For years, the public argument about artificial intelligence focused on visible artefacts: chatbots, image generators, foundation models and the firms that released them. That framing is becoming insufficient. By mid-2026, the more consequential question is less what a model can do than where the underlying compute can legally, physically and economically exist. AI has entered an infrastructural phase. Its politics increasingly resemble those of energy, shipping and telecoms: disputes over capacity, chokepoints, standards, subsidies, security and jurisdiction.
This shift matters because compute is not a neutral substrate. It is a bundle of scarce semiconductors, specialised networking, abundant electricity, water, cooling systems, skilled operators, cloud control planes and regulatory permissions. Unlike software, it does not move frictionlessly. Unlike data, it cannot simply be copied across borders at negligible cost. And unlike traditional industrial plant, it sits inside an ecosystem whose strategic value is amplified by learning effects and military relevance. The result is a new map of power in AI, one organised not only by nations and firms but by what might be called compute territories.
The strategic unit of AI is no longer only the model or the company, but the compute territory in which both are embedded.
From digital abstraction to industrial location
The first wave of internet globalisation encouraged a certain mythology: software was weightless, cloud services were placeless and digital markets would erode geography. AI has reversed that assumption. Training and serving advanced models require concentrations of hardware that are expensive to acquire, difficult to power and politically sensitive to export. The cloud remains abstract from the user’s perspective, but the state increasingly sees racks, substations, pipelines, cables and ports.
That is one reason why AI policy has become entangled with industrial policy. The American CHIPS and Science Act was framed around semiconductor resilience and domestic capability, but its strategic logic extends downstream into who can build and operate large-scale AI infrastructure. Export controls administered by the Bureau of Industry and Security have reinforced this point: access to frontier computing capability is treated not merely as a commercial issue, but as a matter of national security. Europe, meanwhile, has paired regulatory ambition through the AI Act with a more explicit concern for competitiveness, energy availability and sovereign capacity. Across Asia and the Gulf, governments have pursued data-centre expansion, chip packaging, power investments and cloud partnerships as elements of statecraft rather than mere investment promotion.
Why compute has become the real bottleneck
There is a temptation to describe compute scarcity as a transient inconvenience, soon to be solved by larger capex budgets and newer chips. That underestimates the depth of the constraint. Even as semiconductor production expands, advanced AI workloads remain limited by several interlocking factors: leading-edge fabrication capacity, advanced packaging, high-bandwidth memory, interconnects, dependable low-carbon electricity, and the physical time required to permit and construct facilities. A shortage in any one layer can delay the whole stack.
The International Energy Agency has been careful in its treatment of data-centre demand, but the direction is unmistakable. AI is now material to power planning. In multiple jurisdictions, the practical limit on new AI capacity is not investor appetite but grid access and the speed of approvals. This gives electricity regulators and local planning authorities a surprisingly central role in the future of AI. It also means that compute politics will increasingly be fought in arenas far removed from model evaluations or content moderation: transmission queues, water permits, land-use hearings and long-term power purchase negotiations.
The old idea that the frontier was constrained mainly by algorithmic ingenuity has not disappeared. But the balance has changed. Efficient model design may reduce the number of required training runs; it does not remove the need for physical concentration. In fact, improved algorithmic efficiency can make infrastructure more strategically valuable, because the actors with the best access to compute can translate those gains into further scale more quickly than others.
Compute territories are emerging
The strategic unit of AI is no longer only the model or the company, but the compute territory in which both are embedded.
A compute territory is not simply a country. It is a jurisdictional and infrastructural zone in which legal authority, energy systems, data-centre concentration and semiconductor access align sufficiently to support sustained AI activity. Some are national in character, such as the United States’ cluster of hyperscale regions tied to domestic chip policy and export controls. Others are more distributed, as in Europe, where capacity is shaped by the interplay between EU law, national industrial strategies and cross-border energy markets. Still others are city-state or corridor based, linking ports, subsea cable landing stations, special economic zones and energy-rich regions.
This matters because firms do not choose locations on cost alone. They choose them according to a matrix of regulatory predictability, geopolitical risk, talent access, cooling conditions, insurance, tax treatment and grid reliability. The outcome is a world in which the practical geography of AI may diverge from the legal geography of the nation-state. A government may possess strong AI rules yet weak infrastructure; another may have ample power and permissive siting but limited access to advanced chips; a third may have rich capital markets and cloud ecosystems yet remain dependent on external semiconductor chokepoints.
