Construction AI: a Manchester student scheme cleared all 29 blocks through Gateway 2 at once, and Scotland got a £16.5bn bid for the world's biggest data centre

Here's the one to put in front of anyone still treating Gateway 2 as the reason a scheme can't move. On 1 October GRAHAM confirmed it had full Gateway 2 building control approval for every one of the 29 blocks at the University of Manchester's Fallowfield campus redevelopment. Seven of those are higher-risk buildings, 18 metres or taller, the ones that sit under the Building Safety Regulator's direct oversight, and the other 22 are lower-rise. The scheme's worth £530m, runs to 3,300 student beds, with the first thousand rooms due to open in 2028 and the whole thing finished by 2030. Clearing the lot in one go is the part that would have sounded far-fetched a year and a half ago (Construction Enquirer and PBSA News, 1 October).

But the approval isn't really the headline. How they got it is. GRAHAM and the client side, a 50-year partnership with the university alongside Equitix and Rothesay, filed the Gateway 2 applications months before planning consent had even landed, and put the blocks through as a set rather than one at a time. Early, and batched. And that lines up exactly with what the regulator keeps saying actually works. Charlie Pugsley, the BSR's acting chief executive, has pointed to batched applications getting decided in 12 to 14 weeks, while the innovation unit handling new higher-risk buildings is running a 22-week median with a 90 per cent pass rate (BSR figures to 30 May). Andy Roe, who chairs the regulator, reckons Gateway 2 is now much closer to its stated 12-week target, down from something like 48 weeks at the worst of it. I was sceptical the gateway could shift at this pace, and I'll say plainly the direction has caught me out.

What that means for whoever's carrying a higher-risk scheme is a change of plan, not a change of mood. The old move was to sit the building on the shelf and blame the regulator. The new move is to get your Gateway 2 pack drafted early, alongside or ahead of planning, and bundle your blocks so they're assessed together. It's a bit like booking the MOT before the tax runs out rather than after. Same job, done in the right order, and you're not left off the road. The person who feels the difference is the project manager who now gets to plan to a 12 to 22 week window with some confidence, instead of a shrug and a prayer.

Today's action: pull your next higher-risk scheme forward and get the Gateway 2 submission started now, before planning, then ask your approved inspector whether the blocks can go in as one batch.

On 30 September a developer called CX Tech lodged a planning application for a data centre campus at Chapelcross, near Annan in Dumfries and Galloway, on the land next to the decommissioned nuclear station. The numbers are the kind you read twice. A gigawatt of capacity, £16.5bn of investment, 162,000 square metres of floorspace across eight buildings, plus a training centre, an industrial element and its own substation, built out over roughly eight years. If it clears planning and gets built, it would be the largest data centre in the world (developer's figures, via Data Centre Review and The Business Desk, 30 September to 1 October). Treat the headline numbers as the applicant's own for now, because that's what they are.

The power is the interesting bit, and it's where the thinking has moved on. CX Tech isn't asking the grid for a gigawatt and hoping. It's pointing to around 1GW of renewable generation within five kilometres of the site and 2GW within twenty, and proposing a mix of grid supply and private-wire connections to wind, solar and biomass. That's the same lesson the Nscale story taught us a fortnight back: on a build this size the power connection is the project, and the schemes that pencil are the ones that bring their own electrons. The cooling is closed-loop and non-evaporative too, which on the developer's own sums puts water use at about four homes' worth a year per building, a direct answer to the drought-and-data-centre row that's been rumbling all summer.

The planning backdrop matters as much as the kit. Scotland now requires a full environmental impact assessment for any data centre above 50MW, so a 1,000MW campus walks straight into the most demanding version of that process, and it's landed while some at Holyrood are openly pushing for a pause on these schemes. Local campaigners and a Greens MSP have already called the scale of it reckless, which tells you the consent fight is going to be as much political as technical. For anyone bidding data centre work in Scotland, the message is plain: the EIA and the power connection shape your programme before a single foundation goes in. The contractor who wins this stuff isn't the keenest on the shell, it's the one who turns up with the grid and the environmental case already half solved.

On the programme: if you're chasing Scottish data centre work, build the EIA and the power connection into your bid timeline from day one, because above 50MW they're the critical path, not the steel.

And the one from outside the sector. On 30 September Google unveiled Gemini 4 Argon, its first big model in nearly a year. Google says it tops OpenAI's GPT-6 Astra and Anthropic's Claude on several coding and knowledge-work benchmarks, and claims the lowest hallucination rate of any frontier model going. Those are Google's own figures, so weigh them accordingly. VentureBeat framed the launch as Google retaking the benchmark lead. The catch, and it's a real one, is that Argon isn't generally available. It's gone to a small group of cyber-defence partners first, with wider access promised once the testing's done (VentureBeat and Axios, 30 September).

Why does a Google model launch sit in a construction brief. Same reason OpenAI's DevDay did last week: the agents that read your drawings, check a spec or chase an RFI run on models like this one, and they reach you inside the tools you already pay for, not as something you go out and buy. A stronger, more accurate model underneath means the checking tools get better and the slips get rarer. But a benchmark win you can't access does nothing for your Tuesday. What a limited release really tells you is which way the frontier is leaning, not what you can put to work. I'm not convinced the leaderboard gap is as wide as Google's slides suggest, but the direction, models getting more accurate and better at long multi-step jobs, is the thing that matters for anyone leaning on AI to check a compliance pack.

So the practical read for the person choosing software is to skip the horse race and ask the boring question: what can I turn on this month, and does it do the job on the scheme in front of me. The model under the bonnet will keep changing, and often you won't even be told when it does. What shouldn't change is your test for the tool itself, whether it does real work and whether you can walk away with your data if it doesn't.

A practical step: next time a vendor leads with a benchmark, ask which model they're actually running today, and whether you'd notice if it were swapped out overnight.

Three stories, one spine. GRAHAM's gateway win is proof that the compliance route works when you plan for it, not something to dread. Chapelcross is proof that on the big AI-driven jobs the power and the planning case are the project, and the contractor who solves them first wins. And Gemini 4 is a reminder that the cleverest model in the world does nothing until you can actually switch it on. What ties them together is evidence you can use rather than announcements you admire: the regulator's own data, the developer's power numbers, the tools you can genuinely deploy. Admiring the leaderboard is not the same as getting a building through a gate.

For your board pack: one line on each. Where your higher-risk schemes sit in the Gateway 2 queue and whether you've filed early, whether any data centre work in the pipeline has its power and EIA solved, and which AI tools you've actually deployed against the ones you've only trialled.