Construction AI: the slump eased in September, but it's data centres pulling and housing still dragging

So the number everyone will quote this week is 46.1. That's the S&P Global UK construction PMI for September, out on 6 October, up from 44.3 in August and a fair bit better than the 44.9 the City was expecting. It's the slowest the sector has shrunk in eight months. Read that twice, though, because the word doing the work in that sentence is "shrunk". Anything under 50 on this index means activity is still falling. September was a smaller fall, not a rise. Less bad is not the same as good, and I'd keep that straight before it goes anywhere near a board pack.

What's underneath the headline is the bit worth your time. The three big sub-sectors aren't moving together. Civil engineering is where the strength is, and Carly Thorpe at Walker Morris put the modest improvement down to stronger civils work. Commercial building had its smallest drop since May 2025. And housing, again, is the one sat on the floor, with high borrowing costs and a flat market still holding output down. So the recovery, such as it is, is lopsided. It's the infrastructure and data-centre end lifting the average while the homes nobody can afford to build stay still. Tim Moore at S&P Global reckons all three are at least stabilising, which is the glass-half-full reading. New orders tell the other half: still subdued, clients sitting on decisions about the big jobs.

Here's the one an editor might cut. Firms are still cutting jobs, and have been every single month since January 2025. That's not a wobble, that's going on two years of the industry shedding people while the commentary talks recovery. Hold the PMI next to the permissions figure I wrote about on 7 October, the 14-year low in consents, and the shape is consistent: a bit more work going through now, less in the pipe behind it, and headcount coming down to match. For the commercial lead and the resourcing planner, that's the signal, and it isn't the headline. For your board pack: lead with the sub-50, not the eight-month line, and show the sub-sector split so nobody mistakes a data-centre recovery for a housing one. That's the job.

The frontier news this week came out of Paris, not San Francisco. On 6 October Mistral released a public preview of Mistral Large 4, which it has nicknamed "Le Chonk", and the headline specs are big: a trillion parameters with 49 billion live at any one time, natively multimodal, a context window up to a million tokens, and fluent in more than 160 languages including every official EU one. It was trained from scratch on 3,800 of Nvidia's Grace Blackwell chips sitting in Mistral's own European data centres. The weights are due to go open at the end of the month, Reuters says the 27th.

Now the benchmarks, and hold these at arm's length, because most of them are Mistral marking its own homework. It claims a combined coding-agent score of 49.8 per cent, ahead of the latest from DeepSeek and Qwen, and 82 per cent on a test of reproducing software vulnerabilities, where it says the top models from Anthropic and OpenAI scored near zero because they flatly refused the task. The independent read is more sober. Artificial Analysis puts Mistral Large 4 at an intelligence index of 38.4, a big jump on the last version but still only eighth among open-weight models, behind seven Chinese systems, and a long way back from the closed frontier. So it's a real step up and not a leaderboard topper. Both of those are true at once, and I'd be wary of anyone quoting only the half that suits them.

Why does a construction brief care about a French model release. Because the agents that read a drawing, check a spec or chase an RFI all run on models like this one, and they reach you inside tools you already pay for. What Mistral is really selling here isn't the benchmark, it's the location. A capable open-weight model, run under European law, in European data centres, is a direct pitch to any buyer who's been told their project data can't leave the UK or the EU. It's the same question that decided the Procore data-residency move I wrote about on 8 October, where your record physically sits. And it rhymes with the plain promise the challengers make, the broad AI-first platforms taking the big incumbents on with published pricing and your data free to walk out the door: own where your data lives, and keep it portable. Practical bit: if your firm is starting to lean on AI inside its project tools, add one line to the spec, which model and whose data centre, and make a European option something you can actually choose rather than discover after the fact. The person who'll thank you is whoever has to fill in the client's security questionnaire.

Put the week together and a single thread runs through it. The PMI's small lift is civils and infrastructure, a lot of that the data-centre build-out, while housing stays flat. The one corner of the market that's genuinely busy is also the most complicated to deliver and the most crowded with specialist subcontractors. And a piece in Planning, Building and Construction Today on 7 October, which I'll flag is contributed by a payment-software firm so treat it as a vendor making its own case, lands on the right nerve anyway: a data-centre job can run hundreds of subbies and throw off hundreds of payment applications a month, and when that's held together with spreadsheets and email the thing that breaks first is people getting paid on time.

That's the bit that matters more than any model. On the one job that's busy, the specialist electrical and mechanical firms you need are in demand and can pick their main contractor. What keeps them is dull and human: accurate valuations, approvals they can see coming, a payment cycle they can predict. So the AI that earns its keep on a data-centre job this autumn isn't the clever design tool, it's whatever shortens the gap between work done and money in the subbie's account. The comparison only goes so far, but it's a bit like a good crew: they'll come back for the firm that pays straight, long before the one with the fancier kit.

A practical step: before you add another tool to a data-centre job, ask what it does to your month-end payment run, and keep your own payment records in a form you can take with you if you change system. The person doing the paperwork feels that one on the 30th of the month. That's what it's about.