Sovereign by Whose Design? · Jonathan Gompels

A critical reflection on the UK's pursuit of AI sovereignty through Isambard-AI and the Sovereign AI Unit, read through John Gaventa's power cube.

Idea, research and reading by Jono Gompels. Compiled and written by Claude, Anthropic's AI.

Abstract. In October 2026 the Chancellor added £60 million to the UK's Sovereign AI Fund. It's the latest step in a push, centred on the £225 million Isambard-AI supercomputer in Bristol, to make Britain "an AI maker, not an AI taker". This essay asks what that push is for, what AI sovereignty would actually take, and what kind of power is being built along the way. Using John Gaventa's (2006) power cube, I map the spaces, levels and forms of power around the UK's approach. I argue that the UK holds a double position. At home it is the power-holder: it decides behind closed doors and then invites a select few in. Abroad it is a dependent invitee, relying on American chips, firms and base models. The strategy secures visible power, meaning ownership of machines, while leaving hidden and invisible power mostly untouched. I finish by asking what the cube itself misses: material infrastructure, corporate power that ignores the line between levels, and the people this sovereignty is supposedly for.

Contents

Introduction
Why the UK wants AI sovereignty
What AI sovereignty would take
Gaventa's power cube
Mapping the powers at play
Weaknesses in the UK's approach
What the cube fails to show
Conclusion

1. Introduction

On 28 October 2026, in his first Budget as Chancellor, John Healey found another £60 million for the national Sovereign AI Fund (Charles and Gherman, 2026). Next to the rest of the Budget it's a small sum. But it continues a project that has run through British technology policy since the AI Opportunities Action Plan of January 2025: the attempt to make the UK "an AI maker, not an AI taker" (DSIT, 2025a). The most visible piece of that project is a building on the edge of Bristol. Isambard-AI, fully switched on in July 2025, holds 5,448 NVIDIA Grace Hopper superchips paid for with £225 million of public money, and is the UK's most powerful AI supercomputer (University of Bristol, 2025; insideHPC, 2025).

"Sovereignty" is a word about power. It names who gets to decide. So it's strange how rarely the sovereign AI debate asks the questions that people who study power would ask first. Decided by whom? In what kind of room? At what level? With what taken for granted? This essay uses John Gaventa's (2006) power cube to ask them. Gaventa built the cube to help citizens and development practitioners find where power sits and where it might be challenged. It isn't an obvious tool for a national technology strategy, and part of my argument is about what happens when you stretch it that far.

The essay does four things. It sets out why the UK wants AI sovereignty and what it hopes to get from it. Drawing on Charles and Gherman (2026), it asks what AI sovereignty would actually take. It then introduces the power cube and uses its three dimensions to identify the powers at play around Isambard-AI and the Sovereign AI Unit. Finally, it reflects critically, first on the weaknesses of the UK's approach, then on what the cube itself fails to show.

A word on where I'm coming from. I'm a geography student, not a computer scientist, and my dissertation looks at how University of Reading students use AI, and why some choose not to. That work has made me wary of national stories about AI that never reach the scale of one person deciding whether to open a chatbot. That wariness runs through what follows.

2. Why the UK wants AI sovereignty

The government's case rests on three motives: resilience, influence and growth. A fourth, wider reason explains why all three feel urgent now.

2.1 Resilience: not being switched off

Almost every country relies on foreign AI to some degree, and the UK relies heavily on a handful of American companies. Charles and Gherman (2026) report that around three-fifths of the foreign base models used in the government-backed sovereign AI projects tracked by the Center for a New American Security (CNAS) come from the US. The chips that train and run those models come from an even smaller set of suppliers, and trade restrictions or geopolitical tension can cut off access to them (CSET, cited in Charles and Gherman, 2026). If an essential public service runs on a model hosted in another jurisdiction, under another government's export rules, then keeping that service running isn't fully the UK's decision.

The Action Plan frames this as resilience. Publicly owned or allocated compute, it says, "will enable the UK to quickly and independently allocate compute to national priorities", including keeping AI available "for critical services in times of market disruption" (DSIT, 2025a, s.1.1). The first motive, then, is insurance against being switched off.

2.2 Influence: whose values?

The second motive goes beyond defence. The Action Plan argues that "the people and countries who make decisions about how these systems operate and what values they reflect" will "have huge influence over our lives" (DSIT, 2025a, s.3). The UK, it says, should build "true national champions at critical layers of the AI stack", to win both economic upside and "influence on future AI's values, safety and governance" (DSIT, 2025a). Behind this is a real fear: that the defining technology of the century gets made in California and Shenzhen, and Britain becomes a market for it. In the plan's own words, the alternative to having a stake is seeing key decisions "made entirely outside our borders" (DSIT, 2025a).

