK and room for up to 45,000 Nvidia GB200 GPUs. The council objected, the government leaned in, planning got sorted. And the thing is still slipping. Reporting this month has the go-live moving from a Q4 2026 target into 2027, because the grid connection for the site, rising to 90MW, won't be ready in time. Nscale is reported to be in talks with a California fuel-cell maker about generating its own electricity on site instead.

So the constraint everyone assumed would be planning turned out to be power. That's worth sitting with if you're bidding or pricing this work. The £100bn UK data-centre pipeline is real, we covered the 21 schemes actually on site a couple of days ago, but a scheme with £2bn, Microsoft and Nvidia behind it still can't buy its way past a connection queue. What that means on the programme is simple: the date that decides whether you deliver isn't on the planning consent, it's on the grid connection offer, and those two dates can be years apart.

The comparison only goes so far, but it's a bit like fitting out a beautiful kitchen and then finding the water main won't be connected for eighteen months. Every other trade is ready and nothing can commission. If you're carrying data-centre work, or subcontracting into it, get the connection date onto the risk register above the planning date, and ask the developer for the DNO offer, not the promise. That's the number that's actually moving.

xAI released Grok 4.7 on 21 September, its latest frontier model, pitched at coding and long-running agentic tasks. The interesting bit for us isn't the 2.1 trillion parameters or the benchmark chart. It's that on the same day, Grok 4.7 became a selectable model inside GitHub Copilot for the Pro, Business and Enterprise tiers, and it's live in Cursor and the xAI API too. Standard API pricing is $2 per million input tokens and $6 per million output, the same base rates as Grok 4.6, so xAI isn't charging a premium for the jump. Those are vendor-published figures, worth flagging as such.

What that means in practice is that if your business has developers or a data team building anything on top of Copilot or Cursor, they can now route your work through a different model with a couple of clicks, and they don't need to ask anyone to do it. The model itself is fine. That's not the point. The point is governance: which model sees your project data, where it runs, and whether that's a decision your business made on purpose or one that happened by default because it appeared in a dropdown.

I think this is the quiet shift that catches firms out. Everyone's braced for the big platform announcements, the Autodesk-in-Las-Vegas moment we looked at last week. Meanwhile the models are just showing up inside tools people already have open. Decide your line on model choice and data residency now, write it down, and tell the people doing the building. That's the practical bit.

Two adoption signals worth logging. First, Gilbane has selected Trunk Tools as its enterprise site-AI partner and is rolling its agents, submittal checking, text and standard-operating-procedure bots, across more than 200 projects. That's one of the largest single-vendor jobsite deployments announced this year. Trunk Tools is a narrow set of point agents rather than a broad platform, so it sits here under Tools rather than as a challenger, but a contractor putting AI agents on 200-plus live jobs is a real number and a real bet. Second, closer to home, the Contracting Awards 2026 run a dedicated AI category for the first time, with winners announced on 30 September at the Park Plaza Riverbank in London. When the trade awards add a category, the adoption curve has already turned.

So what do you actually do with either of these. The same thing you'd do with Grok landing in Copilot or a data centre slipping over a connection: judge the claim against your own site. A 200-project rollout tells you Gilbane believes it. It doesn't tell you what it did to their rework rate or their RFI turnaround, and those are the only numbers that matter to the person doing the paperwork. Give any new agent one scope, one project, one month, and measure the before and after. That's what it's about.

For your board pack: Ask for one measured pilot result before any firm-wide AI rollout, and make the metric something the site manager already tracks.