The week the ground moved under the tools

The Telegraph reported in the week of 13 July, picked up through the trade press into the weekend, that Nscale's £2bn AI campus at Loughton in Essex has been told its 90MW grid connection will not be ready for the facility's planned 2027 opening. This is the project billed as the UK's largest AI supercomputer, announced by the Prime Minister during the September 2025 state visit, backed by Nvidia and tied to a roughly £22bn Microsoft package. The opening had already slipped once, from 2026. Nscale's answer, per the reporting, is talks with Bloom Energy of California about solid oxide fuel cells running on natural gas, and there's a harder edge: further slippage could expose the firm to financial penalties if it has committed compute capacity to customers by fixed dates.

The instructive part is what didn't fix it. This is the site where a minister wrote directly to the council chief executive to press for approval. The state granted consent over officer objections and gave the scheme prime ministerial billing, and none of it conjures 90 megawatts out of a congested grid. It slots the flagship into the pattern of more than 100 UK data centres now planning on-site generation rather than waiting in a connection queue stretching to roughly 140 sites, and the politics are moving too: the SNP's National Council backed calls for a Scottish moratorium on AI data centre developments, reported around 9 July, with 24 hyperscale projects proposed north of the border. For the contractor or consultant, the energy centre is becoming the critical path: fuel cells, gas connections, acoustic treatment, and a commissioning sequence that has to produce dependable power before a single rack goes live. That's real scope that wasn't in the bid.

The programme note: if you're delivering or bidding a data hall, put the power date on the risk register with a named owner and a weekly status, and ask the client now whether on-site generation is in scope. If the answer is "under review", price the review.

This one is a fortnight old, flagged transparently, and it still hasn't had the attention it deserves. In a written ministerial statement on 9 July (HCWS209), the government set out changes to how the Building Safety Regulator handles building assessment certificates for occupied higher-risk buildings, and buried in it is the number that matters: since April 2024 the BSR has directed nearly 2,000 buildings to apply, and 66% of those applications have been refused. When a regulator refuses two-thirds of the submissions it specifically asked for, the honest reading is that the standard wasn't legible from the outside. This isn't a story about incompetent applicants.

The response is a more proportionate, risk-based approach, focused on organisations managing multiple higher-risk buildings, with updated resources for small resident-led management companies from September 2026. That last part is the humane bit; a right-to-manage company with three retired directors was never going to produce a safety case to a national managing agent's standard. Separately, the government confirmed funding to extend the Cladding Safety Scheme to buildings under 11 metres with serious life-critical cladding defects, with applications opening in August. The AI angle is that there isn't one, directly, and that's the point. The golden thread problem the BSR keeps hitting is a document problem: evidence scattered across handovers, no traceable chain from design decision to as-built, a safety case that reads as assertion rather than proof. That's exactly the work document intelligence is good at, and exactly the work most organisations still do by hand at eleven at night.

The practical bit: if you hold or manage buildings in scope, pull your last refused or pending BAC and map every claim in the safety case back to a source document. Where you can't, that's your gap, and the revised approach will still find it.

GS1 UK and Barbour ABI published a joint report on 14 July putting the cost of poor construction product information at up to £3.8bn a year. Not poor software or poor process. Poor product data: fragmented, inconsistent, and frequently impossible to tie back to a single product identity across the organisations that specified, bought, installed and now maintain it. The compliance numbers are the ones for a board paper: 98% of professionals say they're aware of the Building Safety Act, 21% say they're fully prepared for it, nearly three years in. Nine in ten know the golden thread as a concept. 14% say they fully understand it. And 84% report products being substituted on site, which is the ordinary reality of construction and also the single hardest thing to trace once the building is occupied.

