Time to Reset

Neil D. Lawrence

$$\newcommand{\tk}[1]{} \newcommand{\Amatrix}{\mathbf{A}} \newcommand{\KL}[2]{\text{KL}\left( #1\,\|\,#2 \right)} \newcommand{\Kaast}{\kernelMatrix_{\mathbf{ \ast}\mathbf{ \ast}}} \newcommand{\Kastu}{\kernelMatrix_{\mathbf{ \ast} \inducingVector}} \newcommand{\Kff}{\kernelMatrix_{\mappingFunctionVector \mappingFunctionVector}} \newcommand{\Kfu}{\kernelMatrix_{\mappingFunctionVector \inducingVector}} \newcommand{\Kuast}{\kernelMatrix_{\inducingVector \bf\ast}} \newcommand{\Kuf}{\kernelMatrix_{\inducingVector \mappingFunctionVector}} \newcommand{\Kuu}{\kernelMatrix_{\inducingVector \inducingVector}} \newcommand{\Kuui}{\Kuu^{-1}} \newcommand{\Qaast}{\mathbf{Q}_{\bf \ast \ast}} \newcommand{\Qastf}{\mathbf{Q}_{\ast \mappingFunction}} \newcommand{\Qfast}{\mathbf{Q}_{\mappingFunctionVector \bf \ast}} \newcommand{\Qff}{\mathbf{Q}_{\mappingFunctionVector \mappingFunctionVector}} \newcommand{\aMatrix}{\mathbf{A}} \newcommand{\aScalar}{a} \newcommand{\aVector}{\mathbf{a}} \newcommand{\acceleration}{a} \newcommand{\bMatrix}{\mathbf{B}} \newcommand{\bScalar}{b} \newcommand{\bVector}{\mathbf{b}} \newcommand{\basisFunc}{\phi} \newcommand{\basisFuncVector}{\boldsymbol{ \basisFunc}} \newcommand{\basisFunction}{\phi} \newcommand{\basisLocation}{\mu} \newcommand{\basisMatrix}{\boldsymbol{ \Phi}} \newcommand{\basisScalar}{\basisFunction} \newcommand{\basisVector}{\boldsymbol{ \basisFunction}} \newcommand{\activationFunction}{\phi} \newcommand{\activationMatrix}{\boldsymbol{ \Phi}} \newcommand{\activationScalar}{\basisFunction} \newcommand{\activationVector}{\boldsymbol{ \basisFunction}} \newcommand{\bigO}{\mathcal{O}} \newcommand{\binomProb}{\pi} \newcommand{\cMatrix}{\mathbf{C}} \newcommand{\cbasisMatrix}{\hat{\boldsymbol{ \Phi}}} \newcommand{\cdataMatrix}{\hat{\dataMatrix}} \newcommand{\cdataScalar}{\hat{\dataScalar}} \newcommand{\cdataVector}{\hat{\dataVector}} \newcommand{\centeredKernelMatrix}{\mathbf{ \MakeUppercase{\centeredKernelScalar}}} \newcommand{\centeredKernelScalar}{b} \newcommand{\centeredKernelVector}{\centeredKernelScalar} \newcommand{\centeringMatrix}{\mathbf{H}} \newcommand{\chiSquaredDist}[2]{\chi_{#1}^{2}\left(#2\right)} \newcommand{\chiSquaredSamp}[1]{\chi_{#1}^{2}} \newcommand{\conditionalCovariance}{\boldsymbol{ \Sigma}} \newcommand{\coregionalizationMatrix}{\mathbf{B}} \newcommand{\coregionalizationScalar}{b} \newcommand{\coregionalizationVector}{\mathbf{ \coregionalizationScalar}} \newcommand{\covDist}[2]{\text{cov}_{#2}\left(#1\right)} \newcommand{\covSamp}[1]{\text{cov}\left(#1\right)} \newcommand{\covarianceScalar}{c} \newcommand{\covarianceVector}{\mathbf{ \covarianceScalar}} \newcommand{\covarianceMatrix}{\mathbf{C}} \newcommand{\covarianceMatrixTwo}{\boldsymbol{ \Sigma}} \newcommand{\croupierScalar}{s} \newcommand{\croupierVector}{\mathbf{ \croupierScalar}} \newcommand{\croupierMatrix}{\mathbf{ \MakeUppercase{\croupierScalar}}} \newcommand{\dataDim}{p} \newcommand{\dataIndex}{i} \newcommand{\dataIndexTwo}{j} \newcommand{\dataMatrix}{\mathbf{Y}} \newcommand{\dataScalar}{y} \newcommand{\dataSet}{\mathcal{D}} \newcommand{\dataStd}{\sigma} \newcommand{\dataVector}{\mathbf{ \dataScalar}} \newcommand{\decayRate}{d} \newcommand{\degreeMatrix}{\mathbf{ \MakeUppercase{\degreeScalar}}} \newcommand{\degreeScalar}{d} \newcommand{\degreeVector}{\mathbf{ \degreeScalar}} \newcommand{\diag}[1]{\text{diag}\left(#1\right)} \newcommand{\diagonalMatrix}{\mathbf{D}} \newcommand{\diff}[2]{\frac{\text{d}#1}{\text{d}#2}} \newcommand{\diffTwo}[2]{\frac{\text{d}^2#1}{\text{d}#2^2}} \newcommand{\displacement}{x} \newcommand{\displacementVector}{\textbf{\displacement}} \newcommand{\distanceMatrix}{\mathbf{ \MakeUppercase{\distanceScalar}}} \newcommand{\distanceScalar}{d} \newcommand{\distanceVector}{\mathbf{ \distanceScalar}} \newcommand{\eigenvaltwo}{\ell} \newcommand{\eigenvaltwoMatrix}{\mathbf{L}} \newcommand{\eigenvaltwoVector}{\mathbf{l}} \newcommand{\eigenvalue}{\lambda} \newcommand{\eigenvalueMatrix}{\boldsymbol{ \Lambda}} \newcommand{\eigenvalueVector}{\boldsymbol{ \lambda}} \newcommand{\eigenvector}{\mathbf{ \eigenvectorScalar}} \newcommand{\eigenvectorMatrix}{\mathbf{U}} \newcommand{\eigenvectorScalar}{u} \newcommand{\eigenvectwo}{\mathbf{v}} \newcommand{\eigenvectwoMatrix}{\mathbf{V}} \newcommand{\eigenvectwoScalar}{v} \newcommand{\entropy}[1]{\mathcal{H}\left(#1\right)} \newcommand{\errorFunction}{E} \newcommand{\expDist}[2]{\left\langle#1\right\rangle_{#2}} \newcommand{\expSamp}[1]{\left\langle#1\right\rangle} \newcommand{\expectation}[1]{\left\langle #1 \right\rangle } \newcommand{\expectationDist}[2]{\left\langle #1 \right\rangle _{#2}} \newcommand{\expectedDistanceMatrix}{\mathcal{D}} \newcommand{\eye}{\mathbf{I}} \newcommand{\fantasyDim}{r} \newcommand{\fantasyMatrix}{\mathbf{ \MakeUppercase{\fantasyScalar}}} \newcommand{\fantasyScalar}{z} \newcommand{\fantasyVector}{\mathbf{ \fantasyScalar}} \newcommand{\featureStd}{\varsigma} \newcommand{\gammaCdf}[3]{\mathcal{GAMMA