Construction AI: the UK's flagship AI campus can't get a plug, and China ships the biggest open model yet

Here's the week's most instructive story, and it's about a plug. The Telegraph reported in the week of 13 July, picked up by City AM, Capacity and the data centre trade press through to the weekend, that Nscale's £2bn AI campus at Loughton in Essex has been told its 90MW grid connection will not be ready in time 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 to 2027. Now the date is soft again, and the reason has nothing to do with planning, design or the supply chain. The site simply can't draw the power.

Regular readers will know this site's history. It's the one where a minister wrote directly to the council chief executive to press for approval, which I covered on 8 July, and the one the Guardian visited in March and found a working scaffolding yard. So the state has done everything a state can do for this scheme: consent granted over officer objections, prime ministerial billing, national-importance framing. And none of it conjures 90MW out of a congested grid. Nscale's answer, per the reporting, is to talk to Bloom Energy of California about solid oxide fuel cells running on natural gas, which slots this flagship neatly into the pattern I covered on 18 July, where more than 100 UK data centres are planning on-site generation rather than waiting in a connection queue that stretches to about 140 sites. There's a harder edge too: further slippage could expose Nscale to financial penalties if it has committed compute capacity to customers by specified dates. I'm not sure a fuel-cell bridge closes that gap on the timescale the contracts assume, but the direction is set, and the politics are moving as well. In early July 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. New York last week, Holyrood possibly next. The pipeline's constraint list now reads power, people and politics.

What that means for a contractor or consultant on one of these jobs is that the energy centre is becoming the critical path. Fuel cells, gas connections, acoustic and emissions treatment, and a commissioning sequence that has to produce dependable power before a single rack goes live. That's real scope, real programme, and a consenting question that didn't exist when the job was bid. The person this lands on isn't the developer's investor relations team. It's the project director who has to explain why practical completion means nothing without an energised substation.

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.

Two days after I wrote about DeepSeek V4's rush-hour pricing, China shipped again. Moonshot AI released Kimi K3 on 16 July, a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window, native multimodal input and an API priced at $3 per million input tokens and $15 out. The part that matters is the licence plan: Moonshot has dated the full open weights for 27 July, which, if it holds, makes K3 the largest open-weight model ever released. On launch day it went straight to number one on LMArena's Frontend Code Arena, a 17-place jump on its predecessor, and independent trackers place its general intelligence alongside the strongest closed models from Anthropic and OpenAI, though still behind the very top tier on broad measures.

Now, the honest caveats, because there are a few. Arena rankings are crowd-voted preference tests, not proof a model can hold a fire strategy or a bill of quantities together, and the open-weights date is Moonshot's own until the files actually appear. A 2.8-trillion-parameter model is also not something you run on a workstation under the stairs; self-hosting this class of model is a serious infrastructure decision. But the trend line doesn't need this one release to be perfect. What we've now had is two frontier-class open releases from China inside a fortnight, and each one drags down the price of the closed alternatives it competes with. Think of it like the effect a serious second tier-one bidder has on a framework rate. Nobody has to switch for everyone's price to move.

For a UK construction business the practical question hasn't changed since Thursday, it's just got sharper. Open weights are the route to putting a capable model inside your own boundary, next to contract records, commercial data and golden-thread documentation that you'd rather not send to anyone's cloud, Beijing's or California's. The firms that will move fastest on this aren't the ones picking a winner today. They're the ones whose IT lead can already answer what it would take to host one. That's the capability worth building while the models fight it out.

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 we need to buy, and what would it cost? The answer is probably smaller than you think, and it dates quickly, so ask again in the autumn.

Hold today's two stories together and you get the strange economics of this moment. The intelligence is getting cheaper by the week, and the physical capacity to house it is getting scarcer. Turner & Townsend's global construction market survey, published 9 July and worth the read even at ten days old, found data centres are now the most in-demand and capacity-constrained construction sector in the world, with 87% of markets reporting shortages in the mechanical, electrical and plumbing trades and labour availability now the primary driver of cost escalation globally. UK construction inflation is forecast at 3.7% for 2026 and 4.2% for 2027, and London sits fifth on the world's most expensive construction markets at $6,032 per square metre. Buried in the same report is the adoption signal I'd act on: 66% of markets said AI capability has become more important in tendering and client conversations over the last 12 months. Your next PQQ is more likely than not to ask.

So the discipline stands. The tools are improving faster than the grid, the labour market or the consenting system, which means the returns right now go to firms that apply cheap intelligence to their existing work, bid desks, document control, progress records, rather than firms waiting for the infrastructure story to resolve itself. It won't resolve soon. That's what it's about.

For your next bid: Pull your last three PQQs and count the questions that touch AI, data or digital capability. Then check your standard answers were written this year. If they weren't, that's an afternoon's work with a measurable return.