Iain Walker, GS1 UK's director of industry engagement, put the risk plainly: without fixing product data, the golden thread becomes a repository of documents rather than a trusted record of what was actually specified, installed, substituted and maintained. That sentence is the whole item. It's a bit like a beautifully organised filing cabinet where half the labels were written by someone guessing; the cabinet is fine, and no amount of retrieval software rescues you from the labels. So if you're buying an AI tool to search your building records this year, the question isn't how good the search is. It's whether the records identify the actual installed product, including the substitution the site manager approved on a Tuesday afternoon in 2023 because the specified one was on a twelve-week lead time. Hold this next to the BSR's 66% refusal rate above and you're looking at the same disease from two angles.

Worth doing: pick one completed higher-risk building and try to trace three installed products back to a manufacturer identifier and a current declaration of performance. Time it. That number tells you where you actually stand.

Here's the item that won't make a headline on your site but will show up in a contract you sign this year. On 30 September 2025 Procore rewrote its API terms and published a Developer Policy banning bulk downloads of construction data for training large language models. It then removed Trunk Tools, an AI agent firm used by large US general contractors including Gilbane and Suffolk, from its API, and refunded the booth Trunk Tools had booked at Procore's own conference. On 20 January 2026 Procore completed its acquisition of Datagrid, a rival agentic-AI provider. Protecting customer data with one hand, buying an agent company with the other. Procore frames the API move as security. I'd read it more plainly as a platform defending a moat, and there's nothing wrong with saying so out loud. Trunk Tools answered in June 2026 with Cortex, an intelligence layer built to connect over the systems where project data already sits, including Procore and Autodesk Forma.

Most of the named players are American, and a UK regional contractor isn't in the room for these decisions. But the terms of service are the same terms of service, and the lesson travels without a passport. The same project data that runs your job is the fuel that trains AI agents, so platforms are deciding whether your data can train their agents and which outside agents are allowed to touch it. When you buy a construction platform now, the questions that decide your future flexibility are who can train on your project data, whether an agent you choose can get API access to it, and whether you can export your own records in a usable form if you leave. Those aren't IT questions. They're the difference between owning your golden thread and renting it back.

The procurement filter: on your next software renewal, get three answers in writing before you sign: can the vendor train on our project data, can an AI agent we choose reach it through the API, and can we export it in full on exit. If the answers are vague, that's your answer.

Monumental announced a $32m Series B on 15 July, led by Khosla Ventures with Plural and Hummingbird returning. The Amsterdam firm has more than 150 bricklaying robots in the field, with brickwork delivered on over 100 homes plus a school, a sports centre, a hotel and hundreds of metres of Amsterdam canal wall, and nearly half of those homes were completed in the three months before the raise. Chief executive Salar al Khafaji, who sold his previous company to Palantir, calls the 2024 round "basically still an R&D round" and this one a proof round, which is honest of him. The UK is the primary commercial focus for the next six months, with delivery through VINCI business units including Taylor Woodrow.

But the reason this matters for a UK contractor isn't the hardware. It's the commercial model. Monumental doesn't sell or lease robots. It bids brickwork like a specialist trade, per thousand bricks at rates close to local masonry pricing, and gets paid for finished walls; if the robots can't complete a scope, human crews finish it at Monumental's cost. The customer never holds the technology risk, which is precisely the trap the last generation fell into when FBR's Hadrian X asked contractors to commit nearly $6m per unit for an unproven machine. The labour arithmetic does the selling: the Home Builders Federation reckons the UK needs at least 20,000 more bricklayers for the 1.5 million homes target, against roughly 1,990 completed bricklaying apprenticeships in 2024. I'm not sure the economics survive a wet February on a tight urban site with three trades stacked behind you, but the delivery record is real and the risk allocation is genuinely sensible.

Today's action: if a robotic trade lands on your tender list this year, ask three things before you look at the rate: who carries the technology risk, what happens to the programme if autonomy drops, and who is legally the contractor for that scope.

On Tuesday 28 July the Model Context Protocol publishes its final 2026-07-28 specification, the biggest revision since the protocol launched in November 2024. Most people in construction have never heard of MCP and don't need to, but you're almost certainly using it: it's the standard that lets an AI model reach into a tool and do something, and it's how Bluebeam ships Revu with Claude connected, with Procore, Autodesk and the rest wiring agent access the same way. The headline change is that MCP goes stateless, so any request can hit any server instance, and the authorisation hardening tightens who an agent is allowed to be when it talks to your systems. Beta SDKs for Python, TypeScript, Go and C# have been out since 29 June, and the maintainers are explicit that nothing breaks on the day.