CDF}\left(#1|#2,#3\right)} \newcommand{\gammaDist}[3]{\mathcal{G}\left(#1|#2,#3\right)} \newcommand{\gammaSamp}[2]{\mathcal{G}\left(#1,#2\right)} \newcommand{\gaussianDist}[3]{\mathcal{N}\left(#1|#2,#3\right)} \newcommand{\gaussianSamp}[2]{\mathcal{N}\left(#1,#2\right)} \newcommand{\uniformDist}[3]{\mathcal{U}\left(#1|#2,#3\right)} \newcommand{\uniformSamp}[2]{\mathcal{U}\left(#1,#2\right)} \newcommand{\given}{|} \newcommand{\half}{\frac{1}{2}} \newcommand{\heaviside}{H} \newcommand{\hiddenMatrix}{\mathbf{ \MakeUppercase{\hiddenScalar}}} \newcommand{\hiddenScalar}{h} \newcommand{\hiddenVector}{\mathbf{ \hiddenScalar}} \newcommand{\identityMatrix}{\eye} \newcommand{\inducingInputScalar}{z} \newcommand{\inducingInputVector}{\mathbf{ \inducingInputScalar}} \newcommand{\inducingInputMatrix}{\mathbf{Z}} \newcommand{\inducingScalar}{u} \newcommand{\inducingVector}{\mathbf{ \inducingScalar}} \newcommand{\inducingMatrix}{\mathbf{U}} \newcommand{\inlineDiff}[2]{\text{d}#1/\text{d}#2} \newcommand{\inputDim}{q} \newcommand{\inputMatrix}{\mathbf{X}} \newcommand{\inputScalar}{x} \newcommand{\inputSpace}{\mathcal{X}} \newcommand{\inputVals}{\inputVector} \newcommand{\inputVector}{\mathbf{ \inputScalar}} \newcommand{\iterNum}{k} \newcommand{\kernel}{\kernelScalar} \newcommand{\kernelMatrix}{\mathbf{K}} \newcommand{\kernelScalar}{k} \newcommand{\kernelVector}{\mathbf{ \kernelScalar}} \newcommand{\kff}{\kernelScalar_{\mappingFunction \mappingFunction}} \newcommand{\kfu}{\kernelVector_{\mappingFunction \inducingScalar}} \newcommand{\kuf}{\kernelVector_{\inducingScalar \mappingFunction}} \newcommand{\kuu}{\kernelVector_{\inducingScalar \inducingScalar}} \newcommand{\lagrangeMultiplier}{\lambda} \newcommand{\lagrangeMultiplierMatrix}{\boldsymbol{ \Lambda}} \newcommand{\lagrangian}{L} \newcommand{\laplacianFactor}{\mathbf{ \MakeUppercase{\laplacianFactorScalar}}} \newcommand{\laplacianFactorScalar}{m} \newcommand{\laplacianFactorVector}{\mathbf{ \laplacianFactorScalar}} \newcommand{\laplacianMatrix}{\mathbf{L}} \newcommand{\laplacianScalar}{\ell} \newcommand{\laplacianVector}{\mathbf{ \ell}} \newcommand{\latentDim}{q} \newcommand{\latentDistanceMatrix}{\boldsymbol{ \Delta}} \newcommand{\latentDistanceScalar}{\delta} \newcommand{\latentDistanceVector}{\boldsymbol{ \delta}} \newcommand{\latentForce}{f} \newcommand{\latentFunction}{u} \newcommand{\latentFunctionVector}{\mathbf{ \latentFunction}} \newcommand{\latentFunctionMatrix}{\mathbf{ \MakeUppercase{\latentFunction}}} \newcommand{\latentIndex}{j} \newcommand{\latentScalar}{z} \newcommand{\latentVector}{\mathbf{ \latentScalar}} \newcommand{\latentMatrix}{\mathbf{Z}} \newcommand{\learnRate}{\eta} \newcommand{\lengthScale}{\ell} \newcommand{\rbfWidth}{\ell} \newcommand{\likelihoodBound}{\mathcal{L}} \newcommand{\likelihoodFunction}{L} \newcommand{\locationScalar}{\mu} \newcommand{\locationVector}{\boldsymbol{ \locationScalar}} \newcommand{\locationMatrix}{\mathbf{M}} \newcommand{\variance}[1]{\text{var}\left( #1 \right)} \newcommand{\mappingFunction}{f} \newcommand{\mappingFunctionMatrix}{\mathbf{F}} \newcommand{\mappingFunctionTwo}{g} \newcommand{\mappingFunctionTwoMatrix}{\mathbf{G}} \newcommand{\mappingFunctionTwoVector}{\mathbf{ \mappingFunctionTwo}} \newcommand{\mappingFunctionVector}{\mathbf{ \mappingFunction}} \newcommand{\scaleScalar}{s} \newcommand{\mappingScalar}{w} \newcommand{\mappingVector}{\mathbf{ \mappingScalar}} \newcommand{\mappingMatrix}{\mathbf{W}} \newcommand{\mappingScalarTwo}{v} \newcommand{\mappingVectorTwo}{\mathbf{ \mappingScalarTwo}} \newcommand{\mappingMatrixTwo}{\mathbf{V}} \newcommand{\maxIters}{K} \newcommand{\meanMatrix}{\mathbf{M}} \newcommand{\meanScalar}{\mu} \newcommand{\meanTwoMatrix}{\mathbf{M}} \newcommand{\meanTwoScalar}{m} \newcommand{\meanTwoVector}{\mathbf{ \meanTwoScalar}} \newcommand{\meanVector}{\boldsymbol{ \meanScalar}} \newcommand{\mrnaConcentration}{m} \newcommand{\naturalFrequency}{\omega} \newcommand{\neighborhood}[1]{\mathcal{N}\left( #1 \right)} \newcommand{\neilurl}{http://inverseprobability.com/} \newcommand{\noiseMatrix}{\boldsymbol{ E}} \newcommand{\noiseScalar}{\epsilon} \newcommand{\noiseVector}{\boldsymbol{ \epsilon}} \newcommand{\noiseStd}{\sigma} \newcommand{\norm}[1]{\left\Vert #1 \right\Vert} \newcommand{\normalizedLaplacianMatrix}{\hat{\mathbf{L}}} \newcommand{\normalizedLaplacianScalar}{\hat{\ell}} \newcommand{\normalizedLaplacianVector}{\hat{\mathbf{ \ell}}} \newcommand{\numActive}{m} \newcommand{\numBasisFunc}{m} \newcommand{\numComponents}{m} \newcommand{\numComps}{K} \newcommand{\numData}{n} \newcommand{\numFeatures}{K} \newcommand{\numHidden}{h} \newcommand{\numInducing}{m} \newcommand{\numLayers}{\ell} \newcommand{\numNeighbors}{K} \newcommand{\numSequences}{s} \newcommand{\numSuccess}{s} \newcommand{\numTasks}{m} \newcommand{\numTime}{T} \newcommand{\numTrials}{S} \newcommand{\outputIndex}{j} \newcommand{\paramVector}{\boldsymbol{ \theta}} \newcommand{\parameterMatrix}{\boldsymbol{ \Theta}} \newcommand{\parameterScalar}{\theta} \newcommand{\parameterVector}{\boldsymbol{ \parameterScalar}} \newcommand{\partDiff}[2]{\frac{\partial#1}{\partial#2}} \newcommand{\precisionScalar}{j} \newcommand{\precisionVector}{\mathbf{ \precisionScalar}} \newcommand{\precisionMatrix}{\mathbf{J}} \newcommand{\pseudotargetScalar}{\widetilde{y}} \newcommand{\pseudotargetVector}{\mathbf{ \pseudotargetScalar}} \newcommand{\pseudotargetMatrix}{\mathbf{ \widetilde{Y}}} \newcommand{\rank}[1]{\text{rank}\left(#1\right)} \newcommand{\rayleighDist}[2]{\mathcal{R}\left(#1|#2\right)} \newcommand{\rayleighSamp}[1]{\mathcal{R}\left(#1\right)} \newcommand{\responsibility}{r} \newcommand{\rotationScalar}{r} \newcommand{\rotationVector}{\mathbf{ \rotationScalar}} \newcommand{\rotationMatrix}{\mathbf{R}} \newcommand{\sampleCovScalar}{s} \newcommand{\sampleCovVector}{\mathbf{ \sampleCovScalar}} \newcommand{\sampleCovMatrix}{\mathbf{s}} \newcommand{\scalarProduct}[2]{\left\langle{#1},{#2}\right\rangle} \newcommand{\sign}[1]{\text{sign}\left(#1\right)} \newcommand{\sigmoid}[1]{\sigma\left(#1\right)} \newcommand{\singularvalue}{\ell} \newcommand{\singularvalueMatrix}{\mathbf{L}} \newcommand{\singularvalueVector}{\mathbf{l}} \newcommand{\sorth}{\mathbf{u}} \newcommand{\spar}{\lambda} \newcommand{\trace}[1]{\text{tr}\left(#1\right)} \newcommand{\BasalRate}{B} \newcommand{\DampingCoefficient}{C} \newcommand{\DecayRate}{D} \newcommand{\Displacement}{X} \newcommand{\LatentForce}{F} \newcommand{\Mass}{M} \newcommand{\Sensitivity}{S} \newcommand{\basalRate}{b} \newcommand{\dampingCoefficient}{c} \newcommand{\mass}{m} \newcommand{\sensitivity}{s} \newcommand{\springScalar}{\kappa} \newcommand{\springVector}{\boldsymbol{ \kappa}} \newcommand{\springMatrix}{\boldsymbol{ \mathcal{K}}} \newcommand{\tfConcentration}{p} \newcommand{\tfDecayRate}{\delta} \newcommand{\tfMrnaConcentration}{f} \newcommand{\tfVector}{\mathbf{ \tfConcentration}} \newcommand{\velocity}{v} \newcommand{\sufficientStatsScalar}{g} \newcommand{\sufficientStatsVector}{\mathbf{ \sufficientStatsScalar}} \newcommand{\sufficientStatsMatrix}{\mathbf{G}} \newcommand{\switchScalar}{s} \newcommand{\switchVector}{\mathbf{ \switchScalar}} \newcommand{\switchMatrix}{\mathbf{S}} \newcommand{\tr}[1]{\text{tr}\left(#1\right)} \newcommand{\loneNorm}[1]{\left\Vert #1 \right\Vert_1} \newcommand{\ltwoNorm}[1]{\left\Vert #1 \right\Vert_2} \newcommand{\onenorm}[1]{\left\vert#1\right\vert_1} \newcommand{\twonorm}[1]{\left\Vert #1 \right\Vert} \newcommand{\vScalar}{v} \newcommand{\vVector}{\mathbf{v}} \newcommand{\vMatrix}{\mathbf{V}} \newcommand{\varianceDist}[2]{\text{var}_{#2}\left( #1 \right)} \newcommand{\vecb}[1]{\left(#1\right):} \newcommand{\weightScalar}{w} \newcommand{\weightVector}{\mathbf{ \weightScalar}} \newcommand{\weightMatrix}{\mathbf{W}} \newcommand{\weightedAdjacencyMatrix}{\mathbf{A}} \newcommand{\weightedAdjacencyScalar}{a} \newcommand{\weightedAdjacencyVector}{\mathbf{ \weightedAdjacencyScalar}} \newcommand{\onesVector}{\mathbf{1}} \newcommand{\zerosVector}{\mathbf{0}} $$

