Construction AI: west London says yes to another 64MW shed, and Ofgem starts pricing the queue

Here's the shape of the UK data centre boom in one week. On 6 August 2026 Hounslow Council approved a scheme on the Haslemere Heathrow Estate, a 12.2-acre brownfield site off the M4 currently carrying a row of tired industrial warehouses: three data centre buildings, up to 50,000 sqm, 64MW of electrical demand for 48MW of IT load at a targeted PUE of 1.35. It was brought forward by Brixton (Heathrow Estate) Ltd, a vehicle tied to the California real estate investor Kennedy Wilson, and it needs an 80MVA connection, roughly the draw of tens of thousands of homes, imported from the National Energy System Operator through the Laleham grid supply point that sits outside the borough entirely. Councillors raised the water and energy questions and approved it anyway.

So the consent went through. The connection is the problem. Within days Ofgem, looking squarely at this kind of scheme, put out a consultation on a Data Centre Commitment Fee: somewhere between £237,500 and £712,500 per megawatt, payable when a developer accepts a connection offer, refunded once the project actually connects, and forfeited if it drops out of the queue early. Alongside the money there'd be data-centre-specific milestones, where you keep having to evidence financial capability, commercial maturity and procurement progress to hold your place. The consultation runs to 16 September 2026. It builds on the grid-connection reforms Ofgem floated at the end of July, and it puts a real price on the thing that's actually scarce.

What that does is turn the grid queue from a free option into a paid commitment, which is the whole point. For years a speculative scheme could park a connection request, hold the capacity and decide later, while a job that was genuinely ready to build waited behind it. A refundable deposit in the hundreds of thousands per megawatt sorts the tyre-kickers from the buyers fast. The comparison only goes so far, but it's the ticket touts and the empty seats problem: make people put money down and the queue starts telling the truth. For a contractor or a consultant chasing data centre work, the lesson is that the programme risk has moved. The planning gateway is clearing. The connection date, and now the connection cost, is what decides whether the job is real.

For your board pack: on any data centre pursuit, track the grid connection status and the commitment fee exposure as a live programme risk from day one, not a utilities detail you pick up at RIBA Stage 4.

Reality capture used to mean a 360 camera on a hard hat, walking the floor, comparing what got built against the model. On 6 August 2026 a Nashville startup called SiteVue AI, founded only in August 2025, raised a $7.5m seed co-led by Penny Jar Capital and Overture for something a step past that. Its fixed and wearable cameras run custom-trained models over the footage in real time, and what they're trained to spot isn't just defects and cycle times, it's near-misses, injuries, labour movement and PPE compliance. It installs in days, no infrastructure change, subscription pricing. The company says customers typically see a 3 per cent margin uplift within three months and, in some cases, up to ten times their spend back. Those are vendor figures, and I'd want to see them survive a wet January on a groundworks package before I believed the ten-times.

The safety case is real, and I'm not going to pretend otherwise. A camera that flags a bloke stepping into a swinging load, or a missing harness at height, is doing something a supervisor with eyes on twelve other things cannot. But there's a line here that the sector hasn't drawn yet. The RICS survey that landed this month found the barrier to AI on site that grew fastest in a year wasn't cost, it was trust, with people naming privacy and security concerns jumping from 22 to 30 per cent. A camera that watches the wall is reality capture. A camera on a lanyard that watches the person is surveillance, unless you've said out loud what it records, who sees it and what it's never used for. Get that written down before the kit arrives, because the person it's pointed at is the one whose buy-in you actually need.

Practical bit: if a monitoring camera goes on site, put a one-page notice on what it captures and who can access it into the induction, and keep the footage rules out of the disciplinary process unless you've told people that's where it goes.

Every AI agent the sector's been sold this summer, Procore's Digital Coworkers, Bridgit's workforce planners, the estimating bots, reaches into your data through the same plumbing: the Model Context Protocol. It's the standard that lets an agent read your drawings, your CDE, your programme. This month that plumbing passed 400 million SDK downloads a month, up about fourfold this year (Anthropic's figure), which is another way of saying it's now the default way agents talk to your systems. The enterprise authorisation piece settled back in late July. In August the next brick landed: Claude Security opened in public beta for enterprise customers, and Anthropic retired its experimental prompt-tools APIs on 17 August as it tightened the platform up.

None of that is exciting, and that's rather the point. What we've found is that the question people ask about an agent, is it clever enough, is almost never the one that matters. The one that matters is the golden thread question you already apply to a subcontractor: who can it reach, and can you see what it did afterwards. An agent with a valid token and no audit trail is the digital version of handing a bloke your keys and not writing down which doors he opened. So when a vendor demos an agent that'll draft your daily log or chase your RFIs, park the cleverness for a minute and ask about the authorisation model and the log. If they can show you exactly what it touched and who let it, good. If they can't, that's your answer, and it's got nothing to do with the AI.

The procurement filter: for any agent touching project data, make "show me the access scope and the audit log" a scored question in the tender, not a nice-to-have.

Three items, one thread. The data centre pipeline runs into power and gets a price tag on the queue. The cameras run into people and their willingness to be watched. The agents run into provenance and whether you can prove what they did. In each case the AI or the shed is the easy part, and the constraint is something older: the grid, trust, and accountability. That's what it's about.

None of these needs a strategy deck. They need one decision each. On a data centre bid, whose name is against the grid connection risk. Before a monitoring camera switches on, who signed off what it watches. When an agent goes near your files, where's the log. Sort those three and you're ahead of most of the market, which is still admiring the demo.

A practical step: pick the one of the three that's live on your desk this week, and put a name and a date against it before Friday.