Green party leader Zack Polanski called last week for construction of large AI data centres to stop, and the government rejected it within hours on the grounds that a pause “would be a disaster for jobs and national security”, according to Robert Booth’s reporting in the Guardian on 28 August. Both sides argued from the same two numbers: that data centres take about 2.5% of Britain’s electricity and will take four times as much by 2030, and that the power being requested by data centre projects queuing for a grid connection has tripled inside a year. Neither figure is as firm as the argument built on top of it, and one of them has been published at two materially different sizes in the past five weeks.
Strategic Reality: A moratorium, an industrial strategy and a national planning direction are all being drafted against demand estimates that nobody has yet had to defend under scrutiny. The estimates may well prove right. The point is that no British institution currently owns them.
Why this stopped being a planning question
Until recently, data centre disputes in Britain ran through planning committees one site at a time. Residents in Newport and Cardiff contested particular applications; objectors in Fife, Buckinghamshire and Devon did the same. The arguments were local, the evidence was site-specific, and the numbers that mattered were about noise, traffic and water connections.
That has changed in about six weeks. The Greens now want construction paused nationally, exempting only local-scale sites serving their own communities. Labour has given the sector critical national infrastructure status, which Polanski calls “outrageous” on the grounds that “to most people critical infrastructure means hospitals, water, electricity. It doesn’t mean artificial intelligence.” Scottish ministers have issued a planning direction requiring councils to notify them within seven days of validating any application above 50MW. And the STUC has entered from a third direction entirely, asking not for a ban but for enforceable conditions on jobs and supply chains.
Four positions, four different theories of what is wrong. All four assume the demand is real, large, and arriving on a known schedule.
Where the electricity figure actually comes from
The Guardian’s article sources the 2.5%-rising-fourfold prediction to a page on the Parliament website. Follow it and you reach a press notice dated 16 April 2026 announcing that the Science, Innovation and Technology Committee is launching a short inquiry into low-energy computing. The figure appears in a comment from the committee’s chair, Dame Chi Onwurah: “Data centres currently use about 2.5% of the UK’s electricity, but their consumption is expected to rise four-fold by 2030 as the sector continues to expand.”
That is a chair framing a question, not a committee reporting a finding. The inquiry it launches had a submissions deadline of 14 May and exists precisely because the energy trajectory is uncertain enough to warrant parliamentary attention. Quoting its opening framing as the settled basis for a national moratorium argument inverts what the document is.
Critical Context: The 2.5% figure is not disputed and is not the problem. The fourfold multiplier is the load-bearing part, and it entered general circulation as a scene-setting sentence in an inquiry launch rather than as a published projection with a stated method, baseline and confidence range.
The grid figure has been published at two sizes
The second number is harder. Ofgem’s grid connection data is the closest thing Britain has to a real measurement of intent, because applying for a connection costs something and reveals what developers actually plan to build. It has appeared in three forms this summer:
| Source and date | Figure | What it counts |
|---|---|---|
| Ofgem, reported by the Guardian, 29 July 2026 | 315 data centres queueing, representing 73GW | Data centre projects specifically |
| Reuters, reported in our own coverage, 29 July 2026 | Total connection demand rose 41GW to 125GW (Nov 2024 – Jun 2025), data centres at least 80GW of it | All applicants, with a data centre subset |
| Ofgem, reported by the BBC, 28 August 2026 | Data centre grid requests up from 41GW to 125GW over twelve months | Reported as data centre projects |
The 73GW and the 125GW are not the same measure. One counts data centres in the connection queue; the other, on the Reuters reading, counts the whole queue and then identifies data centres as a subset of it. Read across the coverage, Britain’s most-quoted data centre demand statistic is either 73GW or 125GW depending on which paper carried it, a spread of roughly 1.7 times. Our own reporting on the Ofgem consultation also recorded the regulator’s February assessment at 140 proposed data centres seeking 50GW, which is a third figure again.
Ofgem itself is not confused about this. The distinction between total queue and data centre subset is clear in the underlying material. The confusion is downstream, in the compression that happens when a regulator’s figure becomes a headline and then a debating point.
⚠️ Warning: The 125GW figure is now doing political work in Scotland, Wales and Westminster. Any organisation citing it in a planning submission, an investment paper or a board briefing should state which measure it is using, because a well-prepared opponent can produce a smaller official number from the same regulator.
What the queue actually measures
Even taken at its strongest, the connection queue is a poor proxy for what will be built. Ofgem’s own consultation this summer proposed forcing developers to lodge refundable deposits of roughly £350m to reserve 500MW, explicitly so that scarce capacity goes to “projects most likely to be built, not projects preserving options”. A regulator does not design that mechanism unless it believes a meaningful share of the queue is optionality rather than commitment.
The scale of the mismatch is the tell. The 73GW in the queue sits about 30GW above Britain’s entire peak electricity demand of 45GW. No serious reading of that number treats it as a construction forecast. It is a register of positions taken.
