Two documents, four days apart
On 31 August 2026 the government opened the first four competitions of the £100 million Sovereign AI R&D Procurement Scheme, a joint HM Treasury and Cabinet Office announcement that casts government as an early customer for British AI start-ups. Four days later, Housing Digital reported on a paper by Manchester’s Nathan Critch and Edinburgh’s Darcy Luke in The Political Quarterly that describes almost ten years of UK AI policy as “hyperactive incrementalism”, and warns that AI could widen regional and public-sector inequalities rather than narrow them. Our news report on the study covers its findings. This piece asks a narrower question: when the paper’s constraints are set against the scheme’s design, which ones does the scheme address, and which does it leave where they were?
Strategic Insight: The paper is not new research. Crossref records it as published online on 31 October 2025, ten months before the scheme launched, so it could not have assessed it. That makes it a useful test: a list of constraints written before the policy, against which the policy can be read.
What the paper says is in the way
The paper’s abstract names five structural constraints on delivery: “compute capacity, energy infrastructure, data quality, diffusion barriers and skills shortages”. In a companion blog for The Political Quarterly, the authors describe successive agendas as “short-term, under-evaluated, reactive, fragmented, incremental, and top-down”, producing churn that left underlying problems unaddressed.
The scheme starts from a different diagnosis
The gov.uk release locates the problem somewhere else. Smaller companies, it says, are shut out of national and local government contracts because they cannot show the revenue, reserves or history that bigger suppliers can. The remedy is supply-side: upfront payments where appropriate, winners keeping their intellectual property, and the state acting as first customer for technology at demonstrator stage. AI Minister Kanishka Narayan framed it as opening up the billions the state spends on procurement each year to British AI innovators. Our news report on the launch has the detail of each challenge, and our analysis of Sovereign AI’s funding sets the £100 million against what industry says the job costs.
Neither diagnosis rules out the other. But they point at different bottlenecks, and only one of the first four challenges, on compute efficiency, addresses a constraint from the paper’s list.
| Constraint in the paper | Evidence the paper cites | What the first four competitions do |
|---|---|---|
| Compute and energy | UK holds around 3% of world compute; a twentyfold rise could take AI infrastructure from about 3.6TWh (2020) to 72TWh | One challenge, with ARIA’s Scaling Inference Lab, backs more efficient AI computing infrastructure |
| Public-sector data | 62% of government bodies answering an NAO survey said lack of access to good-quality data was a major obstacle | No challenge targets data quality; the NHS and defence challenges both work with government data |
| Skills | Shortages described as especially severe in local government | No challenge is run with a local authority |
| Diffusion | Only 15% of UK firms used one or more AI technologies in 2022; AI firms and users cluster in the South East and London | The Chancellor wants more of AI’s benefits felt “in every UK postcode”; the release states no regional criteria |
Strategic Reality: The scheme’s four partners are the health department (DHSC), the business and science department (DBIST) with ARIA, the Ministry of Defence and the NCSC. None is a local authority. Two of the paper’s warnings, on skills and on regional inequality, land outside the national bodies the scheme works with.
Where the scheme meets the paper, and where it does not
Compute efficiency is the clearest overlap
The “Driving compute efficiency” challenge is the competition that lines up most clearly with the paper’s list. Its stated purpose is to support technologies that make AI computing infrastructure more efficient, in service of the government’s ambition to expand public AI compute capacity substantially while cutting costs for researchers and firms.
That framing is about cost and capacity. The paper’s concern here is energy: the authors estimate that the Action Plan’s twentyfold compute target could push AI infrastructure’s consumption to over a quarter of total UK electricity use, in tension with net-zero commitments, grid constraints and limited storage. The release does not use the words “energy”, “electricity” or “net zero”. Whatever it achieves on energy, the competition is not framed around it, and it serves the same expansion target the paper costs out.
Critical Context: The paper’s 72TWh figure is the researchers’ own estimate of what a twentyfold increase in compute could mean, as reported by Housing Digital. It is what a twentyfold increase “could” mean, not a forecast of what will be built.
The data constraint sits underneath two of the challenges
The NHS productivity challenge wants AI that can take on workflow automation, care coordination and decision support in health services. The defence challenge wants frontier AI and military data joined up securely across defence systems. Both are demonstrator projects meant to be tested in real operational environments.
The paper’s evidence is that much government data sits on old, fragmented IT, and that council data can be split across incompatible platforms. The companion blog puts it more bluntly: “most public sector data in the UK remains locked behind out-of-date legacy IT systems.” A start-up with upfront funding and a government customer still has to work with whatever data that customer holds.