at Time to Reset, Financial Times, Bracken House, London on Sep 7, 2026 [reveal]

Neil D. Lawrence, University of Cambridge

Abstract

AI could be the UK’s biggest economic opportunity in a generation, but only if its gains spread across the economy rather than pooling in a handful of companies. The returns do not come from hosting the cleverest model. They come from diffusion: millions of firms, public services and workers building on top of it.

This talk argues that UK AI policy has been captured by a supply-side reflex: hyperscale partnerships, compute targets, sovereignty as data centres, while the productivity prize sits on the demand side. Concentration is what stops that value spreading. Sovereignty, in this setting, is not a national champion. It is having alternatives.

The talk suggests that the Digital Markets, Competition and Consumers Act is the right instrument for a market that is still forming. A pro-innovation agenda is not a pro-big-tech agenda. Opening the gateways is the pro-investment move, because it is what lets the next hundred UK challengers raise money and grow.

Not a Race

Concentrated digital markets impose costs. The question is what to do about them. The same AI moment is the biggest growth opportunity in a generation, but only if its value can spread. That will not happen if we treat AI as a race.

I find myself thinking about Life of Brian when I listen to AI policy. The people of Judea, frightened and uncertain, latch on to anyone who arrives with a simple, confident story. The comedy is that they place a regular person as a prophet. The policy analogue is less funny. We are in a moment of technical and economic uncertainty. That uncertainty makes us susceptible to anyone who arrives with a simple story about how to proceed.

The AI race is one of those stories. You can see its footprints in hyperscale partnerships, the sovereignty unit, the compute target. Each one looks like a decision. Most of them are reflexes. A race implies one winner and a field of losers. That is a dangerous frame for the UK. We will not be plucky Brits snatching a bronze. We will be targeting innovations that are not on the entrée of what our citizens and businesses need.

Value in the Spread

The same AI moment is the biggest growth opportunity in a generation. But only if its value can spread. Concentration is precisely what stops it spreading.

The returns from AI do not come from building the cleverest model. They come from diffusion. The genuinely valuable UK products will not come from a handful of frontier labs. They will come from thousands of small firms across the country solving specific, unglamorous problems. That is the growth people actually feel, in the places the last tech wave passed by. Those firms can only build if they can reach a customer without renting access from — or being locked into deals with — the platforms they are competing against.

The Missing Half

If you sit through enough policy meetings on AI, a pattern is hard to miss. Almost everyone advising government has a supply-side interest. They sell infrastructure, or they build models, or they run platforms. Those perspectives matter. But it means the menu of options ministers see is shaped, structurally, by suppliers’ interests. The result is a policy agenda that mistakes the input for the output.

We want broad-based productivity growth, better public services, and a workforce with more agency over its own institutional decision-making. The inputs we have arranged are foreign capital and foreign software. Neither becomes British productivity on its own. The work in between — adoption, institutions, tool-building — is what is missing.

If you are looking for a single intervention, it is not on the supply side. Major US technology firms have enormous and growing lobbying capability. We do not need to worry about whether their interests are being represented. We need to worry about whether everyone else’s are. Mainstream UK businesses left behind by previous digital waves. SMEs that struggle to invest. Councils, hospitals and schools that, left to the market, will be sold platforms rather than supported to build and adapt. That is where the productivity prize lives. It is almost entirely absent from the conversation ministers hear.

The frontier of innovation has also shifted. Recent progress has come from combining and directing existing models rather than training new ones. Models that can be convened in English — or Mandarin, French, Urdu or Spanish — and that can write software tailored to a business’s needs rather than a platform’s bottom line. That drops the cost and skill threshold for institution-led adoption. It means the next decade’s value will be created by the people closest to the problems. None of this happens unless we treat the demand side as a core component of national strategy.

Central Deployment Without Engagement

The Horizon Scandal

[edit]

In the UK we saw these effects play out in the Horizon scandal: the accounting system of the national postal service was computerized by Fujitsu and first installed in 1999, but neither the Post Office nor Fujitsu were able to control the system they had deployed. When it went wrong individual sub postmasters were blamed for the systems’ errors. Over the next two decades they were prosecuted and jailed leaving lives ruined in the wake of the machine’s mistakes.

Figure: The still is from the 2015 select committee.

See Lawrence (2024) Horizon scandal p. 371.

When I gave evidence to the Business and Trade Select Committee I was asked about the UK’s weaknesses. I said one of them is a lack of confidence in our own people, our own businesses, and our own universities. We look across the Atlantic at organisations that are brilliant. We want them here. But by over-emphasising what they can do for the economy, we forget that the growth will not come from them. It will come from our businesses adopting, assimilating and innovating.

The track record of centrally deploying technology on people, without engaging them, is the Horizon scandal. That is not a metaphor I reach for lightly. It is what happens when a system is imposed, the people who use it have no voice, and the institution that deployed it cannot admit the machine is wrong.

Building from the Ground Up

[edit]

Work across Cambridgeshire demonstrates effective ground-up innovation in action. In local government, Greater Cambridge Shared Planning collaborates with universities to develop AI tools that analyse public consultation responses. This partnership has potential to reduce staff analysis time from over a year to just two months, while maintaining planning expertise at the heart of the process.

