The Announcement Was the Deliverable

The Lanarkshire AI growth zone, and the gap between the purpose a system advertises and the one it performs.

In January 2026 the UK government announced an £8.2bn AI datacentre complex in Lanarkshire, built by the American firm CoreWeave and the Scottish company DataVita, and promised it would be powered from on-site renewables and running by 2030.¹ Six months later an investigation by freelance journalist Aisha Down established through freedom of information requests that the site had no grid connection, that its renewable energy parks existed mostly as a website, and that within weeks of the announcement Scotland’s First Minister had been privately conceding to the developer the power problem the announcement waved away in public.² DataVita’s own materials described energy parks “directly connected” to the datacentres, written in the present tense. Nothing was built.³ To supply the gigawatt it advertised, the site would need somewhere between forty and a hundred square kilometres of generation, on the investigation’s own range. It has applied for roughly two, and its parent’s one posted wind scheme, up to nineteen turbines, would cover a small fraction of it.⁴

The easy reading is bad faith, and it is the wrong one to reach for first. The grid is genuinely constrained, and the levers that might ease it sit at Westminster, not Holyrood. The political timetable is real and not of the officials’ making. Compute is mobile capital that will build in Virginia if Scotland hesitates. Against all that, wanting the announcement out, and out fast, is not cynicism; it is close to the rational move for someone who means it. The question is not whether anyone lied. It is what the machinery produces when everyone means well and only the announcement is rewarded.

The natural reading is that the planning system failed, and there is a strong case for it. Scotland’s National Planning Framework rests on an evidence base written in 2022, before ChatGPT existed, which never modelled the electricity demand of an AI datacentre.⁵ The consent DataVita holds for its Chapelhall site is a Permission in Principle, an instrument that grants the yes and defers the binding detail by design.⁶ Environmental impact assessments are mostly screened out: of the seven datacentre applications to reach this stage in Scotland, a countryside charity found five were told none was required.⁷ In Edinburgh this year a datacentre pitched itself as “green” to the council while its own plans showed rows of diesel backup generators, by that same charity’s count around two hundred of them. The planning committee could find no definition of a “green datacentre” in national policy to test the claim against, and refused the scheme only after its own planners had recommended approval.⁸ The word carries no statutory definition here, so it means whatever the applicant needs it to mean. On every count, the gate that should turn a promise into an enforceable obligation is standing open.

Everyone says the system is broken, which is the surest sign the words have stopped meaning one thing. “Not fit for purpose” holds three complaints whose cures run in opposite directions. To Westminster it is too slow: Rachel Reeves promises to back “the builders, not the blockers”, and the UK’s Planning and Infrastructure Bill takes powers to overrule councils and cut the time schemes spend in court; the cure is to strip the friction out, which is what a growth zone’s “planning support” is for.⁹ Holyrood sings the same tune a tone quieter, Kate Forbes pressing to shorten the wait for planning and consenting and declaring Scotland “open for business”, Swinney’s own letter to DataVita wanting the development to “proceed at pace”. To the planners’ own institute it is starved: planning is the most-cut service in local government, its workforce at a five-year low, carrying forty-nine duties nobody funded, and the cure is money.¹⁰ A third complaint is quieter but well evidenced: the system is neither slow nor starved but toothless. Enforcing a planning condition is a discretionary act, and the obligation a developer accepts to win consent is not reliably delivered: in England, 17,432 affordable homes secured through section 106 agreements held detailed permission but remained uncontracted in October 2025, agreed and unbuilt.¹¹

Of the three, only the first is winning, and Lanarkshire is where it won. Nothing here was slow, because there was next to nothing to decide. One small consent came through on time; the flagship was announced with no application on file. The system’s quickest work was to say yes, and it asked for nothing back.

So calling it a failure smuggles in an assumption: that we agree what the system was for. Watch it against a different purpose. The press release counted 3,400 jobs and £8.2bn of investment, both figures, in the government’s own words, “supplied by DataVita as part of their proposals.”¹² The community fund headlined at £543m holds, as of this summer, no money; it is to be filled from the company’s revenues, if it generates them.¹³ The jobs figure, a charity established under FOI, was lifted from estimates for a different site in Northumberland and multiplied up.¹⁴ The designation came stamped with its own escape clause, “conditional on delivery against agreed milestones”, left in the notes to editors.¹⁵ Every one of these is a failure if the goal was a working datacentre. None of them is a failure if the goal was the announcement. Measured against the press release, the system performed immaculately. The photo was taken, the number was banked, the minister moved on. The announcement was the deliverable.

