The AI Scale Race: Are We Competing for Volume Instead of Value?

AI has acquired an industrial-age measure of ambition. Bigger models, more compute, more data centres, more capital: scale has become shorthand for seriousness.

When countries talk about competing in AI, the conversation quickly moves towards infrastructure. Who has the processing power? Who can train the largest models? Who controls the data centres, the chips and the energy required to run them? Those things matter, and Britain cannot simply opt out of the infrastructure on which a growing part of its economy will depend.

But infrastructure is only one layer of the AI economy, and it may not be where the most interesting value is created.

Twenty-five years of solving increasingly specific problems

At a recent B4 Corum lunch, Robert Pilkington, joint CEO of Ebbon Group, spent some time explaining how the business had evolved. Ebbon was founded in 2000 and has spent the past quarter of a century building technology primarily for the automotive sector. Rob described how the business initially focused on software that solved problems within the automotive interests of its shareholders, and how the capability developed there was gradually taken into the wider market.

One of the early products was Leaselink, connecting leasing companies, vehicle manufacturers and dealers through the procurement process. Other products followed: moDel addressed vehicle delivery and collection, while Licence Check and DriverCheck moved the Group further into driver and vehicle compliance. There were also choices about what not to pursue. In 2015 Ebbon-Dacs sold its dealer management system business to Pinewood so it could concentrate on fleet connectivity, a decision that is easy to overlook precisely because it runs against how technology businesses usually get discussed.

Technology businesses are usually discussed through addition: more products, more markets, more capability. But strategy also involves subtraction, and the ability to build something does not mean it is the right thing to build. Ebbon appears to have spent much of its history going deeper into a field it already understood rather than continually broadening the field itself. The result is a business operating in a relatively specific part of the economy while handling activity at considerable scale. Ebbon currently reports 1.2 million digital vehicle handovers, €20 billion of vehicle purchases and more than two million licence checks through its platforms each year.

Niche, in this context, does not mean small. It means specific.

Specificity can scale

There is a tendency to think about scale geometrically. A bigger market creates a bigger opportunity, a more general technology reaches more users, a larger model can answer more questions. That logic works, until specificity itself becomes the advantage.

Leaselink is valuable because it does not need to solve every procurement problem. It needs to understand vehicle leasing exceptionally well. Driver compliance software does not need to manage every part of an organisation. It needs to make a complicated and recurring responsibility significantly easier to control. In both cases the market can remain narrow while the value created within it becomes substantial.

This matters because much of the current AI race is pulling in the opposite direction. Foundation models are becoming increasingly general, built for breadth, needing to perform across huge numbers of subjects, users and applications, and that requires extraordinary amounts of capital and infrastructure. The businesses using them do not necessarily have the same requirement. For many organisations, the valuable question is not how much intelligence they can access, but how precisely that intelligence can be applied to something they understand unusually well.

AI arrived after the knowledge

Ebbon Intelligence makes this part of the story more interesting. It was created in 2021 through a partnership with Oxford University spin-out Zegami, initially exploring applications of AI in areas including vehicle damage and condition assessment, more than twenty years after Ebbon was founded. This sequence matters. This wasn't an AI company searching for an automotive problem. It was an automotive technology business adding AI to more than two decades of accumulated knowledge about customers, processes, systems and industry friction.

Over the last two years Ebbon has evolved Ebbon Intelligence from effectively its AI/Robotics R&D arm into an important future commercial entity for the enterprise, with RALPH, its autonomous mini-drone-based automotive visual marketing product, sitting alongside the Group's established platforms. The AI is important, but it does not arrive alone. It arrives with domain knowledge, with relationships, with an existing route into customers, with an understanding of how automotive businesses operate, and with twenty-five years of learning about which problems are irritating enough, expensive enough or repetitive enough to be worth solving. That combination may prove more significant than the model itself.

Intelligence is becoming cheaper. Knowing what to ask of it isn't.

This connects to another change taking place in strategy. For decades, businesses have operated with constrained analytical capacity: leadership teams could consider only a limited number of strategic options before time, money and human attention forced them to narrow the field. AI changes those economics.

Felipe Csaszar, writing in Harvard Business Review, argues that AI can radically increase the number of strategic alternatives an organisation can generate and examine. It can create richer models of markets, test assumptions and expose plans to structured challenge before a leadership team commits resources. His useful shorthand is that AI expands the search while humans choose, and that has implications well beyond corporate strategy departments.