2.3 Growth: an industrial strategy

Third, sovereignty is sold as industrial policy. The Sovereign AI Fund, launched in April 2026 with £500 million and based in the Department for Science, Innovation and Technology (DSIT), works much like a government-backed venture capital fund. It gives promising UK AI firms money, time on supercomputers such as Isambard-AI, research support and openings to work with government (Charles and Gherman, 2026). The Action Plan is open that this needs "a more activist approach", modelled on Japan's MITI rather than "the 'invisible hand'" of the market (DSIT, 2025a, s.3). That's a striking phrase to find in a document about power, and I'll come back to it.

2.4 Why now

There's also a wider reason sovereignty has become urgent. Jennings (2026) is sceptical of AI doom talk, but he still takes one risk seriously: powerful AI handing new capabilities to governments, corporations or other groups, with access spread unequally. That is a concentration of power. If AI does concentrate power, being on the dependent end of it is a strategic problem as well as an economic one. The global response shows how widely this is felt. CNAS recorded one government-backed sovereign AI project in January 2023, and 184 across 67 countries by mid-2026 (Charles and Gherman, 2026).

So the UK wants three things from sovereignty: resilience (not being switched off), influence (a hand in the values built into AI) and growth (a share of the industry). These goals don't always point the same way. Growth, for instance, is fastest if you invite in whoever has the most capital, and that usually means the very firms you are meant to be less dependent on. That tension sits at the heart of what follows.

3. What AI sovereignty would take

3.1 The stack

AI isn't one thing a country can own. Charles and Gherman (2026) list what it depends on: chips, data, electricity, skilled workers, internet connections and cybersecurity. A country may control some of these and depend on foreign suppliers for the rest. Together they make up the "AI stack", the layers of hardware, infrastructure, models, data and applications needed to build and run AI, each with its own owners. A decade before the current rush, Bratton (2015) argued that planetary-scale computing produces its own layered geography of sovereignty, which cuts across states rather than sitting neatly inside them. Sovereign AI policy is, in effect, an attempt to pull some of those layers back inside the border.

3.2 Four meanings of sovereignty

Charles and Gherman (2026) suggest that four ideas usually get bundled together under the label:

Ownership: who owns the chips, data centres and models?
Jurisdiction: whose laws apply, and where is the data held?
Capability: can the country build, adapt and run AI itself?
Resilience: can essential services carry on if a supplier withdraws access?

Much of the argument about sovereign AI, they suggest, is really an argument about which of these matters most. The academic debate on digital sovereignty shows the same pattern. Pohle and Thiel (2020) trace how the idea has stretched from a defensive claim about state control over networks into a much broader and vaguer language of self-determination, used by actors with very different aims. Couture and Toupin (2019) find that states, companies, social movements and individuals all invoke "sovereignty" over the digital, and mean quite different things by it. The word is doing a lot of work, and we should ask whose work.

3.3 A dial, not a switch

On a strict test almost every country fails. If sovereignty means owning every layer of the stack, Charles and Gherman (2026) point out, hardly anyone qualifies. Researchers at Stanford suggest it is better understood as managing dependencies than eliminating them: backing domestic firms, using open-source models and spreading reliance across several suppliers (cited in Charles and Gherman, 2026). Charles and Gherman conclude that sovereignty looks "less like a switch and more like a dial". The real test is whether a country has meaningful choices. Can it switch suppliers without breaking essential services? Can it enforce its own laws? Can it adapt AI to local languages, priorities and needs?

Other countries have turned the dial in different directions. India has gone for scale: its IndiaAI Mission has deployed around 38,000 GPUs, alongside BharatGen, a multilingual model, and AIKosh, a national platform for anonymised training data. Switzerland has chosen control over raw power. Its Apertus model is open about its training data, code and development, and the canton of Ticino uses it to translate sensitive government documents without relying on a commercial provider (Charles and Gherman, 2026).