The demand curve itself is contested
Separately from how much capacity is queued, there is an open question about how much will be needed. As we reported on Sunday, a contrarian case is forming in the trade press that AI’s centre of gravity is shifting from training, which concentrates in a few large clusters, to inference, which is lighter and increasingly runs on local hardware. AMD’s acquisition of the Toronto startup Taalas in August 2026, whose chips bake a model permanently into silicon, points the same way. Vendors making that case are selling the alternative and their figures deserve the usual scepticism, but the direction is not implausible.
Set that beside the largest UK AI infrastructure contract of the year. Nscale, a British company, agreed in late August to rent Anthropic roughly 460MW of capacity over six years in a deal reported at about £33bn. The site is in West Virginia. Whatever else that transaction demonstrates, it does not demonstrate that British demand is being met by British capacity, which is the premise of the sovereignty argument the government used to reject a pause.
Strategic Insight: Two things can be true at once. AI compute demand can be growing fast, and the specific British forecasts driving British planning decisions can be soft. Confusing the first with the second is how an industry ends up building against a number rather than against a customer.
Who carries the risk of a wrong number
The interesting question is not whether the forecasts are right. It is who absorbs the cost if they are not, and the answer differs sharply by stakeholder.
| Stakeholder | If demand is overstated | If demand is understated | What to watch |
|---|---|---|---|
| Developers and their lenders | Stranded capacity, forfeited connection deposits, impaired assets | Missed positions, capacity secured at a premium later | Whether deposit forfeitures start appearing in results |
| Host communities | Grid and water capacity consumed by sites that underuse it | Faster expansion than local infrastructure can absorb | Conditions attached to consent, not the headline pledge |
| UK businesses buying compute | Capacity glut, softer pricing, better terms in 2028 onward | Sustained scarcity pricing and long procurement lead times | Whether frontier pricing diverges from self-hosted cost |
| Government | A sovereignty case built on capacity that arrives late or elsewhere | Dependence on foreign compute for health and defence workloads | Where contracted capacity is physically located |
Note the asymmetry in the third row. For most British organisations, an overstated demand forecast is not a threat. It is the more favourable of the two outcomes. The organisations genuinely exposed to the forecast being wrong are the ones whose capital is already committed against it, and the lenders who are now pricing local opposition into credit decisions have worked that out.
The consent constraint is firmer than the compute constraint
There is one number in this story that is not soft, and it is the one about people. Objections have moved from individual sites to coordinated national demands, a pattern we set out in our analysis of the AI resistance catalogue. Scottish campaigners at more than twenty proposed sites stopped fighting separately and started drafting national rules. Holyrood’s lobbying register recorded 31 logged encounters between the AI infrastructure industry and Scottish parliamentarians across roughly twenty months. Friends of the Earth wants a UK-wide pause until there are limits on what the sector may consume.
Compute scarcity is a market condition, and markets clear. Consent scarcity is a political condition, and it does not. An organisation planning AI capacity in Britain over the next three years should treat the second as the binding constraint and the first as a price.
Success Factor: The firms that will move fastest are not the ones with the best demand forecast. They are the ones whose plans survive their forecast being wrong in either direction, because they bought optionality rather than a position.
What to do with a number you cannot verify
The practical response is not to pick a side in the forecast argument. It is to stop letting a single unverified figure carry structural weight in your planning.
💡 Implementation Framework: Forecast-independent capacity planning
Phase 1: Trace your numbers (2 weeks)
- List every demand or capacity figure your AI plan depends on
- For each, identify the primary document, its date, and its stated method
- Mark any figure whose primary source is a press notice, a launch statement or a secondary report
Phase 2: Test the plan against both directions (4 weeks)
- Model the plan assuming capacity arrives 40% cheaper than budgeted
- Model it assuming your preferred region issues a moratorium
- Identify which commitments are irreversible under either outcome
Phase 3: Buy optionality, not position (ongoing)
- Prefer contract structures with exit points over long fixed commitments
- Separate the decision to adopt AI from the decision to site capacity
- Review quarterly against Ofgem queue data rather than press coverage of it
Priority actions by exposure
For organisations buying AI as a service
- Treat capacity pricing as volatile in both directions: A capacity glut in 2028 is at least as plausible as sustained scarcity. Multi-year prepaid commitments made on scarcity assumptions carry real downside.
- Route workloads by cost and sensitivity now: Every workload that runs acceptably on a smaller or self-hosted model reduces exposure to whichever forecast turns out wrong.
- Ask suppliers where the capacity physically sits: A British supplier is not the same as British capacity, as the Nscale contract makes plain.
For organisations building or hosting capacity
- Assume your headline figure will be checked: Any demand number in your planning submission that traces to a press notice is a liability once an organised objector group reads it.
- Price the consent timeline explicitly: Model delay as a cost with a duration, not as a probability of refusal.