Seen in this light, debates about sovereignty are becoming more granular. The relevant question is not whether a state is “in” or “out” of AI, but which parts of the stack it controls, which it leases, and which it can credibly protect during a crisis.
The return of strategic dependence
Globalisation did not disappear in AI; it became more selective. Very few jurisdictions can independently provide every component of the compute stack. Even large economies remain dependent on external suppliers for lithography, fabrication equipment, speciality chemicals, memory, GPUs, network gear or cloud software. The OECD’s work on semiconductor supply chains has made this point repeatedly: resilience in this sector cannot be achieved through simple replication because the ecosystem is too specialised and capital-intensive.
That creates a dilemma. States want to reduce dependence on strategic rivals without paying the cost of full technological autarky. The result is a layered approach: subsidise domestic production where feasible, cultivate trusted partners, screen foreign investment in sensitive areas, and reserve emergency powers for extreme cases. Yet this patchwork can itself increase friction. Export controls intended for security purposes may complicate legitimate research collaboration; local content requirements may slow deployment; subsidy races may overbuild in politically favoured regions while underinvesting in transmission or skills.
States once worried about data localisation; they now face the harder problem of compute localisation without technological autarky.
The phrase “digital sovereignty” once referred mainly to data governance and platform dependence. In 2026 it increasingly concerns the terms on which compute can be procured, scaled and governed. The politics are more awkward because compute cannot be localised by decree alone. It requires an industrial base, cheap and reliable power, trusted hardware pathways and a market large enough to justify utilisation.
Law is moving down the stack
Regulation, too, is becoming infrastructural. The EU AI Act remains principally a framework for AI systems and their uses, especially those deemed high-risk or prohibited. But the wider policy environment around it is beginning to shape where advanced compute is deployed. Cybersecurity obligations, energy disclosure, public procurement rules, environmental permitting, competition investigations and cloud sovereignty provisions all influence the attractiveness of a location for AI infrastructure.
Elsewhere, the same pattern appears through different legal instruments. In the United States, security-oriented tools have had outsized impact, from export controls to investment screening and federal procurement preferences. In Asian markets, strategic planning often moves through industrial ministries, utility arrangements and state-supported infrastructure vehicles rather than through omnibus AI statutes. The legal consequence is that AI governance can no longer be read solely through specialised “AI law”. One has to read electricity regulation, trade law, critical infrastructure protections and financial sanctions as part of the same field.
This is one reason the governance debate can appear disjointed. Policymakers discuss safety in one room, chips in another, and energy in a third, even though firms experience them as a single operating environment. The next generation of serious AI governance will need to integrate these layers rather than treat them as separate domains.
States once worried about data localisation; they now face the harder problem of compute localisation without technological autarky.
Energy policy is now AI policy
If there is one area where the materiality of AI is impossible to ignore, it is energy. Large-scale training and inference require persistent, high-quality power. The question is not merely total electricity consumption, but the temporal and locational profile of demand. AI facilities often need rapid connection, high uptime and predictable expansion paths. That pushes them towards regions with flexible generation, robust transmission, storage capacity and politically manageable permitting.
For governments committed simultaneously to decarbonisation and digital competitiveness, this creates a difficult balancing act. Fast data-centre growth can raise local concerns about land use, water stress and the crowding-out of other industrial loads. It can also force uncomfortable trade-offs between clean-energy objectives and the desire for immediate baseload reliability. Some jurisdictions are likely to discover that their most ambitious AI strategies depend less on software talent than on whether they can build substations fast enough.
The energy-AI nexus also alters international alignment. Power-rich states with stable institutions may become more attractive as hosts for compute, even if they were not previously seen as central digital players. Conversely, countries with strong research talent but constrained grids may find themselves specialising in design, evaluation or applications while relying on external compute hubs for heavy workloads. The map of comparative advantage is being rewritten in electrical rather than purely digital terms.
The uneven geography of compliance
One underappreciated consequence of compute concentration is the concentration of compliance capability. Advanced AI governance is not costless. It requires auditing tools, logging systems, model documentation, cybersecurity controls, legal expertise and supply-chain assurance. Large infrastructure operators can spread these costs across extensive operations. Smaller providers and public-sector actors often cannot. This may lead to an unintended consolidation in which the most compliant actors are also the largest, not because they are intrinsically more virtuous, but because they can absorb the overhead.
That creates a strategic challenge for governments seeking both innovation and accountability. If compliance is too fragmented, governance becomes porous. If it is too centralised, the market may harden around a small set of infrastructural gatekeepers. The same concern applies internationally. Countries with the administrative capacity to monitor compute use and enforce standards may become preferred venues for high-value AI activity. Those without such capacity risk becoming regulatory blind spots or, conversely, being bypassed altogether by risk-averse operators.