The paperwork confirms it. On the day of the announcement the 500-megawatt datacentre at the heart of the zone had no planning application before the council at all, only a proposal-of-application notice, the form a developer files to say it intends to apply, lodged three weeks earlier. The one datacentre permission the company actually holds, on a separate plot nearby, is a Permission in Principle. On the planning register, the thing the government announced as delivered existed as a notice of intent.¹⁶

That Chapelhall consent bears this out to the letter. When North Lanarkshire granted it in January, it attached seventeen conditions, on coal-mining safety, contamination, noise, air quality, drainage, trunk-road visibility, parking, landscaping. Not one of them touches how the datacentre is to be powered. The renewable energy the whole scheme is sold on is not a condition of consent, not softened, not even deferred to the later-details stage; it simply is not the planning system’s concern. The single condition that gestures at decarbonisation asks only for a feasibility report on reusing the building’s waste heat, and a feasibility report is a study, not an outcome, discharged by the party who pays for it, sometimes by concluding the thing is not feasible, the same escape route that has quietly dissolved affordable-housing obligations in England for years.¹⁷

Lanarkshire is not the anomaly, it is the genre. The reporter who unpicked it had already found the UK’s flagship AI programme resting on what she called phantom investments: a £1.9bn contract the government announced and then never signed, a promised supercomputer that turned out to be a scaffolding yard.¹⁸ The machinery counts pledges at the moment of announcement and does not check them afterwards, because checking is not what the machinery is for. The social scientist Donald Campbell’s law holds that once an indicator becomes the basis for reward, institutions begin to optimise the indicator rather than the thing it was meant to measure.¹⁹ AI investment announcements look increasingly like that kind of metric. If ministers are rewarded for billions announced rather than datacentres connected to the grid, then the announcement ceases to describe delivery. It becomes the deliverable.

Blaming planning alone lets two larger culprits walk. The physical bottleneck is the grid. Around 100GW of datacentre projects sit in the connection queue, and because most cannot connect, more than a hundred are now asking to burn gas instead, some permanently, enough to put the 2030 target of keeping unabated gas below five per cent out of reach.²⁰ No planning reform shortens that queue. The levers that would, connection policy, energy strategy, are reserved to Westminster, while the consenting power that absorbs the consequences is devolved to Scotland. Planning looks like the root cause because it is the only lever within reach. A diagnosis that lands on the only thing you can touch deserves suspicion.

What Scotland nets

Set the £8.2bn aside and ask what reaches the public purse, what happens to bills, and how much of the money stays. The purse sees little, and the design intends it. Corporation tax on DataVita and CoreWeave is reserved to Westminster, and on a build this heavy in plant it is close to nil for years anyway, because full expensing lets a company write the whole of its qualifying kit off against profit in the year it is bought.²¹ The servers cancel the tax. What Holyrood can touch is non-domestic rates and Scottish income tax, and both are thin. Rates are devolved, but Scotland’s own Business Growth Accelerator Relief holds a new building off the roll until about a year after it opens and relieves the improvements after, so the state’s own hand lightens the bill.²² Income tax falls on the permanent staff, and a datacentre turns electricity into computation, not people into wages: a 500MW hall runs largely unmanned, on a few dozen technicians who come in to swap or rewire failed kit, not the thousands in the headline, most of which were construction and, the charity’s FOI showed, borrowed from a site in Northumberland.

Set beside England, Scotland captures less still. A designated English growth zone keeps every pound of its business-rates growth for twenty-five years; Scotland has arranged no such capture, so the recurring public revenue behind the £8.2bn banner is plausibly low tens of millions a year, unstateable more precisely because a datacentre that does not yet exist has no published rateable value.

Bills move the wrong way. A 500MW load bolted onto a constrained grid adds balancing and network cost, and that is spread across every bill, not the operator’s. The only discount actually on the table runs the other way: central Scotland’s datacentres, this one among them, are in line for up to £24 off each megawatt-hour of grid power from 2027, a saving for the datacentre, not the household.²³ DataVita’s advertised low-cost power is its own private-wire cost, and where the public purse lowers an operator’s price that is a subsidy travelling one direction. The single arrangement that would let a datacentre lower the system’s cost, a flexible load that switches off when the wind drops and drinks it when it blows, is the one the application does not contain.

Most of the £8.2bn is ephemeral. It is a single self-supplied number spread across three components, datacentre halls, a 1GW renewables park and ‘innovation parks’ of labs and robotics, with no line-item split published and two of the three, the renewables and the innovation parks, still unbuilt.²⁴ The one line with a real invoice is the chips, and it is a wasting one: NVIDIA silicon designed in California and fabricated in Taiwan, £1.5bn from an American operator, on a product line the maker now replaces yearly and that critics put at two or three years’ useful life, though the operator’s own books say six.²⁵ It enters as Scottish investment and leaves as a purchase order abroad. Scotland keeps the concrete and the construction wages while they last, the land, a thin tail of power and maintenance, and the lightened rates. The computation is sold on to customers mostly elsewhere, and whatever profit survives is easy to move: the standard route is an intra-group licence fee, the local company paying not the operator but a sister company that owns the platform, the brand and the software and is domiciled wherever the tax is lowest, an IP box in Ireland or Luxembourg, a holding company in Bermuda or the Caymans, until the Scottish profit is a rounding error and the earnings surface somewhere they are barely taxed at all. Policing that, transfer pricing, diverted-profits rules, the global minimum rate, is a reserved matter, imperfectly enforced, and Scotland holds no share of whatever it recovers.