Similar successes emerged during COVID-19, when NHS England East collaborated with researchers to develop AI models supporting hospital resource allocation. This partnership showed how AI can enhance operational decisions when developed alongside those who understand local needs.

In healthcare, clinician-led development of AI tools for cancer diagnosis demonstrates how frontline expertise can ensure AI enhances rather than replaces professional judgment. Key success factors include:

Local expertise driving development priorities
Solutions addressing concrete operational challenges
Professional judgment remaining central
Iterative development based on real-world feedback

In Cambridge we have been working with South Cambridgeshire District Council on planning. Officers had already formed AI clubs of their own. They wanted to understand the technology and use it in their work. Researchers from Liverpool and Cambridge have been helping them build planning tools that they themselves design, deploy and maintain. The work solves a real operational problem, builds durable capability inside the council, and is generating expertise that could catalyse a business serving local authorities across the country. The budget is on the order of twenty thousand pounds.

What do we hear four months later? Big announcements that Google is going to solve planning across the country. On the day of that announcement I got an email asking: what does this Google thing mean for us? That is concentration strangling diffusion in one sentence. Not because Google is a villain. Because a supply-side announcement, made over the heads of the people already doing the work, chills the very capability we say we want to build.

Concentration Strangles Diffusion

One of the most significant features of the modern AI landscape is the extreme concentration of computing resources required for cutting-edge AI development. This concentration has profound implications for the structure of the AI ecosystem and for policy considerations.

The scale of compute required for training state-of-the-art AI models has grown exponentially. According to some estimates, the compute used for the largest AI training runs has doubled approximately every 6-10 months since 2012, far outpacing Moore’s Law. Training a model like GPT-4 likely required resources in the hundreds of millions of dollars.

This creates a compute barrier that effectively limits who can participate in foundation model development to a small club of well-resourced organizations - primarily large technology companies and a handful of well-funded research labs. Even academic researchers at leading institutions often depend on partnerships with industry to access the necessary resources.

The hardware that powers AI systems is similarly concentrated. NVIDIA has established a dominant position in the market for AI accelerators, with its GPUs being essential for most large-scale AI training. While competitors are emerging, the combination of specialized hardware and software ecosystems creates strong network effects that reinforce concentration.

Cloud computing services add another layer of concentration, with AWS, Google Cloud, and Microsoft Azure controlling much of the market for AI-oriented computing resources. For many organizations, these platforms represent the only viable way to access the hardware needed for AI development and deployment.

This concentration creates several policy challenges. It raises competition concerns, as control over compute resources can translate into control over AI capabilities more broadly. It creates potential vulnerabilities in supply chains and infrastructure that may have national security implications. And it raises questions about equitable access to a technology that is increasingly seen as having broad societal importance.

Policy responses to compute concentration might include public investment in research computing infrastructure, measures to ensure fair access to commercial computing resources, support for hardware diversification and open-source alternatives, and international cooperation to prevent destructive competition over compute resources.

IPPR’s Bottleneck Britain report, launched in the days before this event, sets out how concentrated digital infrastructure already taxes UK business. Search, advertising, cloud, app distribution: a handful of firms set the terms on which everyone else builds. If those terms do not change, AI will not open the market. It will be bundled through the same gateways.

The report’s argument and the committee evidence land in the same place. A country that hosts data centres but cannot diffuse the technology captures almost none of the prize. The window for structural intervention in AI is still open. It will not stay open.

Sovereignty Means Having Alternatives

[edit]

There is little point investing hundreds of millions in a sovereign AI fund to build national champions while leaving untouched a market structure that will see them acquired or outcompeted. Diane Coyle has called the outcome `policy incoherence’ Coyle (2026), pg 70. She is right.

A country whose firms can only build on someone else’s terms is not sovereign, however many data centres it hosts. Sovereignty in this setting is not a national champion. It is having options: a British business that is not left stranded when the terms change in a US boardroom.