It reduces the cost of exploring. A specialist business can consider more hypotheses before committing capital, challenge assumptions earlier, and combine its own data and experience with external intelligence to test routes that would previously have been discarded because investigating them was too expensive. The strategic advantage may therefore move away from simply having more information, and increasingly come from knowing which questions deserve asking. That favours expertise, and potentially it favours businesses that have spent years becoming exceptionally knowledgeable about a relatively narrow field.

The same pattern is already reshaping organisations from the inside. When intelligence stops being scarce, judgement becomes the scarce resource in its place, and working out who has the authority to exercise it becomes the harder problem to solve.

Not every useful AI requires hyperscale infrastructure

The infrastructure argument is changing as well. Large frontier models require extraordinary computational resources, and that is unlikely to disappear, but smaller models increasingly operate under very different constraints.

Google's Gemma 3n, for example, is designed to run on everyday devices including phones, laptops and tablets, using techniques intended to reduce memory and compute requirements. Microsoft's current Foundry Local documentation describes small language models that can run on-premises using constrained hardware, trading some breadth of knowledge for efficiency on focused tasks.

That does not make hyperscale infrastructure irrelevant. It changes the assumption that every valuable AI application must depend on it to the same degree. Some applications will need frontier capability continuously; others may operate predominantly within a far smaller computational environment, calling on larger models only when wider intelligence is required. The architecture becomes distributed, and perhaps the economics become distributed too.

Britain still needs infrastructure

It would be easy to take this argument too far. Britain needs compute, it needs energy, it needs data infrastructure, and it needs the skills and investment required to remain an active participant in AI rather than becoming entirely dependent on technology developed elsewhere.

The UK government's own AI Opportunities Action Plan explicitly argues for world-class compute and data infrastructure and for serious domestic participants at multiple layers of the AI stack. That seems reasonable.

The harder question is whether infrastructure volume should also become our primary measure of success. The United States and China possess advantages of capital, scale, energy and industrial capacity that Britain cannot replicate across every part of the technology stack, and trying to match them everywhere risks spreading resources across a race defined by somebody else's strengths.

There is another possibility. Britain may be better placed to combine increasingly accessible AI with specialist knowledge, intellectual property, engineering capability, trusted relationships and deep understanding of particular markets. Not to avoid scale, but to find a different route to it.

The wrong definition of scale

Ebbon's history provides an interesting version of what that can look like. A specific problem becomes software; software becomes a platform; a platform creates relationships and knowledge; those relationships expose adjacent problems; new capability allows those problems to be addressed, and the field deepens. Eventually AI becomes another form of intelligence available within it. This is not scale achieved by becoming more general. It is scale achieved by becoming increasingly valuable within a chosen field.

There is a broader strategic idea here. We tend to treat technology as the scarce resource and applications as what gets built afterwards, but AI may invert some of that logic. As intelligent capability becomes cheaper and more accessible, the scarce assets may increasingly be context, judgement, proprietary knowledge and an understanding of where genuine value sits. The same AI model can be available to thousands of businesses, and they will not all know what to do with it.

Csaszar makes a similar argument about competitive advantage. As access to general-purpose models becomes widespread, differentiation shifts towards the proprietary data, processes and strategic capabilities organisations build around them. Owning the intelligence is not the same as owning the value created from it.

Volume is an input. Value is an outcome.

This is the distinction the AI debate risks losing. We can measure model size, compute capacity, investment, data-centre construction, token volumes. Those numbers tell us something about capability. They tell us much less about what that capability ultimately creates.

The defining capability of modern organisations is no longer simply access to intelligence, but their ability to convert intelligence into value. AI increases the intelligence available. It does not determine which problem matters, decide what an organisation should stand for, or establish which trade-offs are acceptable, and it does not automatically create the judgement, decisions and ownership required to turn an opportunity into something useful.

The same applies to countries. A nation can possess enormous technological capability and still struggle to convert that capability into broad, sustainable economic value. Another may control less of the underlying infrastructure but become exceptionally good at applying it in areas where it already possesses expertise, credibility and intellectual advantage. Neither route is risk-free, but they are different strategies.

Perhaps there is more than one AI race

The Ebbon Group has not spent twenty-five years becoming an AI company. It has spent twenty-five years becoming progressively better at recognising where technology can create value within a market it understands. AI is the latest capability available to it, and that distinction may matter far beyond automotive.

The AI scale race is real. The largest models, the largest data centres and the companies financing them will shape a significant part of what comes next. But scale has more than one form: there is scale of infrastructure, scale of intelligence, scale of distribution, and scale of value. Britain needs enough of the first two to remain capable. Its greater opportunity may lie in becoming exceptional at the last.

Perhaps the question is not whether Britain can produce as much AI as
the United States or China.
Perhaps it is whether we can become better at deciding what all that
intelligence is actually for.