3.4 Where Isambard-AI fits

According to CNAS, only a handful of jurisdictions (India, South Korea, the UK and Taiwan) run working programmes across compute, models and data (cited in Charles and Gherman, 2026). In the UK, Isambard-AI is the compute layer. It is run by the Bristol Centre for Supercomputing and housed at the National Composites Centre. HPE built it around 5,448 NVIDIA GH200 superchips, and it delivers roughly 21 exaflops of AI performance (University of Bristol, 2025; insideHPC, 2025). With Dawn in Cambridge it makes up the AI Research Resource (AIRR), which the government's Compute Roadmap commits to expanding twentyfold by 2030, alongside up to £750 million for a new national supercomputer in Edinburgh (DSIT, 2025b). Researchers, SMEs and start-ups apply for free time on it.

Against the four meanings, Isambard-AI scores well on ownership, since the state paid for the machine, and on jurisdiction, since it sits under UK law in Bristol. It also adds real capability. On resilience it is weaker, because its most critical parts, and the software ecosystem built around them, come from one American company. The Action Plan admits the wider limit itself: sovereign compute "will almost certainly be the smallest component of the UK's overall compute portfolio" (DSIT, 2025a, s.1.1). Most British AI will still run on someone else's servers.

So AI sovereignty would take control over enough layers of the stack, and enough choice over the rest, to keep deciding for yourself when circumstances change. That raises the question the rest of this essay is about: deciding for whom, and decided by whom?

4. Gaventa's power cube

John Gaventa developed the power cube at the Institute of Development Studies in response to a problem in development practice. By the 2000s "participation" was everywhere. Even the World Bank and the IMF talked about "partnership" and "shared ownership". Gaventa (2006, p. 23) asked whether this new terrain marked "a real shift in power", or simply made the status quo look more legitimate. His route into the question was a mining valley in Appalachia in the 1970s, where enormous injustice met almost no visible protest; over time, the voices "had been silenced altogether" (p. 24). That became Power and Powerlessness (Gaventa, 1980), an application of Steven Lukes' (1974) three dimensions of power.

The cube links three dimensions, each with three parts (Figure 1).

Spaces describe how an arena for decision-making was created.

Closed spaces are where decisions are made behind closed doors, "without any pretence of broadening the boundaries for inclusion" (p. 26). Bureaucrats, experts or elected representatives decide, then provide for "the people".
Invited spaces are those that people are invited into by some kind of authority: a government, a supranational agency or an NGO (Cornwall, 2002, cited in Gaventa, 2006, p. 26).
Claimed or created spaces are made by less powerful actors, either taken from the power holders or built on their own, from social movements to the places where people gather to debate and resist (p. 27).

The key point about spaces is that "those who create it are more likely to have power within it" (p. 27).

Levels describe where power sits: local, national or global. Gaventa insists these aren't separate spheres. They are "increasingly interrelated" (p. 28), and globalisation keeps shifting where power lies.

Forms describe how visible power is. They come from Lukes, as reworked by VeneKlasen and Miller (2002):

Visible power: the formal rules, institutions and procedures of decision-making.
Hidden power: controlling who gets to the decision-making table and what makes it onto the agenda.
Invisible power: shaping meaning and what seems normal, so that some issues are kept even "from the minds and consciousness" of the people they affect. Gaventa calls it "probably the most insidious" form (Box 1, p. 29).

Figure 1. The power cube, redrawn after Gaventa (2006, p. 25). The front face is the visible slice of power. The two highlighted cells mark where the UK's approach to AI sovereignty mostly operates: national-level, visible power, exercised through closed and invited spaces.

The dimensions turn like a Rubik's cube. You can start from any side, but moving one face moves the others (p. 26). Transformative change, Gaventa argues, happens in "rare moments" when actors manage to work across all three dimensions at once (p. 30). And the cube is "not a checklist" (p. 31). It is a device for reflection, including reflection on your own position. Donors who placed themselves in the cube found they had become "part of the power equation" (p. 31).

Two features make the cube useful for thinking about sovereign AI. First, it was built for exactly the situation Gaventa describes on its first page: powerful actors adopting the language of participation and ownership in ways that hide "inequalities of resources and power" (p. 23). "Sovereignty" has become that kind of word. Second, the cube is relational. It doesn't ask who has power in general. It asks who has power in this space, at this level, in this form. That matters for a country that is powerful in some rooms and dependent in others.

5. Mapping the powers at play

5.1 Spaces: who built the room?

Closed spaces. The core decisions about UK sovereign AI were made in closed spaces. The Action Plan was written by a single commissioned author, the tech entrepreneur Matt Clifford, at the request of the Secretary of State (DSIT, 2025a). Budget allocations, the procurement of Isambard-AI from HPE and NVIDIA, and the Sovereign AI Unit's investment choices are all made by ministers, officials and appointed experts. These are Gaventa's "provided" spaces, where elites decide and deliver. That isn't unusual, or necessarily illegitimate; much economic and security policy works this way. But it matters that the basic definition of sovereignty for the UK (compute-first, led by national champions, aimed at growth) was settled in rooms the public never entered.