When I was asked whether we had outsourced AI model development to private billionaires with no loyalty to the UK, I said yes — with a caveat. These businesses are not evil. They are incentivised differently. When we talk about AI alignment we usually mean aligning models with our values. We should also talk about corporate alignment. These corporations are not aligned with the interests of UK citizens. That is why there is such a gap between what is promised and what is delivered.

The Instrument We Already Have

We already have the tool. The Digital Markets, Competition and Consumers Act gives the CMA powers to keep markets contestable for small and medium suppliers. Its role has been misunderstood in some quarters. A pro-innovation agenda is not the same as a pro-big-tech agenda. Pro-innovation, properly understood, means making sure there is room for the next wave of UK entrants to build, validate and scale. That requires using the powers we already have.

The Act is the right instrument for AI precisely because it is targeted and can move with the market. The market is forming faster than the enforcement around it. Traditional competition law is too slow: by the time an investigation concludes, the harm is entrenched. The DMCCA was designed for that problem. The ask now is not another bill. It is pace, political backing, and a focus on remedies that create competitors, not only fairer terms for dependents.

The Investment We Should Worry About

This is the pro-growth move. The fear has it backwards. Opening the gateways is not the anti-investment play. It is the pro-investment one, because it is what lets the next hundred UK challengers raise money and grow.

The investment we should worry about losing is the investment that never happens, because a founder already knows the gatekeeper can copy, bury or tax them at will. Hyperscalers complain first about energy, planning, employment rights and data protection. Competition enforcement is low on that list. The AI labs the UK most wants to attract are often trying to reduce their dependence on established players. Enforcement that keeps deeper layers of the stack competitive can help them.

Proportionality remains the right test where investment and diplomatic risks exist. But that test still points to enforcement that is bolder than we have seen. Government cannot outsource the judgement of what is proportionate to the incumbents it is supposed to be regulating.

Campus UK

Universities as Innovation Bridges

[edit]

Universities occupy a unique position in the innovation ecosystem. As neutral conveners, they can bring together diverse communities of expertise, from engineering to ethics. Their access to research capabilities and human capital makes them vital for the next wave of AI innovation.

Universities can create spaces where people working to deploy AI in public services and industry can collaborate with diverse communities of expertise. This role is particularly important for:

Facilitating dialogue between technology developers and users
Providing ethical oversight and governance frameworks
Supporting knowledge transfer between research and practice
Building capacity through education and training

See this opinion piece from Neil Lawrence and Jess Montgomery.

If you draw a circle around the educational and research expertise that can be reached by train within the United Kingdom — Southampton, Cambridge, Manchester, Oxford, Glasgow, Derby, Sheffield, and a great many places in between — the concentration of knowledge inside that circle is unmatched. We are, by geography, unusually well placed to translate research into adoption.

We do not currently reward our universities for delivering on that. The incentives push us towards the cover of Nature rather than the cover of the local paper. That forgets why many of these institutions were developed. Our civic universities were founded to advance learning for the benefit of their city and region. AI is a chance for them to do just that. Lincoln’s work with farmers through agricultural robotics. Nottingham Trent’s work on inclusion for students with learning disabilities and autism. That is Campus UK. It is not as glamorous as a frontier-lab announcement. It is meant to be useful.

A Reset

This is not an argument against the large technology companies. They do serious work, and we want them present in the UK. But let us not confuse their presence with a strategy. On its own, their presence does not solve the productivity problem of a British SME, the workflow problem of a planning officer, or the staffing pressure on an NHS ward.

A reset means taking the demand side seriously as a matter in its own right. Not as a downstream consequence of supply. As the battleground where the economic and public value of AI must be realised. The DMCCA is how we keep that battleground open. Perhaps that is a less dramatic framing than a global race. It is a framing that could work for the UK.

Thanks!

For more information on these subjects and more you might want to check the following resources.

References

Coyle, D., 2026. The AI arms race and economic security, in: Harding, V. (Ed.), Reimagining the AI Arms Race. AIxGEO, Bennett School of Public Policy, University of Cambridge, pp. 63–73. https://doi.org/10.17863/CAM.132048

Lawrence, N.D., 2024. The atomic human: Understanding ourselves in the age of AI. Allen Lane.