The familiar story about artificial intelligence has been told through three variables: data, talent and algorithms. That framing is now incomplete. By mid-2026, the more revealing lens is infrastructure. Advanced AI depends on a stack of physical and institutional systems that are expensive to build, difficult to substitute and increasingly entangled with national strategy. The real question is no longer simply who can design the best models, but who can secure sustained access to the compute required to train, deploy and adapt them.
This shift matters because compute is not a single commodity. It is a compound asset made up of semiconductors, advanced packaging, networking equipment, hyperscale cloud architecture, specialist software, cooling systems, land, grid access, and patient capital. It is governed partly by markets and partly by export controls, procurement rules, competition policy and energy planning. In that sense, compute sovereignty is emerging as a more precise term than digital sovereignty. It focuses attention on the material bottlenecks that shape what countries, firms and public institutions can actually do.
For policymakers, this is an awkward development. Many governments entered the AI debate through ethics frameworks, safety institutes and data governance. Those remain necessary. But they do not answer a harsher question: what happens when a country has world-class researchers, a robust regulatory regime and strong demand for AI adoption, yet depends on a narrow group of foreign suppliers for the infrastructure underpinning modern systems. The answer, increasingly, is that policy autonomy begins to narrow long before any formal crisis arrives.
From software politics to infrastructure politics
For much of the past decade, AI policy was discussed as if software were the decisive layer. It was natural to think this way. Software appears scalable, portable and relatively independent of geography. Governments could imagine that domestic capability might be cultivated through education, research grants and startup finance. The arrival of foundation models altered that assumption. Training frontier systems required concentrations of compute so large that only a few firms and states could absorb the capital cost and operational complexity.
The result has been a subtle but important relocation of power. Influence now sits not only with model developers but with whoever controls chip design, fabrication, interconnects, cloud distribution and the energy systems feeding data centres. The economics resemble previous infrastructure transitions more than classical software markets. Scale advantages compound over time. Switching costs are high. Capacity is often pre-committed. Public agencies, universities and smaller firms become price takers unless states intervene intelligently.
The most consequential AI bottleneck is not always the algorithm; it is the physical system that trains, hosts and powers it.
This is why AI industrial policy has broadened. Measures once treated separately, such as semiconductor incentives, foreign investment screening, spectrum policy, transmission planning and public cloud procurement, now interact. A nation can no longer assess AI readiness simply by counting research papers or venture capital rounds. It must ask whether its institutions can obtain secure, affordable and lawful compute at strategic moments.
Why sovereignty now means optionality
Sovereignty is often misunderstood as self-sufficiency. In advanced computing, that is unrealistic for almost everyone. Semiconductor production is globally distributed by necessity; no major economy controls every stage from electronic design automation to lithography, fabrication, packaging and deployment. The more practical objective is optionality: the capacity to avoid single-point dependence and to preserve room for political and economic choice.
That definition is gaining traction because dependencies in AI are layered. A state may have domestic data centres yet rely on foreign-designed accelerators. It may subsidise fabrication yet lack advanced packaging. It may host cloud regions yet find that pricing, software ecosystems and technical standards are still set elsewhere. Each dependence may look manageable in isolation. Together they can create strategic lock-in.
Optionality therefore has several dimensions. The first is supply diversity, especially across chips, cloud providers and networking equipment. The second is institutional competence: governments need procurement, technical and regulatory capabilities equal to the complexity of the systems they buy and oversee. The third is legal resilience, including rules on access, portability, cybersecurity and continuity under geopolitical stress. A country that lacks these features may discover that nominal market access is not the same as operational control.
Compute sovereignty does not mean autarky. It means retaining meaningful choice under pressure.
The most consequential AI bottleneck is not always the algorithm; it is the physical system that trains, hosts and powers it.
The hidden economics of concentrated capacity
Much commentary still treats compute as if it were purchased on demand like any standard cloud service. That view obscures how concentrated advanced capacity has become. Frontier training requires enormous up-front reservation of hardware, specialised engineering teams and long planning horizons. Even inference at scale increasingly depends on high-end accelerators and optimised software stacks. The nearer one moves to the frontier, the less the market resembles abundant commodity computing.
These economics produce second-order effects. Public research institutions struggle to compete for access when commercial demand surges. Smaller countries may find that domestic AI strategies rest on infrastructure they do not influence. Start-ups can become technically sophisticated yet commercially dependent on a handful of providers whose pricing, credits, service design and interoperability rules shape survival. Competition authorities have begun to look more closely at cloud markets for precisely this reason: concentration may not only affect prices, but innovation pathways and strategic dependence.
There is also a temporal problem. Compute shortages matter most at moments of transition, when a new model class, chip generation or optimisation technique changes the feasible frontier. Those with immediate access can learn and adapt faster; those without fall behind not linearly but discontinuously. The gap is difficult to close because infrastructure investment cycles are long, while algorithmic progress is fast.
