Why the next AI revolution won't live entirely in the cloud

Paying for occasional interactions with a frontier model is one thing. Operating agents that make millions of interactions across a business is something quite different. CTI CEO Chris Burgess examines what this means for local AI in your organisation and how UK PLC competes.

For the past few years, we have largely treated artificial intelligence as somewhere we go.

We open ChatGPT, use Copilot, call an API or upload a document to a service sitting somewhere in the cloud.

Intelligence has become something we access rather than something we own.

That model has been extraordinarily successful. It has put capabilities that would once have required specialist teams and enormous computing resources into the hands of almost anybody with a browser. It has also accelerated the adoption of AI at a pace few technologies have achieved before.

But I don't believe it represents the end state.

AI is getting bigger and smaller at the same time

We are approaching another important shift in computing. AI is beginning to move from being predominantly a service that organisations consume towards becoming infrastructure that organisations can increasingly operate and control themselves.

Some of the world's most powerful models will continue to run in enormous data centres, but at the same time another layer of intelligence is moving much closer to the businesses and people actually using it.

That might mean AI running within sovereign UK infrastructure. It might sit inside an organisation's own private cloud or data centre. Increasingly, it can also mean substantial AI capability running on hardware inside an office and, quite literally, on a desk.

The interesting thing is that two apparently contradictory trends are happening at once. At the frontier, AI is becoming enormously more computationally intensive. The UK Government's AI Opportunities Action Plan argues that access to compute is becoming a pillar of economic security and has committed to increasing the capacity of the UK's AI Research Resource by at least twenty-fold by 2030. The Government has subsequently committed £1 billion towards that expansion.

At the same time, useful AI is becoming dramatically smaller and more efficient.

Stanford's 2025 AI Index provides a striking illustration. In 2022, the smallest model able to exceed 60 per cent on the widely used MMLU benchmark contained 540 billion parameters. By 2024, Microsoft's Phi-3-mini crossed the same benchmark threshold with just 3.8 billion parameters. That is a 142-fold reduction in model size in only two years. It does not mean those models are equivalent in every respect, but it demonstrates just how quickly capable intelligence is becoming more compact.

The hardware is following the same trajectory. NVIDIA's DGX Spark, a computer compact enough to sit on a desk, has 128GB of unified memory and can run inference locally on models containing up to 200 billion parameters. It can also fine-tune models of up to 70 billion parameters. This isn't a hypothetical future product; systems of this kind are already shipping.

Intelligence is therefore simultaneously becoming enormously bigger at the frontier and dramatically smaller, cheaper and more deployable at the edge.

That has some profound implications for businesses.

From personal productivity to organisational orchestration

The first phase of generative AI adoption has largely been about personal productivity. Employees have used AI to write emails, produce content, summarise meetings, generate software, analyse documents and help answer questions. The next phase will go much further. AI will increasingly operate inside business processes rather than simply alongside them.

Agents will interact with corporate systems. They will interrogate organisational knowledge, analyse customers, monitor processes, generate reports, write and test software, trigger workflows and potentially operate continuously rather than waiting for a person to type a prompt.

Once AI begins to operate in that way, where that intelligence resides stops being a technical detail. It becomes a commercial decision.

A business running hundreds of AI agents will need to think about where its information is being processed, what those agents are permitted to see, how their activities are audited, what happens if a provider changes its pricing or terms and whether the organisation can move to a different model in the future.

There is also an economic question. Paying for occasional interactions with a frontier model is one thing. Operating agents that potentially make millions of interactions across a business is something quite different. In some circumstances the economics of owning or controlling more of that compute may become increasingly attractive.

Sovereign AI and attractive economics

This is where I think the conversation around sovereign AI becomes much more interesting than simply asking where a data centre is located.

At a national level, the UK Government clearly regards sovereign AI capability as strategically important. Its Sovereign AI Fund is backed by up to £500 million and has been given a mandate to support UK companies in strategically important parts of the AI value chain. Government language around the programme has been unusually direct: Britain should become an AI “maker, not just an AI taker”.

The UK's National Security Strategy goes further. It acknowledges that complete independence in frontier technologies such as AI is unlikely to be realistic, but argues that Britain should maintain a baseline sovereign capability and sufficient strategic depth within the wider international ecosystem. That is an important distinction. Sovereignty does not have to mean isolation.

I think the same logic applies to individual organisations.

A British company does not need to build its own foundation model in order to have greater sovereignty over its AI. It does not need to disconnect itself from American technology companies, abandon the cloud or attempt to recreate every component of the AI stack internally.

It does, however, need to understand who ultimately controls the intelligence becoming embedded in its business.

Can it determine where sensitive information is processed? Can it choose which models are appropriate for different workloads? Can it protect its intellectual property? Can it audit what agents are doing? Can it change models without rebuilding its entire architecture? Can it continue to operate if a supplier, commercial arrangement or geopolitical situation changes?

Those are questions of sovereignty too.

This is why I think it is a mistake to reduce Sovereign AI entirely to data residency. Keeping information within the UK can clearly matter, particularly for sensitive or regulated workloads, but sovereignty is ultimately about control, choice and resilience.

Future architecture is unlikely to be simply cloud or local

Cloud or local? AI architecture will be both.

There will be occasions where an organisation wants access to the most capable frontier model available and is perfectly comfortable calling that model through a global cloud provider.