The most consequential closed spaces aren't in Britain at all. US export controls on advanced chips, NVIDIA's decisions about who gets its next generation of hardware, and the boardrooms of the frontier labs whose base models other countries adapt all shape what the UK can do. Britain has no seat in any of them.

Invited spaces. Around those closed cores the UK has built a ring of invited spaces. The AIRR is the clearest example. Researchers, SMEs and start-ups are invited to apply for compute on terms set by government, with allocation steered by "mission-focused" programme directors (DSIT, 2025a, rec. 3). The Sovereign AI Unit invites selected firms in as investees. Consultations on AI and copyright, or on AI Growth Zones, invite organised interests to respond. Each of these is real, and each opens something up. But Gaventa's warning applies. Whoever creates an invited space sets its terms, and a new participatory design without countervailing power "might simply be captured by the already empowered elite" (p. 27). In practice, free compute goes to those already able to write winning proposals: university labs and well-connected start-ups. A promotional video for Isambard-AI, embedded in Charles and Gherman's (2026) article, opens by declaring that the machine is "for everyone". Its access process suggests it is for those already inside the research system.

The UK is also an invitee. The Tech Prosperity Deal, signed during the US state visit in September 2025, brought around £31 billion of commitments from American firms. These included Microsoft, Google and NVIDIA, and a "Stargate UK" partnership between OpenAI, Nscale and NVIDIA in the North East (Tom's Hardware, 2025; Computer Weekly, 2025). The government presented the deal as strengthening British sovereignty. Critics argued that it tied the UK more tightly to a US strategy for AI dominance (Tech Policy Press, 2025). In cube terms, the UK accepted a place in a space created by others: American capital chose to build in Britain, on its own terms. As Gaventa would predict, those who created the space hold the power within it.

Claimed and created spaces. Claimed spaces are thin, but they exist. The clearest came from the creative industries. In February 2025 news publishers launched the Make It Fair campaign, soon backed by musicians and writers, against government plans to let AI firms train on copyrighted work unless creators opted out. In Parliament, Baroness Kidron's amendments to the Data (Use and Access) Bill, which would have made AI companies disclose the copyrighted works they used, were backed repeatedly by the House of Lords. The government removed them, and the Act passed without them in June 2025 (Music Business Worldwide, 2025; The Bookseller, 2025). This was a claimed space in Gaventa's sense: actors outside government forcing an issue onto an agenda that had left it out. It lost the vote, but it made the question of whose work trains "British" AI impossible to ignore. Open-source communities are another kind of created space, and Switzerland's Apertus shows what happens when a state treats openness, rather than raw power, as the point (Charles and Gherman, 2026).

What is largely missing is any claimed space for the public as users of AI, or for the communities that host its infrastructure.

5.2 Levels: where does the power sit?

Global. This is where AI power is most concentrated. The race between the US and China sets the pace (Jennings, 2026). A handful of firms control the chips, the most critical layer of the stack. And three-fifths of the foreign base models in the world's sovereign AI projects are American (Charles and Gherman, 2026). Lehdonvirta, Wú and Hawkins (2024) map public cloud GPU capacity and find it sharply uneven. They sort the world into a "Compute North" that hosts the most advanced chips, a "Compute South" with older ones and "Compute Deserts" with none, and they argue that hosting compute is what gives a state leverage to govern AI. The UK is firmly in the Compute North, but mostly as a host of American hardware rather than a maker of its own.

National. The sovereign AI project is, by definition, an attempt to pull power back to the national level and make the British state the actor that decides. Isambard-AI, the AIRR, the Sovereign AI Unit and the Compute Roadmap are all national instruments. Here the state plays the mediating role Gaventa describes: it negotiates with global firms above and allocates resources to institutions below.

Local. The local level is where the machines physically land, and it barely features in the policy. Isambard-AI alone is a 5-megawatt facility (DatacenterDynamics, 2025), and the Compute Roadmap aims for at least one AI Growth Zone above a gigawatt by 2030 (DSIT, 2025b). Growth Zones such as Culham in Oxfordshire and sites in the North East (Prime Minister's Office, 2025; Computer Weekly, 2025) involve decisions about land, grid capacity, water and jobs in particular places. Those decisions are largely made nationally, with private partners. Meanwhile, the benefits of sovereign AI are meant to reach local public services, classrooms and workplaces, but there is no visible way for those places to feed back up into decisions. Gaventa's point that the levels are interrelated (p. 28) cuts both ways. National strategy is shaped by global firms from above, but it also lands on, and depends on, local communities that never had a say.

