Drax, Teesside and the power bill behind Britain's AI datacentre plans

On 3 November the FTSE 250 owner of the Drax power station in North Yorkshire announced a £20bn deal with the infrastructure investor Brookfield and the AI infrastructure platform Radiant, according to the Guardian. The plan puts a datacentre on the site of the plant the newspaper calls Britain's most polluting. Its eventual capacity is more than 1 gigawatt. Phase one, up to 100MW of compute, is due to be operational from the first quarter of 2028.

That is a power story wearing a datacentre coat. The Guardian reports Drax will start by supplying land and access to the national grid through existing connections on the site. It hopes to begin powering the facility directly by the early 2030s. Its chief executive, Will Gardiner, said the project would "help secure the long-term future of the site and protect hundreds of jobs" while "supporting the north's reindustrialisation and strengthening sovereign AI capability". The arithmetic behind the enthusiasm is not complicated. Drax burns woody biomass pellets and produces almost 5% of the country's electricity, according to analysts quoted by the Guardian. That makes it the single largest emitter of carbon dioxide. Those emissions are not counted against the UK government because Drax claims the equivalent carbon is absorbed by growing trees. Ministers have promised a "much more limited role" for the plant after 2027 to 2031, when it will have access to only half its current £2m a day in subsidies.

So the datacentre is a hedge as much as an expansion.

The regulatory question is what happens if direct supply does not clear approvals before 2031, when the subsidy cut lands. The company's market value climbed almost 9% on the London Stock Exchange to almost £3bn on the news, the Guardian reported. Equity analysts welcomed it.

Teesside, and a grid queue that decides who gets to build

The Drax announcement came two days after National Grid moved on a second, larger power question. On 1 November the Guardian reported that the company had agreed to buy 29.5 hectares (73 acres) of land at Dorman Point on Teesworks, a local private-public partnership site. The plan is for what would be Britain's biggest electricity substation. The deal was first reported by Sky News, the Guardian said. It follows reports last year that the government had picked Teesworks for a new AI datacentre, with planning applications submitted for a project spanning nearly 500,000 sq metres.

Capacity is the headline number. Ben Houchen, mayor of the Tees Valley combined authority, told Sky News that having up to 3.2GW at the site means the region can compete for projects. "Everyone who has tried to build a datacentre in this country knows the biggest problem is getting enough power to the site quickly," he said. He added that while Britain has been stuck in the queue, the US, the Gulf and parts of Europe have been "hoovering up the investment".

Matt Pearce, head of land at National Grid, said the deal "reflects the constructive work between all those involved to enable this vital infrastructure". He said the company is committed to progressing the development as quickly as possible. The Guardian noted Balfour Beatty was reported to be in line to deliver the substation, with Steel River Power also involved. None of that answers the harder question: whether the power arrives before the public patience runs out.

Both announcements landed within a week of the government publishing a draft national policy statement on datacentres. It described them as "as fundamental to our economy as the railways and electricity networks on which we have long depended". The same document said ministers must "take control over the development of datacentres in this country" so that host communities benefit. That is a striking thing to write in the same fortnight that a biomass plant with a contested emissions account is being repositioned as AI infrastructure.

The models are getting cheaper, the buildings are not

There is a version of this story where the demand forecast softens. On 3 November, the Register reported on a Gartner document, "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps". It argues that most agent work will not run on frontier models at all. Gartner's Mark Petty and Raymond Paquet wrote that "nine in 10 agent workflow steps are routine, and open-weight models handle them at parity for a quarter of the price of the cheapest closed model".

The pair put the best case at a twentieth of the cost of running a frontier model on every step. They concede that "current open models can be more verbose and retry more often, so the saving per task against a frontier model is closer to five or six times". Their planning assumption is that by 2030, 90% of enterprise operational agent workflow steps will execute on open-weight models.

Read that next to Google's week. Tom's Hardware reported on 3 November that Gemini 4 Carbon, released the previous week, scores 59 on Artificial Analysis' Intelligence Index V5. That is one point ahead of Claude Opus 5.5 and two ahead of OpenAI's GPT-6 Astra, at $7.76 per task on that test. It also tops the firm's AA-Omniscience v1.1 hallucination and recall benchmark, though the best score on a -100 to +100 scale was 54. Google is giving cyber defenders first access through its Fairwind program before wider availability.

Cheaper, better models per task are exactly the argument for building fewer, or at least differently shaped, datacentres. The counterargument is that total demand can rise even as unit cost falls. It is also that the grid connection, not the GPU, is the binding constraint. Neither Drax nor National Grid is publishing a model-by-model forecast of what the site will host.

The environmental bill is not only carbon

Water sits awkwardly in this story. None of the primary documents on Drax or Teesside sets out cooling design or water withdrawal figures. The Guardian's reporting does not mention water at all. The wider debate is loud, with headlines in the last 72 hours about cooling, water stress and local moratoria. But those are context, not numbers we can attach to these two projects.

What the record does support is the political risk. The Guardian reported that in October AI companies told ministers they had 12 months to head off public opposition to datacentres or risk US-style protests. It reported that campaign groups have protested against planned developments from Auchtertool, north of Edinburgh, to Brick Lane in east London. The Teesside project has a mayor arguing for speed and a planning system that has to be seen to consult.

There is also a quieter technical thread. A paper submitted to arXiv on 29 October by Wendi Li and co-authors proposes agentic world modeling for test-time scaling without replicating the environment. It reports gains of 40.07% on average over a single-shot simulation baseline across tau-bench, AppWorld and OfficeBench. That is research, not procurement. But the direction of travel is the same one Gartner describes: agents that plan against read-only queries instead of parallel sandboxes, with fewer expensive executions and more cheap ones.

Separately, a 29 October arXiv paper by Zhifei Hu, Alexandra I. Cristea, Adel Bibi and Jialin Yu looked at 3,000 predictions from ten frozen open models judged by humans. It found that 86.1% were semantically plausible, yet plausibility did not translate into safe authorisation decisions. The lesson generalises past web agents. A forecast that looks right is not the same as a forecast you can build a 1GW facility on.

What to watch before 2028

Three dates matter. The first quarter of 2028, when up to 100MW at Drax is meant to be live. The 2027 to 2031 subsidy window, when Drax's support halves. And whatever the government's final policy statement says about who pays for grid reinforcement, since neither announcement puts a figure on that.

For now the money is moving faster than the electrons. Drax has land, a grid connection and a partner with Brookfield's balance sheet. Teesside has 73 acres, a substation plan and a mayor who has been waiting in a queue. Neither has published the water numbers, the cooling design or a supply contract that survives 2031. Those are the documents to ask for.