In this sense, sovereignty in AI is partly bureaucratic. It depends on whether a state can inspect, verify and govern technical systems that are both globally connected and locally situated. Administrative weakness can be as limiting as hardware scarcity.
Public compute and the problem of access
As compute becomes strategic, demands for public or quasi-public access have grown louder. Universities, scientific institutes and start-ups argue, with reason, that frontier capability should not be determined solely by corporate balance sheets. Several governments have therefore explored national compute facilities, sovereign cloud arrangements or publicly supported access programmes. The impulse is understandable: if compute is a foundational input to research and economic development, leaving its allocation entirely to market power may narrow the innovation base.
Yet public compute is harder to organise than public research grants. Hardware depreciates quickly. Operating costs are high. Demand is volatile. Security requirements are significant. There is also a persistent risk that politically announced capacity does not translate into practically usable access because software stacks, queueing policies and support services are inadequate. The challenge is not only to buy accelerators, but to build institutions that can govern scarce compute fairly and efficiently.
The more subtle issue concerns purpose. Public compute can serve different ends: scientific discovery, industrial incubation, public-interest model development, defence-related research or national resilience. These goals are not identical. A country that fails to prioritise among them may build facilities that satisfy none particularly well.
The real constraint on advanced AI is becoming an awkward triad of chips, electricity and permitting.
Security doctrine is broadening
Security thinking around AI initially centred on misuse, cyber risks and the military applications of autonomous systems. Those concerns remain real. But infrastructure concentration is forcing doctrine to widen. A large compute cluster is also a critical asset vulnerable to espionage, supply interruption, sabotage, ransomware, insider compromise and cascading energy failures. In strategic terms, it resembles a hybrid of data centre, power-intensive industrial plant and communications node.
This has at least three implications. First, cloud security and national critical infrastructure policy are converging. Second, model security cannot be separated from hardware provenance and facility operations. Third, resilience planning must consider not only catastrophic scenarios but also chronic friction: delayed maintenance, shipping disruptions, software patch dependencies and insurance constraints in geopolitically exposed regions.
For defence planners, the issue is broader still. Access to reliable compute increasingly shapes intelligence analysis, simulation, cyber operations and logistics. As a result, civilian AI infrastructure may acquire strategic significance even when it is not formally part of the defence industrial base. The old distinction between commercial digital capacity and national security capability is becoming less tidy.
A planetary system with regional centres
The phrase “planetary AI” can imply a seamless global layer hovering above politics. The reality is more terrestrial. AI is becoming a planetary system in the same way shipping or telecommunications are planetary: globally interconnected but regionally concentrated, politically contested and governed through overlapping regimes rather than a single authority. There will be no neat division between open global markets and closed national fortresses. Instead, the world is likely to settle into a mixed order of allied corridors, regulated interdependence and selective exclusion.
For business, this means strategy must account for jurisdictional topology as much as technical roadmaps. For policymakers, it means that competitiveness cannot be reduced to subsidies for model development while neglecting permitting, workforce, transmission and standards capacity. For civil society, it means the governance of AI power will increasingly depend on institutions that traditionally attracted little attention in digital debates: utilities commissions, competition authorities, customs agencies, export-control offices and planning boards.
The real constraint on advanced AI is becoming an awkward triad of chips, electricity and permitting.
What to watch in the second half of the decade
Three questions will help determine how this geography evolves. The first is whether advanced compute markets become more interoperable across trusted jurisdictions or fragment into rival blocs with incompatible compliance and security expectations. The second is whether energy systems can expand quickly enough, and cleanly enough, to accommodate AI without provoking political backlash. The third is whether public institutions can gain enough technical competence to govern compute concentration before dependency hardens into a new form of private infrastructural power.
Much commentary still treats AI as if its decisive battles will be fought at the level of applications or consumer adoption. Those battles matter, but they are downstream. The deeper contest concerns the territorialisation of compute: who hosts it, who finances it, who secures it, who regulates it and who may be denied it. That is a more prosaic story than the mythology of machine intelligence, but in 2026 it is the one most likely to shape the political economy of the field.
In the coming years, the winners in AI may not simply be those with the best models. They may be those that best align semiconductor access, electrical abundance, legal predictability and institutional competence. That combination is rarer than enthusiasts suppose. It is also why the next map of AI will look less like a leaderboard and more like infrastructure geopolitics.