5.3 Forms: visible, hidden and invisible

Visible power. The visible face of UK AI sovereignty is impressive and easy to count: a £225 million machine, a £500 million fund, an extra £60 million, a twentyfold expansion target. These are formal decisions by identifiable authorities, announced in Budgets and press releases, and their visibility is part of the point. A supercomputer is a policy you can photograph.

Hidden power. Hidden power works by deciding who gets to the table and what gets onto the agenda (Gaventa, 2006, p. 29). Three examples stand out.

Defining the problem. The agenda was set as a question about compute and companies. Other questions largely fell off it: whose data trains the models, who does the labelling, who gets access, and who bears the energy and water costs. Once "sovereignty" means more GPUs and national champions, those questions start to look off-topic.
Vendor dependence. The Action Plan itself says the public compute ecosystem should host "a range of hardware providers to avoid vendor lock-in" (DSIT, 2025a, rec. 2). Yet the flagship sovereign machine is built entirely around one US supplier's chips, and with them that supplier's proprietary CUDA software ecosystem, which shapes what researchers can easily build. Every model trained on Isambard-AI is trained on NVIDIA's terms. This kind of agenda-setting never appears as a line in a Budget.
A corporate seat at the table. Deals like the Tech Prosperity Deal give large US firms a privileged position in shaping UK infrastructure. Whittaker (2021) shows how academic AI research's dependence on big tech's compute and data lets those firms shape the research agenda itself. Sovereign compute was partly meant to answer this, but at the scale the UK is building, it supplements corporate compute rather than replacing it.

Gaventa's opening observation fits almost exactly. Powerful actors using words like "partnership" and "shared ownership" invite engagement on a "level playing field" while hiding the inequalities behind it (p. 23). Swap in "sovereign AI", a label that, as the Stanford researchers note, can be attached to commercial products that still rely on foreign providers (cited in Charles and Gherman, 2026).

Invisible power. The deepest layer is what everyone involved takes for granted. Three assumptions do most of the work.

That AI is an inevitable race. The debate is framed as a contest between the US and China, in which Britain chooses whether to run, not whether the race is worth running or what it's for. Jennings (2026) notes that dramatic AI scenarios grab attention because they're hard to rule out. The race framing works the same way, and makes speed feel like a necessity.
That sovereignty means infrastructure. The idea that being sovereign means owning machines and backing firms is now so widespread (184 projects in 67 countries) that it reads as common sense.
That citizens are users. "Maker, not taker" casts the nation as a producer and the public as consumers of whatever gets made. Siemens (2023) argues that AI matters less as a stand-alone intelligence than as something "embedded in and part of our knowledge network". It becomes part of how people think, write and decide. If that is right, the values built into models shape what seems normal for millions of people. That is invisible power in Gaventa's sense, and it is exactly what the Action Plan says it wants to influence: "what values they reflect". Yet the plan locates that influence in national champions, not in the people whose thinking is being shaped.

The Action Plan's own wording contains a small irony here. It rejects "the 'invisible hand'" of the market in favour of an activist state (DSIT, 2025a, s.3). But swapping the market's invisible hand for the state's visible one doesn't touch invisible power in Gaventa's sense: the shared assumptions about what AI is for, who it's for, and who gets to say.

Table 1. The UK's approach to AI sovereignty, mapped onto the power cube.

Dimension	Part	In UK AI sovereignty	Who holds power
Spaces	Closed	Budget decisions; procurement of Isambard-AI; Sovereign AI Unit investment choices. Abroad: US export controls, NVIDIA's allocation decisions, frontier-lab boardrooms	Ministers, officials, appointed experts; the US state and US firms
	Invited	AIRR compute calls; Sovereign AI Fund investees; consultations. The UK itself invited into the Tech Prosperity Deal	Whoever sets the terms: DSIT at home, US firms abroad
	Claimed	Make It Fair and the Kidron amendments; open-source communities	Creative workers, peers, developers. Marginal, and mostly defeated
Levels	Global	The US–China race; chip bottlenecks; US base models (about three-fifths of foreign models in sovereign projects)	US firms and the US state; the chip supply chain
	National	Isambard-AI, AIRR, Compute Roadmap, Sovereign AI Unit	The UK state, mediating between levels
	Local	Bristol, Culham, the North East; users, workers, students	Little formal power. They carry the costs and wait for the benefits
Forms	Visible	£225m machine, £500m fund, +£60m, a twentyfold target	Formal decision-makers, in public
	Hidden	Sovereignty framed as compute plus champions; single-vendor hardware and software; corporate seats in deals	Agenda-setters in government and industry
	Invisible	AI as an inevitable race; sovereignty as infrastructure; citizens as users	Shared common sense, largely unexamined

