The Register reports that Bracken, who founded the Government Digital Service and served as the UK’s first Government Chief Data Officer, sees dependency accumulating through individually reasonable technology decisions. City A.M. also reported his warning on September 21, describing his concern that organisations are surrendering control of critical systems as they adopt external technology and AI services. Those reports establish the public warning, rather than independently demonstrating the extent of the problem.
There is a concrete policy backdrop: the UK government announced a £500 million Sovereign AI programme in April to support domestic AI companies, including investment and access to publicly backed computing resources. But Bracken’s argument concerns a different measure of success. Building British AI capacity and preserving a particular institution’s freedom to choose are related objectives; achieving the first does not automatically establish the second.
Mike Bracken puts supplier choice at the centre of sovereign AI
“Institutions rarely lose sovereignty in a crisis. They lose it one reasonable decision at a time,” Bracken said, according to The Register. His warning is about cumulative commitments: an organisation makes decisions that work for its immediate needs, then discovers later that changing direction has become prohibitively difficult.
In this context, digital sovereignty means the capacity to make and implement decisions about an organisation’s technology. Public Digital, the consultancy Bracken co-founded after leaving government, defines it as the “agency and capacity” to make informed choices about a digital future. Its published position explicitly says that local data centres, bringing supply chains back within national borders and additional regulation do not by themselves establish that capacity.
That definition separates two questions often bundled into the sovereign-AI debate. Where infrastructure sits is one question. Whether its customer can choose another approach—and has the people, resources and practical ability to carry out that choice—is another. A domestic location does not answer both.
Bracken develops the argument with Public Digital colleagues in Digital Sovereignty: The Power to Decide. Public Digital lists the book for sale in autumn 2026 and describes it as drawing on the authors’ experience of technology contracts, supplier relationships and institutional decision-making. This is a practitioner’s argument associated with a forthcoming book, not a published measurement of how many British organisations are locked into AI suppliers.
That boundary matters to the conclusion. The warning does not establish that the UK’s sovereign-AI programme has failed, or that any named platform prevents customers from leaving. It offers a way to judge technology decisions before their consequences become difficult to reverse.
The £500 million Sovereign AI programme addresses national capacity
The Department for Science, Innovation and Technology’s April 16 announcement described Sovereign AI as a £500 million effort to back homegrown AI companies. Its stated ambition was to make Britain an AI “maker, not just an AI taker,” retaining expertise, decision-making and economic value in the country while reducing reliance on a small number of foreign technology companies.
The programme goes beyond constructing data centres. The launch package included direct investment, access to supercomputing, research and development support, assistance attracting talent, and help with data access and early procurement opportunities. That breadth is important: the government’s published approach already recognises that useful AI capability requires more than physical infrastructure.
At launch, the government named infrastructure startup Callosum as its first equity investment. Six other companies—Prima Mente, Cosine, Cursive, Doubleword, Twig Bio and Odyssey—were allocated access to the AI Research Resource supercomputer network. The announcement offered fully funded computing access of up to one million GPU hours per supported startup, giving recipients time on specialised processors used for AI workloads.
Those are commitments to build and support supply. They are not evidence that a future customer of a supported company will find its services interchangeable with competitors. Equally, the allocations do not demonstrate that customers will become locked in. The announcement establishes what the government is backing, while the customer’s freedom to move remains a separate question.
The useful comparison is therefore between the policy’s national objective and Bracken’s institutional test:
| Decision | What the evidence establishes | What it does not establish |
|---|---|---|
| Back British AI companies. | The government intends to strengthen domestic capability and retain more economic value in Britain. | That an individual customer can readily replace a supported supplier. |
| Provide publicly backed computing access. | Selected companies receive resources to develop and test AI technology. | That services developed with those resources will be portable between platforms. |
| Keep infrastructure within national borders. | Location forms part of the sovereign-AI discussion. | That the customer possesses the capability to change its technology or strategy. |
These distinctions allow the government’s investment and Bracken’s warning to be considered together without manufacturing a dispute. Expanding the available supply of AI technology can create opportunities. Customers still have to examine what accepting one of those opportunities commits them to.
AI lock-in becomes an operational question after adoption
According to The Register, Bracken’s concern is that switching becomes harder once externally controlled AI models and platforms are embedded across an organisation’s infrastructure and workflows. The relevant unit of analysis is therefore the service an institution depends on, rather than simply the model it initially selects.
The practical inference is straightforward: a decision-maker should distinguish between choosing an AI supplier and retaining the ability to replace the resulting service. Those are different stages of the technology lifecycle. An attractive initial choice does not establish how difficult a later transition would be.
Consider an architectural review as a thought exercise, rather than a claim about any particular product. If a team says it can change its AI supplier, the useful follow-up is to ask what “change” includes. Does the assessment cover the working service and the people who operate it, or only the availability of another model? Bracken’s definition requires the organisation to be able to act on its decision, so the answer has to extend beyond identifying an alternative.
Public Digital’s book material supports this broader institutional emphasis. It describes dependencies involving systems, skills and supply chains, some chosen deliberately and others inherited without examination. It also warns against treating ownership or bringing systems in-house as a universal solution. Taking responsibility for more technology does not, by itself, demonstrate that an organisation can manage it effectively.
This is why the warning is relevant to enterprise IT without needing a Microsoft-specific allegation. The material identifies no Azure, Copilot or Windows defect, contractual change or affected version. The decision facing IT leaders is a procurement and operating-model question: whether an AI commitment preserves the choices the organisation considers important.
Nor does the argument require avoiding foreign suppliers. Public Digital explicitly rejects isolation as a substitute for sovereignty. The defensible objective is to understand and deliberately accept dependencies, while retaining sufficient capability to change the ones that could constrain important future decisions.
AI procurement should make freedom to change explicit
Organisations considering a major AI commitment should use Bracken’s warning to sharpen their assessment of future choices, rather than treat “sovereign” as a sufficient purchasing criterion. Neither the reporting nor the official programme announcement supports an urgent migration, a supplier blacklist or a universal technical remedy.
A practical application of the argument is to ask for evidence at the next significant procurement, renewal or expansion decision. These are analytical questions derived from Bracken’s definition, not a documented migration procedure or an audit of any vendor.
- Ask the supplier to explain precisely what its use of “sovereign AI” promises, rather than assuming the term guarantees freedom to switch.
- Separate the requirement for domestic infrastructure from the requirement to change supplier, technology or strategy, and assess each on its own terms.
- Identify which dependencies were consciously accepted and which were inherited, reflecting Public Digital’s distinction between deliberate and unexamined commitments.
- Assess whether the organisation has the skills and resources to act on an alternative choice, rather than counting the existence of another supplier as sufficient.
- Treat government investment or computing support as evidence of support for a company’s development, not as proof that its customers can migrate easily.
This approach also accommodates a decision to stay with an existing platform. Bracken’s test does not demand constant movement or the elimination of every dependency. It asks whether the institution can make a meaningful choice and carry it out. A deliberate commitment, understood by the people accountable for it, is different from discovering that earlier decisions have removed the available alternatives.
Britain’s sovereign-AI investment is intended to increase the country’s capacity to build and benefit from AI. Bracken’s warning adds a customer-side measure that spending announcements cannot settle: whether institutions retain the ability to decide what happens next. For enterprise IT, the next contract or expansion proposal is the concrete place to apply that test—before a useful service becomes a commitment the organisation can no longer realistically reconsider.