Hidden Cost: The scheme funds the supplier side of a pilot. The paper’s evidence suggests the buyer’s data is often the harder part, and nothing in the release says who pays to make it usable.
”Hyperactive incrementalism” is a test the scheme has not yet faced
The paper’s charge is a pattern of repeated, short-horizon initiatives with limited evaluation. Sovereign AI presents itself as something new: the release’s notes for editors describe a unit unlike its government-backed predecessors, operating as a state-backed venture investor.
The same notes describe the first competitions as a trial of the delivery model, with more challenges to come. Whether that is a break from incrementalism or another instance of it depends on what happens to the test. The release says nothing about evaluation or about publishing results.
⚠️ Warning: A scheme described as a test, with no stated evaluation, is exactly the shape of initiative the paper describes. That is not proof the scheme will repeat the pattern. It is the question to put to it.
Who is affected
Start-ups and SMEs
They are the scheme’s intended beneficiaries. The release offers upfront payments where appropriate, retention of intellectual property, and applications judged by Sovereign AI, the departments taking part and outside technical experts. What it does not say is where those firms are. The paper found that AI firms and the businesses using AI cluster in the South East and London, so a supply-side scheme open to “British” firms is likely to draw on that concentration unless something in its design pulls the other way.
Local government and housing
Housing Digital singled out organisations in housing and local services as the readers for whom the paper’s data and skills findings are likely to resonate: council data split across systems that do not connect, and digital skills shortages that the researchers describe as especially severe in local government. The release itself names local government contracts as one place small firms are shut out. None of the first four challenges involves a council.
Regions and sectors already behind
The researchers warn of a risk that the most productive firms and places pull further ahead, leaving weaker areas and sectors behind. It names hospitality, health and retail as sectors where uptake was low. Health is the one of those the scheme reaches, through the NHS challenge.
| Group | What the scheme changes for them | What is still unknown |
|---|---|---|
| AI start-ups and SMEs | Upfront payments, IP retention, a government first customer | Whether location features in assessment |
| Councils and housing providers | Nothing directly in the first round | Whether a later challenge is run with local government |
| NHS bodies | A challenge aimed at workflow automation and care coordination | Whether data access is resolved before demonstrators are tested |
The open challenges
A first customer needs the capacity to be one
Winning firms build their technology alongside government departments, so the scheme relies on those departments as buyers. The paper documents persistent digital skills shortages across the public sector. The release does not say how participating departments will be resourced to specify, test and adopt what the competitions produce.
Efficiency does not settle the energy arithmetic
The compute efficiency challenge targets cost and efficiency. The paper’s tension is between total consumption under a twentyfold target and net-zero commitments. The release frames the challenge as supporting that expansion, so the energy question the paper raises remains with the expansion target, not with this scheme.
The paper predates the scheme
Critch and Luke were writing about the AI Opportunities Action Plan and the government’s agenda as it stood in 2025. Their critique applies to the scheme only by inference. A fair reading holds the scheme to the constraints the paper documents, not to a verdict the authors never gave on it.
Reality Check: According to Housing Digital, the researchers call Labour’s plans the most comprehensive attempt by any UK government to make use of AI. Their argument is about the depth of the barriers, not the absence of ambition.
What to watch
The release says further challenges are expected as the programme expands. Three things in the next round would show whether the scheme is moving towards the paper’s constraints or away from them:
- A challenge run with local government. The first round has none. A council-partnered challenge would take the scheme to where the paper aims its skills and data warnings.
- A published evaluation of the first round. The government calls these competitions a trial of the delivery model. Results published with what worked and what did not would answer the “under-evaluated” charge directly; a second round with no account of the first would fit it.
- Where the winners are based. If the government publishes successful firms, their locations can be set against the paper’s finding that AI activity clusters in the South East and London.
The researchers conclude, as Housing Digital reports it, that AI may end up “another problem for an already problem-laden government, rather than the solution it is hoped to be.” The procurement scheme is a reasonable answer to one bottleneck, small firms shut out of public contracts. Its first round reaches one of the paper’s five constraints directly, compute, and leaves data, skills, energy demand and uneven adoption to whatever comes next. The next round of challenges will show whether the scheme moves towards the rest.
Source: AI rollout risks deepening UK inequalities, study warns (Housing Digital, 2026). Paper: Nathan Critch and Darcy Luke, A Critical Appraisal of Labour’s AI Agenda, The Political Quarterly 97(1), 2026. Scheme: £100 million competition to back British AI companies to fix public services (HM Treasury and Cabinet Office, 2026).
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the scheme’s next challenges reach local government and the regions the paper says AI has yet to reach. Read more analysis at Insights, or get in touch.