5.4 Turning the cube: the UK's double position

Put the three dimensions together and the UK's position turns out to be double (Figure 2). At home it is the power-holder. It builds closed spaces for strategy, creates invited spaces for researchers and firms, and exercises mostly visible power at the national level. Abroad it is the less powerful actor. It is an invitee in spaces created by American firms and the American state, dependent on hidden power it doesn't control (chip supply, export rules, software ecosystems) and working within invisible assumptions (the race, the stack) set elsewhere.

Global

Create the space: the US state (export controls), NVIDIA, frontier labs, hyperscalers.

Invited in: the UK, which buys the chips, adapts the models and hosts the investment.

↓ invites the UK in · ↑ the UK seeks a seat

National

Create the space: the UK state: Treasury, DSIT, the Sovereign AI Unit.

Invited in: university researchers, SMEs and start-ups (AIRR time, Sovereign AI Fund).

Pushing in from outside: Make It Fair, open-source communities.

↓ decisions and machines land · ↑ no route back up

Local

Host the machines and live with the results: Bristol, Culham, the North East; users, workers, students.

Their own space: none, invited or claimed.

Figure 2. The UK's double position: an invitee at the global level and the creator of closed and invited spaces at the national level, with little connecting either to the local.

Gaventa built his framework for citizens trying to open up states. Applied here, it shows a state in a similar position relative to global tech power: pushing from an invited space towards a closed one. At the same time, the UK reproduces at home the very closed-and-invited structure it is trying to escape abroad. It wants a seat at tables it didn't build, while in its own house it has built tables few people can sit at.

The double position also shows why Gaventa's Rubik's cube analogy is apt. A move that lines up one face can scramble another. Inviting US capital in delivers more compute, faster, which aligns the visible face of sovereignty: more machines on British soil. But it scrambles the hidden face by deepening dependence on the same firms. Building Isambard-AI aligns ownership and jurisdiction while leaving resilience exposed. Gaventa argues that transformative change needs alignment across all three dimensions at once (p. 30). On this reading, UK AI sovereignty is aligned on one face only: visible, national, and closed or invited.

6. Weaknesses in the UK's approach

Sovereignty of the state, not the public. The UK treats sovereignty as something held by the nation-state and its champion firms. Gaventa, drawing on Hayward (1998), offers a different yardstick. Freedom means taking part in "shaping the social limits that define what is possible", which includes the right not only to join a space but "to define and to shape that space" (Gaventa, 2006, p. 26). Measured that way, sovereign AI policy has given the public no part in setting its limits. Citizens get a stronger state, not more say.

Ownership without resilience. The flagship is owned but not independent. Isambard-AI depends on NVIDIA's hardware and software, HPE's systems and global supply chains. A machine built from one foreign supplier's parts is a strong case of the first of Charles and Gherman's four meanings, ownership, and a weak case of the fourth, resilience. The Action Plan's own warning about vendor lock-in hasn't been applied to its own flagship.

Rhetoric that runs ahead of reality. By the government's own account, sovereign compute is the smallest part of the UK's compute. Most British AI will run in American-owned data centres, many of them built under the Tech Prosperity Deal. Calling the whole package "sovereign" risks the very problem Gaventa diagnosed: powerful actors adopting a language that obscures where power actually sits (p. 23). As Charles and Gherman (2026) conclude, the label itself tells us little about how much control a country really has.

Invited spaces without countervailing power. The AIRR and the Sovereign AI Unit open doors, but mainly for people already inside the research and venture system. Without claimed spaces pushing from outside (unions, civil society, creative workers, local communities), invited spaces risk capture by "the already empowered elite" (Gaventa, 2006, p. 27). The copyright fight showed what happens when outsiders do push: the government closed the space again.