There will be other tasks where a smaller specialist model running privately is more than capable of doing the job. Some sensitive workloads might remain inside an organisation's infrastructure. Others might sit on sovereign UK infrastructure. Increasingly, some AI workloads may run on powerful machines inside an office.

The goal should not be to put walls around artificial intelligence. It should be to retain doors.

Businesses should be able to use extraordinary global AI innovation without unnecessarily surrendering control of their information or locking themselves permanently into a single provider, model or architecture.

There is a useful parallel here with cyber security.

Twenty years ago, cyber security was still widely regarded as an IT problem. Firewalls, antivirus software and password policies were matters for technology teams. Today, no serious board treats cyber risk that way. Cyber security is part of corporate governance because a cyber incident can affect revenue, reputation, customers, regulation and the organisation's ability to continue operating.

AI is beginning a similar journey.

The original question was simply what AI could do. The next questions concern what AI can access, what it can act upon, where it operates and who is accountable for it.

Understanding the intelligence supply chain

As AI becomes embedded across businesses, boards will increasingly need to understand their intelligence supply chain in much the same way as they understand their financial, technology and cyber dependencies. The governance issues identified in the original sovereign AI argument - where data is hosted, who owns it and how its use is monitored, are becoming business decisions rather than purely technical ones.

Paradoxically, I think organisations that establish those controls will ultimately be able to use more AI rather than less.

When businesses don't know what employees can safely put into AI systems, the understandable reaction is restriction. When they have clear governance, appropriate technology and trusted environments, people can experiment much more confidently.

But there is another part of this transition that may prove even more significant.

The most useful AI for a company may not always be the biggest model.

A frontier model has been designed to understand an extraordinary breadth of human knowledge. An organisational AI has a rather different purpose. It needs to understand the business.

Its products, customers, systems, terminology, processes, documents and accumulated knowledge matter enormously. It needs context. It also needs permissions. It should understand not simply what the organisation knows, but what an individual employee is allowed to know and which systems they are entitled to use.

When AI is considered in those terms, the model itself becomes only one component of a much larger system.

A smaller model with the right organisational knowledge, access to the right tools and a clearly defined task can be far more useful in a particular business process than an enormous general-purpose model with no context about the organisation.

That is a fundamentally different way of thinking about enterprise AI.

Instead of simply buying AI tools, businesses begin to build intelligence infrastructure.

Imagine joining a company in the future and being given access not only to email, Microsoft 365 and a CRM, but to the organisation's AI.

It understands the company's products and services. It understands years of corporate knowledge. It knows the systems you are authorised to use and the information you are permitted to access. It can find information, answer questions and increasingly perform tasks across those systems on your behalf.

At that point it becomes something quite different from giving every employee a subscription to the same public chatbot.

Parts of that intelligence might use frontier models in the cloud. Parts may run on UK-hosted infrastructure. Parts could operate inside the company's own environment. And increasingly some of the compute could sit on specialised AI hardware physically inside the business.

This is the direction that has informed our work on Croft at CTI Digital. Croft has been designed around the idea that organisations should be able to access advanced AI, organisational knowledge and agentic workflows while maintaining greater control over where their data, documents and AI activity reside. Depending upon the organisation, that can mean UK-hosted infrastructure, infrastructure controlled by the customer or systems deployed on-premise.

But Croft is one implementation of a much bigger change.

The really interesting question is what happens when the same principle extends from organisations to individuals.

Personal computers put computing power on our desks. The internet gave us access to information. Smartphones put computing and permanent connectivity into our pockets.

The next stage may be personal intelligence.

Your own, personal intelligence

Picture an AI that understands your documents, preferences, previous work, communications and context. An intelligence that doesn't have to rediscover who you are every time you open another application. It might use enormous cloud models when it requires their capabilities, smaller local models for everyday activity and private memory that stays under your control.

That idea is becoming technologically plausible much faster than many people realise.

And I think it is particularly relevant to Britain.

We are unlikely to win an AI race by trying to spend more money building ever larger general-purpose models than the United States or China. Nor do I believe success requires a British version of every American technology product.

Our opportunity can be broader than that.

The UK can become extremely good at applying intelligence: specialist models, AI infrastructure, orchestration, governance, security and systems that actually solve business problems. Government policy already recognises the need for domestic compute and strategically important AI capability alongside access to international infrastructure.

That hybrid approach makes sense for British businesses too.

Use the world's best intelligence when you need it. Build specialist capability where it creates advantage. Keep sensitive information under appropriate control. Avoid unnecessary dependency. Make the intelligence layer as replaceable as possible, because today's leading model will not necessarily be tomorrow's.

Above all, businesses should start considering AI not simply as another category of software but as part of their future operational infrastructure.

For the last few years the question has been: Which AI tools should we use?

The question increasingly becomes: What intelligence does our organisation need, where should it live and how much of it should we control?

That is the real significance of Sovereign AI.

It is not an argument against the cloud. It is not an argument against frontier models or global technology companies. And it is certainly not an argument for cutting ourselves off from the extraordinary pace of AI innovation taking place around the world.

It is an argument for choice.

The future of AI will include hyperscale data centres and enormous frontier models. It will include national sovereign infrastructure and global cloud platforms.

But it will not exist exclusively inside them.

Some intelligence will run nationally. Some will belong inside individual organisations. Increasingly, some will become personal.

And sooner than many businesses expect, when someone asks where their AI actually lives, part of the answer may be surprisingly simple:

It's that box sitting on the desk.

Find out more about CTI Digital's private AI and LLM services.