I believe them. I'd still want it tested rather than trusted, because "nothing breaks" is a claim you want to have verified in a test branch rather than discovered on a live job. And the authorisation changes are the ones your IT people should actually read: if agents have been getting into your CDE on vaguely scoped credentials, the direction of travel is that vagueness stops being tolerated. That's a good thing. It's also work.

The discipline: if anyone in your organisation runs an MCP server, or your CDE vendor does, ask this week whether they've run the beta against real traffic and pinned their versions. If the answer is "our vendor handles that", get it in writing with a date on it.

Moonshot AI released Kimi K3 on 16 July, a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window, priced at $3 per million input tokens, with full open weights dated for 27 July. If the weights land, it's the largest open model ever shipped, and the second frontier-class open release from China inside a fortnight after DeepSeek V4. Arena rankings are crowd-voted preference tests, not proof a model can hold a bill of quantities together, and the weights date is Moonshot's own until the files appear. But each release like this drags down the price of the closed alternatives, the way a serious second tier-one bidder moves a framework rate. Nobody has to switch for everyone's price to move.

Worth asking this week: put one question to whoever runs your IT: if we needed a capable open-weight model running privately against our project data within six months, what would it take and what would it cost? The answer dates quickly, so ask again in the autumn.

On 21 July Google released Gemini 3.6 Flash, 3.5 Flash-Lite and a security-tuned 3.5 Flash Cyber. The pitch isn't a smarter model, it's a cheaper one: about 17% fewer output tokens than 3.5 Flash and $7.50 per million output tokens, down from $9, all Google's own figures. That matters more than a flagship for construction, because the high-volume work, drawing takeoff, RFI drafting, document sorting, runs on Flash-class models and bills by the token. The other half of the story, an update on what I covered last week, is what didn't ship: Gemini 3.5 Pro missed its widely reported 17 July date, its third slip since May, after Bloomberg's 16 July report that DeepMind scrapped and rebuilt the base model over hallucination rates. Google teased Gemini 4 instead.

The takeaway: if a supplier's answer to a capability gap is a model that isn't generally available yet, treat it as absent and price the deal on what runs today. And when a cheaper Flash-class model lands, make sure someone is allowed to switch the pipeline over and test it.

One figure from Turner & Townsend's 9 July global construction market survey deserves its own entry, distinct from the capacity-squeeze headlines covered a fortnight ago: 66% of markets surveyed said AI capability has become slightly or much more important in tendering and client discussions over the past twelve months. Not in innovation workshops. In tender conversations, where the money is. The backdrop makes it bite: 87% of markets report MEP trade shortages, London is now the world's fifth most expensive construction market at $6,032 per square metre, and UK construction inflation is forecast at 3.7% this year and 4.2% in 2027. When a client asks what AI you use, the answer being scored isn't a tool list. It's whether you can say what it does, who checks it, and what happens when it's wrong.

Today's action: write down, on one page, the three AI tools your bid team actually uses, the named person accountable for each, and the check before anything goes out. If you can't fill the page in twenty minutes, your next PQQ will find the same hole.

Two smaller signals on the pipeline. The SNP's National Council backed calls for a Scottish Government moratorium on AI data centre developments, reported around 9 July, with 24 hyperscale projects proposed north of the border; the pipeline's constraint list now reads power, people and politics. And mid-July reporting suggests 30 to 50% of the global data centre capacity planned for 2026 could slip on power and equipment, so the sector proving so many of these tools is also the one most likely to teach you how a fixed date becomes a moving one.

The procurement filter: if a single sector is more than a fifth of your forward book, treat it like client concentration risk, and check where each scheme's power and funding actually sit before committing long-lead resource.