The local is missing. Growth Zones and supercomputers land in particular places, with particular costs. The benefits are assumed to trickle down to public services and workers. There are none of the "vertical links" Gaventa calls for (p. 28) that would let local communities, or ordinary users, shape national choices. From where I sit, studying how students in one university actually meet AI, this is the most striking gap. Sovereign AI policy has a great deal to say about exaflops and almost nothing to say about whether the people it's meant to serve want, trust or can use what is being built.

Silence on global inequality. Only rich states can afford a sovereign AI strategy at scale. The UK pursues its own compute in a world split into Compute North, South and Desert (Lehdonvirta, Wú and Hawkins, 2024). The material costs of AI, from mining and manufacturing to data labelling and energy, fall unevenly and often on the Global South (Crawford, 2021). Kwet (2019) calls the resulting dependence on Northern-owned digital infrastructure a form of "digital colonialism". Britain's sovereignty project worries a great deal about its own dependence on the US, and says nothing about other countries' dependence on countries like Britain.

7. What the cube fails to show

Using the cube on a national technology strategy also exposes its limits. Gaventa invites this. He calls the cube "not a checklist" and admits it was criticised as "a bit too static" (2006, pp. 31, 25).

It sees actors and arenas, not things. The cube asks who is in which space. But much of the power in AI sits in material infrastructure: chips, fabrication plants, undersea cables and the grid. Mann's (1984) idea of infrastructural power, a state's capacity to reach into and coordinate society through its infrastructure, captures something the cube can't, as does Bratton's (2015) idea of the Stack. Owning the layers is itself a form of power, whoever sits at which table. NVIDIA's position isn't mainly about setting the agenda in a meeting. It comes from controlling a bottleneck.

Its levels assume territory. Local, national and global form a nested, territorial scale. Corporate AI power doesn't sit at any one of these levels. A hyperscaler is global in ownership, national in regulation and local in its data centres, all at once. Gaventa accepts that the levels interconnect (p. 28), but the cube still has no obvious place for a networked actor that cuts across all of them.

It is a public-sphere tool. Gaventa notes that the cube "focuses primarily on power in the 'public sphere'" and accepts that this makes it incomplete (2006, p. 32, note 8). Much AI power lives in private space: proprietary models, commercial contracts, trade secrets. These aren't closed spaces in Gaventa's sense, because they never claimed to be public in the first place.

It is mostly about power over. The cube is largely about domination: who has power over whom. Sovereignty debates are just as much about power to, meaning capability, the ability to build and adapt (Charles and Gherman's third meaning). Gaventa defines power to, with and within (2006, p. 24), but the cube doesn't really put them to work. His later writing tries to connect them (Gaventa, 2021).

It starts from citizens. The cube explicitly "takes citizen action and participation as its starting point" (Gaventa, 2006, p. 26). Using it to analyse a state is a stretch. It can show that the UK is both powerful and dependent, but it can't tell us whether the push for sovereignty is a good thing: whether a more sovereign British state would be better for British people, or for anyone else. Gaventa says the cube isn't prescriptive (p. 31), but sovereignty is an unavoidably normative question.

It diagnoses but doesn't prescribe. The cube doesn't say how to turn the dial. Even if you accept that UK AI sovereignty is aligned on one face only, the cube has nothing to say about what full alignment would look like in a field that changes this fast. Spaces in AI governance open and close in months, not decades.

8. Conclusion

The UK wants AI sovereignty for understandable reasons: to avoid being switched off, to have a hand in the values built into a defining technology, and to own a share of the industry. Achieving it, as Charles and Gherman (2026) show, means controlling enough of the stack, and keeping enough choice over the rest, to go on deciding for yourself. Isambard-AI is real progress on ownership, jurisdiction and capability. It is weak on resilience, and it is a small part of the compute the UK actually uses.

Read through Gaventa's power cube, the UK's approach looks lopsided. It builds visible power at the national level, through closed and invited spaces, while leaving the hidden power of vendors and agendas, and the invisible power of the race narrative, mostly untouched. The UK is an invitee abroad and a gatekeeper at home. Claimed spaces are few and quickly closed. The local level, where the machines land and the people live, is barely in the picture. The cube has blind spots of its own: it struggles with material infrastructure, with corporate power that ignores levels, and with private space. But it asks the right first question.

Gaventa opened his paper by asking whether the new terrain of participation represented "a real shift in power" (2006, p. 23). For UK AI sovereignty, the honest answer is a partial yes: a shift, within narrow limits, from American firms towards the British state. It isn't yet a shift towards the people who will live with the results. A sovereignty worth the name would add what is currently missing. That means claimed spaces that are listened to, not shut down; transparency about whose data and whose work trains "British" AI; a real say for the communities that host the machines; and open models that the public can inspect, on the Swiss example. If the UK wants AI that reflects its values, it will need to ask its people what those values are. Sovereignty, after all, is supposed to belong to someone.