- Attach mechanisms to your community commitments: The government has said clean power, responsible water use and British supply chains are expectations. The STUC’s argument is that expectations without enforcement produce nothing.
For public bodies and their suppliers
- Separate the sovereignty claim from the siting claim: Sensitive workloads need capacity in a specific jurisdiction, which is a narrower and more tractable requirement than total national capacity.
- Commission a defensible baseline: The absence of an owned, methodologically stated UK demand projection is the root problem, and it is fixable.
- Publish the scoping alongside the figure: Most of the current confusion is a scoping problem, not an arithmetic one.
Resource Reality: Tracing your plan’s figures to primary sources takes a competent analyst about two days for a typical mid-sized programme. It is the cheapest risk reduction available in this area, and almost nobody does it.
Four things this makes harder than it looks
The number you cannot replace
Finding that a figure is soft does not give you a better one. There is currently no British demand projection with a published method that you could substitute, which means the honest position is a range rather than a point, and ranges are difficult to put in a board paper or a planning submission.
Mitigation: State the range and the reason for it, then plan against the decision points rather than the forecast. A board can act on “we do not commit capital until the connection offer is confirmed” without knowing whether demand quadruples.
Both political sides need the number to be firm
The moratorium case and the growth case both depend on the demand being large and certain. Nobody in the current argument has an incentive to say the projection is weak, which is why it has survived this long without being examined.
Mitigation: Do not expect the correction to come from the debate. Read the regulator’s own consultation documents, which are more careful than the coverage of them, and treat trade-press figures as leads rather than evidence.
Water is a smaller argument than it appears
The government says data centre water use in England is under 1% of non-household consumption, and Water UK’s estimate of roughly 6.6 million litres a day is equivalent to about 20,000 homes. Those are real numbers and they are modest at national scale. But water disputes are local, and a national percentage answers nothing about a specific catchment during a drought.
Mitigation: Never rebut a local water objection with a national statistic. It reads as evasion, and it is the argument most likely to harden a contested application into a refusal.
Self-generation quietly moves the goalposts
Foxglove has flagged that developers facing long connection queues are proposing their own gas-fired generation, which its co-executive director Rosa Curling’s team treats as a climate step backwards. If self-generation becomes the standard workaround, the grid queue stops being a useful measure of anything, because capacity gets built outside the number everyone is arguing about.
Mitigation: Track planning applications for on-site generation separately. They are the leading indicator that the connection queue has ceased to describe the market.
Reality Check: None of this resolves within the current political cycle. Expect the figures to stay contested, the moratorium arguments to continue, and the first genuinely authoritative UK projection to arrive after several irreversible siting decisions have already been made.
What this is actually about
Britain has arrived at a national argument about AI infrastructure without an agreed measurement of what it is arguing over. That is not unusual for a fast-moving sector, and it is not a scandal. It becomes a problem at the moment the argument starts producing binding decisions, which is where it now is: designations granted, planning directions issued, moratoria proposed, industrial strategies demanded.
Three things determine whether an organisation navigates this well.
- Source discipline: Knowing which of your numbers came from a primary document with a stated method, and which arrived via a headline. This distinction is invisible until someone contests it, and then it decides the outcome.
- Directional symmetry: Building a plan that works whether capacity turns out scarce or abundant. Most current AI capacity planning is one-directional, which is a bet, not a plan.
- Consent literacy: Understanding that the constraint on British AI infrastructure over the next three years is permission, not silicon, and that permission has a timeline you can influence but not purchase.
Strategic Insight: The useful question is not “will data centre demand quadruple by 2030”. It is “which decisions am I making now that only work if it does”. For most British organisations, the honest answer is very few, and that is a more comfortable position than the current debate suggests.
Your next steps
Immediate actions (this week):
- Identify every demand or capacity figure in your AI plan and trace each to a primary document
- Flag any figure whose primary source is a press notice, launch statement or secondary report
- Confirm where any contracted AI capacity you rely on is physically located
Strategic priorities (this quarter):
- Model your capacity plan against both a 2028 glut and a regional moratorium
- Move at least one workload from frontier pricing to a smaller or self-hosted model
- Replace point forecasts with stated ranges in board and planning documents
Long-term considerations (this year):
- Track on-site generation applications as a separate signal from grid queue data
- Review whether any irreversible commitment depends on a figure you cannot verify
- Build consent timelines into capital planning as a cost with a duration
Source: Labour rejects Zack Polanski’s call to ‘slam brakes’ on building AI datacentres by Robert Booth, UK technology editor (The Guardian, 28 August 2026). The electricity projection is traced to the Science, Innovation and Technology Committee’s low-energy computing inquiry launch of 16 April 2026. Grid connection figures are drawn from Ofgem as reported by the Guardian and the BBC, and from Reuters via prior Resultsense reporting.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will keep watching for the first UK data centre demand projection published with a stated method, a baseline and a confidence range, because the planning decisions being made now deserve one. Read more analysis at Insights, or get in touch.