The £543m community fund, promised from revenues “if it generates them”, pegs the town’s benefit to reported local profit, the one figure a group this size has every legal means to minimise. Ordinary accountants save a company tens of millions; the structures open to a multinational save it hundreds, comfortably more than the fund is meant to hold. The fund and the Treasury are waiting on the same number, and it is the number the tax system is best at making small.

Look last at the transaction itself, the datacentre stripped out. What the developer spent to reach the announcement is small, and what it bought is durable. The planning fee was about £13,000 for the one consented site, a Permission in Principle over 1.8 hectares at Chapelhall, with even the 500MW flagship capped at around £93,000 when its application lands.²⁶ That permission can sit unbuilt for five years before it lapses, an asset to hold, sell or develop as the grid queue allows, not a duty to build.²⁷ The renewables remain a page on a website, the turbines a pre-planning sketch. Set the public side beside it, £5m and a government designation, and the shape is plain: the state gave the consent and the credibility at signing and took back promises that cost nothing to abandon.²⁸ If the grid connection never comes, the developer keeps the consent and the headline. The public keeps the press release, rewritten as news.

If we weren’t starting from here

Ask what the state is actually buying, because the current strategy never does; it defaults to the only number in the room, the developer’s. For Scotland the thing worth buying is grid value and verified capability, not jobs; the jobs are mostly construction, and the headline was borrowed from another site. Decide that first and the conditions worth imposing follow; leave it undecided and you are bargaining over someone else’s press release.

So invert the failure: make the permission reversible, and check the promise before it counts. Verify before you announce. The state cites no figure it has not independently audited, and the ones that matter are binary: is the connection secured, is the land controlled, is the capital committed. That one rule ends the phantom-investment loop. It also takes the announcement away from the politician, which is exactly why it would be the hardest sentence in the bill to keep.

Stage the designation, and let it lapse. Growth-zone status stops being a single gift and becomes a sequence of gates, each releasing one benefit, planning help, then connection priority, then fiscal support, against one verified milestone: land assembled, connection secured, a first tranche of firm and renewable generation consented and financed. Miss a gate, lose the status. That is use-it-or-lose-it applied the whole way down, and it fits the connection reform the regulator is already attempting: order the queue by system value, and make the growth-zone prize a place in that order, granted only against real conditions. The developer who means it signs; the one selling a promise walks.

Condition only what cannot move. Compute is mobile and impatient; bind everything and it builds in Virginia while Scotland gets nothing, which is the real price of a reflexive refusal. So bind what cannot follow the data across the water: where the load sits, whether it can be interrupted, its call on the grid, the jurisdiction it answers to, and a public stake in the load itself. Leave nationality and local content alone, because those conditions break the procurement treaties Scotland is signed into.²⁹ And give “green” a definition with teeth, so a shed ringed by diesel fails it on sight.

None of this is free, and pretending otherwise is how such schemes die on contact. Verification assumes a technical capacity the state has just shown it lacks; an independent infrastructure-assurance body, whose audited numbers are the only ones a minister may cite, is a new institution, slow to build and easy to capture. The reserved-devolved split means a clean Scottish design has to be assembled from the levers Holyrood actually holds, planning conditions, the consenting regime above fifty megawatts, procurement, public equity through the national investment bank, the same short list the country keeps arriving at. And there is a curve of conditionality no one can plot in advance: too few conditions and you are back to gifts, too many and the capital is already gone.

The hardest objection is the one that dissolves the complaint. All of this assumes the strategy is failing. It may be succeeding at a purpose it never prints. If the point is to be seen to win the AI race, to signal to mobile capital that Scotland is open for it before a rival does, then press-release industrial strategy is not broken. It is optimal. To call it a failure quietly assumes we agree on the purpose, and the evidence is that we do not. The system serves the announcement perfectly. It fails only against the purpose written on the tin.

Which leaves the reform that actually matters, and it is not a planning reform. It is to make the advertised purpose the real one, by making delivery, not announcement, the thing that pays. Until a minister loses more from an unbuilt datacentre than from an unannounced one, the announcement will remain the deliverable.

airt.scot.