References

Bratton, B.H. (2015) The Stack: On Software and Sovereignty. Cambridge, MA: MIT Press.
Charles, V. and Gherman, T. (2026) 'The UK wants "AI sovereignty" to cut its reliance on foreign tech – here's how that can be achieved', The Conversation, 29 October.
Computer Weekly (2025) 'UK government signs US partnership to deliver Europe's largest AI factory'. Available at: computerweekly.com.
Cornwall, A. (2002) Making Spaces, Changing Places: Situating Participation in Development. IDS Working Paper 170. Brighton: Institute of Development Studies.
Couture, S. and Toupin, S. (2019) 'What does the notion of "sovereignty" mean when referring to the digital?', New Media & Society, 21(10), pp. 2305–2322.
Crawford, K. (2021) Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press.
DatacenterDynamics (2025) 'UK's 5MW Isambard-AI supercomputer goes live in Bristol'. Available at: datacenterdynamics.com.
DSIT (Department for Science, Innovation and Technology) (2025a) AI Opportunities Action Plan (independent report by M. Clifford), 13 January. Available at: gov.uk.
DSIT (2025b) UK Compute Roadmap, July. Available at: gov.uk.
Gaventa, J. (1980) Power and Powerlessness: Quiescence and Rebellion in an Appalachian Valley. Oxford: Clarendon Press.
Gaventa, J. (2006) 'Finding the spaces for change: a power analysis', IDS Bulletin, 37(6), pp. 23–33.
Gaventa, J. (2021) 'Linking the prepositions: using power analysis to inform strategies for social action', Journal of Political Power, 14(1). doi:10.1080/2158379X.2021.1878409.
Hayward, C.R. (1998) 'De-facing power', Polity, 31(1), pp. 1–22.
insideHPC (2025) '£225M HPE-NVIDIA "Isambard-AI" supercomputer launched in UK', July. Available at: insidehpc.com.
Jennings, N. (2026) 'AI poses three big types of risk – here's how worried we should be about them', The Conversation, 27 October.
Kwet, M. (2019) 'Digital colonialism: US empire and the new imperialism in the Global South', Race & Class, 60(4), pp. 3–26.
Lehdonvirta, V., Wú, B. and Hawkins, Z. (2024) 'Compute North vs. Compute South: the uneven possibilities of compute-based AI governance around the globe', Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), pp. 828–838. doi:10.1609/aies.v7i1.31683.
Lukes, S. (1974) Power: A Radical View. London: Macmillan.
Mann, M. (1984) 'The autonomous power of the state: its origins, mechanisms and results', European Journal of Sociology, 25(2), pp. 185–213.
Music Business Worldwide (2025) 'UK's House of Lords votes to strengthen copyright protections in AI, dealing blow to government's plans'. Available at: musicbusinessworldwide.com.
Pohle, J. and Thiel, T. (2020) 'Digital sovereignty', Internet Policy Review, 9(4). doi:10.14763/2020.4.1532.
Prime Minister's Office (2025) 'Prime Minister sets out blueprint to turbocharge AI', 13 January. Available at: gov.uk.
Siemens, G. (2023) 'Has GPT-4 really passed the startling threshold of human-level artificial intelligence? Well, it depends', The Conversation, 31 March.
Tech Policy Press (2025) '"Tech Prosperity Deal" binds UK to US AI dominance strategy'. Available at: techpolicy.press.
The Bookseller (2025) 'Industry leaders urge prime minister to enforce copyright law after AI amendment rejected'. Available at: thebookseller.com.
Tom's Hardware (2025) 'UK cosies up to big tech with $42 billion data center and AI investment deal'. Available at: tomshardware.com.
University of Bristol (2025) 'Lifting the lid on Isambard-AI, the UK's fastest AI supercomputer'. Available at: bristol.ac.uk.
VeneKlasen, L. and Miller, V. (2002) A New Weave of Power, People and Politics: The Action Guide for Advocacy and Citizen Participation. Oklahoma City: World Neighbors.
Whittaker, M. (2021) 'The steep cost of capture', Interactions, 28(6), pp. 50–55.

Written for my own interest, not as assessed coursework. Page numbers for Gaventa (2006) refer to the printed IDS Bulletin pagination. Figures and policy details are correct as of early November 2026.