Energy is becoming AI policy by another name
A second blind spot in AI strategy has been energy. Data centres have always required power, but the scale and intensity associated with advanced AI have made electricity availability a strategic constraint rather than an operational detail. Grid connection queues, transmission bottlenecks, local permitting, water constraints for cooling and volatility in power markets now shape where capacity can be built and how reliably it can operate.
This has altered the geography of AI. Regions with abundant low-carbon electricity, grid stability and faster permitting regimes have acquired an advantage not captured by traditional innovation metrics. In some cases, the contest is less about who has the best tax incentive than who can deliver power at the right quality and timeline. This is one reason energy ministries, transmission operators and environmental regulators have become unexpectedly central to AI policy.
The political consequences are substantial. If national AI ambitions require very large additions of data-centre load, then public authorities must balance industrial strategy against household prices, local environmental concerns and decarbonisation commitments. Cheap rhetoric about becoming an AI leader encounters the practical realities of substations, land-use conflicts and balancing intermittent generation. In several jurisdictions, the bottleneck is no longer capital alone, but the state’s ability to coordinate infrastructure systems that were planned for a different era.
The return of the state, but in a more technical form
Governments have returned to industrial policy with unusual force, from semiconductor subsidies to strategic procurement and foreign investment screening. Yet the state’s role in compute sovereignty is not simply to spend more. It must become more technically literate. Poorly designed interventions can lock in weak architectures, crowd out private investment or subsidise prestige projects with little strategic value.
The more durable approaches share a few traits. They distinguish between layers of the stack rather than treating AI as a single sector. They recognise where domestic advantage is plausible and where alliances are more sensible. They use procurement to create dependable demand for secure, portable and auditable computing services, especially for public-interest research and critical functions. And they connect compute planning to skills, energy and competition policy rather than leaving each domain to proceed separately.
This is a demanding agenda because state capacity itself has become part of the compute problem. A government may announce a national strategy, but if ministries cannot negotiate cloud contracts, assess chip dependencies, model energy demand or enforce interoperability, sovereignty remains rhetorical. The challenge is administrative as much as financial.
Europe’s dilemma: regulation without enough infrastructure
Europe offers the clearest case of this tension. The continent has been influential in digital regulation, from the GDPR to the AI Act, and has a strong scientific base in several relevant fields. Yet its recurring weakness lies in scaling infrastructure at the pace and concentration now required. The issue is not absence of capability but fragmentation across capital markets, energy systems, procurement practices and national industrial priorities.
Compute sovereignty does not mean autarky. It means retaining meaningful choice under pressure.
The Draghi report sharpened the diagnosis by linking competitiveness to lagging investment and incomplete integration. In AI, that diagnosis becomes concrete. Europe can shape rules for trustworthy deployment, but rule-setting does not automatically produce domestic compute capacity. Nor does semiconductor policy by itself solve dependence if advanced packaging, cloud orchestration and software ecosystems remain externally controlled.
The strategic risk is a divided outcome: Europe becomes a significant regulator and adopter of AI, but not a decisive shaper of the infrastructure on which its own public services, firms and researchers depend. That would leave it with considerable normative power but reduced industrial leverage. The policy response increasingly points towards shared supercomputing resources, public-private infrastructure consortia and stricter attention to portability and competition in cloud markets. Whether that becomes a coherent continental strategy remains uncertain.
Middle powers and the politics of selective depth
Not every country can or should aspire to end-to-end capability. For middle powers, the more realistic path is selective depth. That means identifying a limited number of strategic layers where domestic competence matters most and where comparative advantage is plausible. For some, this may be energy-rich data-centre hosting; for others, specialised chip design, photonics, trusted testing environments, or public compute infrastructure for research.
The key is not to imitate the largest economies but to avoid becoming a passive consumer of external stacks. Selective depth can create bargaining power. A country with robust public compute for universities and critical sectors, clear rules on data portability, and credible continuity planning is less exposed than one that relies entirely on opaque commercial arrangements. Equally, countries with strong positions in grid infrastructure, subsea connectivity or advanced materials may find that their strategic role in the AI economy is larger than headline model rankings imply.
This is an unexpected but important shift in the development conversation. For years, digital ambition was often equated with building platforms or incubating apps. The AI era rewards some older strengths: engineering discipline, utilities planning, exportable industrial components and patient institutions.
Standards, interoperability and the quiet architecture of dependence
Strategic dependence is not created only by physical scarcity. It is also encoded in interfaces, APIs, software toolchains and operational standards. Once developers, enterprises and public agencies build around a given environment, switching becomes painful even if alternatives exist. This is where standards policy, often treated as technocratic housekeeping, becomes geopolitically significant.
