Construction AI: McLaren puts robot dogs on UK sites, and Whitehall leant on a council to approve the Loughton data centre

McLaren Construction announced on Monday 6 July that it has partnered with FieldAI, the California robotics firm, to deploy autonomous quadruped robots across its UK sites. The first units are Boston Dynamics Spot robots running FieldAI's Field Foundation Models, and McLaren has already tested the setup on a job with some pedigree: the Passivhaus refurbishment of the London School of Economics' 35 Lincoln's Inn Fields building. The robots capture 360-degree imagery and point cloud data, verify progress, compare the as-built against the design model to flag deviations, patrol for safety compliance and run quality assurance checks. Construction News, Building, the Construction Enquirer and the Construction Index all carried it on the day, and the deal marks FieldAI's entry into the British market.

What makes this one different from the Spot-on-site pilots we've seen since 2020 is the autonomy claim. FieldAI says its models combine data-driven AI with physics-based reasoning, and crucially don't depend on prior maps, supporting infrastructure or pre-planned routes, so the robot adapts as the site changes rather than needing someone to re-teach it every week. The company says its technology is already working on hundreds of sites across Europe, Asia and North America. That's vendor-reported and I'd want to see the definition of "working" before I repeated it in a bid, but McLaren testing on a live Passivhaus refurb, one of the fussier building types for tolerances and airtightness detail, is a reasonable signal this is more than a photo opportunity.

So, the practical bit. A robot that walks your site on its own schedule, recording everything, is a wonderful progress record and a data governance question in a hi-vis vest. Where does the imagery live, who can pull it in a dispute, and what's your position when the footage happens to capture an operative doing something the HSE would rather they didn't? Those are contract and policy questions, and they're easier to answer before the first patrol than after the first incident. The site manager who used to walk the floor with a 360 camera on a stick gets that hour back, which is the point. But someone in the office inherits the archive, and that job needs a name on it.

The procurement filter: Before any autonomous capture lands on your site, get three things in writing from the vendor: where the data is stored, who holds the retention decision, and whether footage is used to train their models. A supplier who hesitates on the third one is telling you something.

Monday's other story sits one junction up the M11 from most of my readers. The Telegraph reported on 6 July, corroborated by Capacity, that Lord Stockwood, the Minister for Investment, wrote to Epping Forest district council's chief executive Andrew Small in November to "emphasise our strong support" for the Nscale and Microsoft data centre at Loughton, reminding him the project was announced by the Prime Minister during the September 2025 state visit. Planning correspondence from a minister direct to a council chief executive is, as the reporting politely puts it, highly unusual. The council granted permission for the enlarged scheme in June, despite objections from the town council over water and power supplies, from its own landscape and urban design teams, and from residents. Officers noted the letter in their approval and described data centres as critical infrastructure of national importance.

There's history here that regular readers will remember. Loughton is the same site the Guardian's phantom-investment reporting examined, part of the roughly £22bn Microsoft package from the state visit, and it has already slipped from being operational this year to 2027. Nscale, valued at around $15bn, says it remains committed. Meanwhile the pattern is spreading: a Kent data centre site was approved the week before through a route that bypassed local councillors, and the Local Government Association responded on Monday morning by insisting councils "must be treated as key partners" in AI growth zones and infrastructure planning rather than obstacles to be written to.

I'm genuinely of two minds, and I'll say so. The UK does need this capacity built, and the construction pipeline attached to it is one of the few unambiguous growth stories the sector has. But planning decisions that arrive pre-endorsed from Whitehall change the risk picture for everyone involved, because approvals won under pressure are approvals that invite judicial review, local resistance, and the sort of programme uncertainty that reaches the contractor eventually. What it means for the supply chain is simple enough: the work is real, the political backing is real, and so is the local opposition, so price the delay risk honestly. That's what it's about.

For your board pack: If data centre work is in your pipeline, split your risk register into technical risk and consent risk, and track the second one per scheme. Loughton shows consent can be won politically; it also shows why won-politically is its own risk category.

Google's Gemini 3.5 Pro entered the second week of July still in limited preview on Vertex AI, available to a small group of enterprise customers plus testers on LMArena, with no confirmed general availability date, no published benchmarks and no final pricing. Leaks reported in late June and early July point to a target around 17 July, after the model slipped from a June window. The headline specification is a 2 million token context window, roughly double the frontier norm, plus a Deep Think reasoning mode for Ultra subscribers. Google hasn't confirmed the date or the specs, so file all of it under reported rather than promised.

The reason for the delay is the interesting part. The reporting attributes the slip to three linked issues: token efficiency concerns flagged by early testers, coding performance not yet at flagship standard, and long-horizon multi-step reasoning falling short of what Google showed at I/O. Regular readers will recognise the first one, because it's the same story as last Friday's Sonnet 5 tokenizer item, where Anthropic's new model quietly counted about 30% more tokens for the same text. Across the industry, cost per completed task is overtaking benchmark scores as the number enterprise buyers actually evaluate, and Google appears to have decided that shipping a model that burns tokens faster than expected, as some users reported with the earlier 3.5 Flash, would cost it more than a fortnight's delay. I think that's the right call, and it tells you the labs know buyers have started doing the arithmetic.

For a construction business the guidance is unglamorous and unchanged. If your vendor's roadmap leans on Gemini 3.5 Pro for a mid-July feature, ask them what ships if it slips again. Evaluate any model on the cost to complete your task, a takeoff, a document check, a schedule query, not on the context window in the press release. And don't hold a procurement decision hostage to an unconfirmed launch date; generally available models that do the job today beat previewed models that might do it better in a fortnight. The person doing the paperwork needs the tool to work on Tuesday, not to be impressive at I/O.

A practical step: Take one real workflow, run it through the model you use now, and log tokens in and out against the result. That single number, cost per completed task, is your comparison baseline for whatever ships on the 17th.

Three items, and the same discipline threads through them. McLaren's robots are worth watching precisely because a live contractor deployment generates evidence a demo never will; ask for that evidence in six months, not the launch quotes today. The Loughton letters tell you data centre work carries a political layer, so underwrite it like one. And Gemini's delay confirms what we've found all year: the model market rewards buyers who measure cost per completed task and punishes ones who buy the benchmark.

None of this needs a strategy day. It needs a name against the robot's data archive, a consent-risk line in the register, and one logged workflow with a token count. Small, boring, checkable. That's the job.

Worth doing: Pick whichever of those three applies to you and do it this week, before the next launch cycle gives you a fresh excuse not to.