Interoperability can reduce lock-in, but only if it is pursued consistently across procurement, competition and cybersecurity. Public buyers matter here. Governments are large consumers of cloud and AI services, and the terms they accept can either reinforce concentration or create pressure for portability, auditing and modularity. The issue is not hostility to scale; some concentration is inherent in advanced infrastructure. The issue is whether users retain enough technical and contractual freedom to adapt if circumstances change.
Standards also affect safety and accountability. Reliable documentation, benchmark transparency, secure logging and incident reporting are not peripheral governance issues. They are preconditions for operating complex AI systems across institutional boundaries. In this sense, technical governance and sovereignty converge. A country that cannot inspect, migrate or meaningfully assess the systems it relies upon has less control than formal ownership might suggest.
Security policy is merging with economic policy
Export controls and investment restrictions made explicit what had long been implicit: advanced compute sits at the junction of economic competition and national security. The distinction between civilian and security-relevant capability has blurred, not because every AI application is military, but because the underlying infrastructure is general purpose. Capacity built for commercial services can support cyber operations, intelligence analysis, defence logistics and dual-use research.
This does not mean every country should securitise its entire AI agenda. Over-securitisation can chill collaboration and distort commercial decision-making. But it does require a more integrated policy frame. Defence planners, civil regulators and economic ministries are now dealing with overlapping assets: chips, cloud contracts, energy resilience, undersea cables, specialised talent and software supply chains. Fragmented governance leaves dangerous blind spots.
In AI, dependence accumulates quietly: through chips, through cloud contracts, through power markets, and through technical standards.
NATO and other security institutions have already recognised emerging and disruptive technologies as central to alliance resilience. The next step is operational. Resilience in AI will depend less on abstract declarations than on whether allied countries can ensure continuity of compute under stress, share trusted infrastructure where appropriate, and reduce common dependencies that could be exploited coercively.
What markets still get right
It would be a mistake to conclude that compute sovereignty requires heavy-handed state control over every layer of the stack. Markets still perform essential functions. They allocate capital, reward engineering efficiency and enable rapid experimentation. Much of the progress in semiconductors, cloud architecture and optimisation has been driven by competitive pressure and the search for commercial advantage.
The challenge is that market outcomes do not automatically align with public resilience. Left alone, firms optimise for returns, not necessarily for strategic redundancy, equitable research access or continuity under geopolitical strain. This is why a mixed model is emerging. States set boundaries, support critical capabilities and shape incentives; markets deliver much of the underlying execution. The hard part is calibration. Too little intervention leaves dependence unaddressed. Too much can freeze innovation and misallocate resources.
The most capable governments are therefore becoming market shapers rather than market substitutes. They identify failure points, reduce coordination problems and use public purchasing, standards and targeted finance to alter incentives without pretending to out-engineer the private sector wholesale.
The next phase of AI competition
By mid-2026, AI competition is entering a less theatrical and more structural phase. Public attention still gravitates towards model releases and headline valuations, but the deeper contest is over enduring capacity: fabs, packaging plants, cloud regions, transmission lines, software ecosystems, and the legal mechanisms governing access to them. This phase is slower, more technical and more consequential than the previous one.
It also changes how success should be measured. The relevant questions are not only who publishes the most impressive benchmark, but who can support broad-based deployment across public services and industry; who can maintain research access during shortages; who can keep critical systems running through supply disruptions; and who can bargain from a position of institutional competence rather than dependency.
In AI, dependence accumulates quietly: through chips, through cloud contracts, through power markets, and through technical standards.
The countries that navigate this transition best are unlikely to be those with the loudest rhetoric. They will be the ones that treat compute as a strategic system: material, legal, financial and energetic at once. That requires a broader imagination than conventional tech policy has usually supplied. In the coming decade, sovereignty in AI will be determined less by aspiration than by infrastructure, and less by slogans than by the patient organisation of capacity.
Research agendas for a compute-constrained world
For scholars and public-interest researchers, this new geography of compute suggests a revised agenda. One line of inquiry concerns measurement: how to assess national and regional compute capacity without relying on opaque commercial disclosures. Another concerns governance design: what institutional arrangements best preserve research access, portability and resilience without dulling innovation. A third concerns energy and environmental trade-offs, including how to integrate large data-centre demand with decarbonisation pathways and local legitimacy.
There is also a need for more comparative work. Different jurisdictions are assembling compute sovereignty through different combinations of subsidy, competition oversight, public procurement and alliance-building. Comparing those models will be more useful than recycling generic claims about AI leadership. The political economy of compute is still young as a field of study. But it is already clear that the future of AI will be shaped as much by the governance of infrastructure as by the science of models.
That may be the least glamorous insight in the current debate. It is also one of the most important.
