For years, artificial intelligence was discussed as if it floated above geography. The important questions, it was said, concerned data, talent, algorithms and capital. Location mattered mainly as a shorthand for industrial clusters: Silicon Valley, Shenzhen, Bengaluru, London. By mid-2026 that abstraction has become harder to sustain. The constraints that now shape frontier AI are increasingly physical. They concern transmission lines, substations, advanced packaging capacity, water rights, permitting battles, export controls and the politics of who gets access to scarce megawatts.
This shift does not mean software has stopped mattering. Nor does it diminish the significance of the European Union’s AI Act, NIST’s risk-management work or the OECD’s efforts to harmonise policy language. But it does suggest that the most consequential map of AI now looks less like a venture-capital slide and more like an infrastructure atlas. To understand who can train, deploy and govern advanced systems at scale, one must ask where compute can actually be built, and under what conditions.
From digital myth to industrial fact
The mythology of the internet era treated computation as placeless. Cloud services encouraged this view by hiding hardware behind elegant interfaces. Yet large-scale AI has exposed the material base of the digital economy with unusual clarity. Training and serving powerful models require concentrated energy demand, specialised chips, resilient network connections and operational expertise. These are not infinitely replicable. They depend on industrial ecosystems that evolve slowly and are vulnerable to disruption.
That materiality matters because AI is moving from episodic experimentation to continuous infrastructure. A laboratory can borrow capacity for a single run; an economy-wide AI system for healthcare, logistics, manufacturing or public administration needs durable access to compute over years. The question is no longer merely whether a model can be trained. It is whether nations, firms and institutions can secure long-term computational sovereignty without incurring untenable dependence on a handful of regions and suppliers.
Artificial intelligence is becoming less weightless than its mythology suggests.
The hidden bottleneck is not only chips
Public debate often reduces the AI race to semiconductor fabrication. That is understandable. Advanced chips remain a central bottleneck, and policy interventions such as the US CHIPS Act and the European Chips Act were explicit acknowledgements that semiconductors are strategic assets. But fabrication is only one layer of the stack. Packaging, memory, networking equipment, clean-room tooling and software optimisation all shape usable compute. So do the prosaic systems that receive less attention: transformers, switchgear and grid interconnection queues.
In several advanced economies, the pace of data-centre demand is colliding with electricity systems designed for a different era. The issue is not always generation in the aggregate. Often it is local capacity, transmission congestion, or the time required to connect large loads safely. This creates a new hierarchy. Regions with abundant power but weak transmission may struggle. Regions with strong policy ambition but slow permitting may lag. Regions with existing industrial loads, flexible generation and politically acceptable expansion pathways may surge ahead.
Electricity becomes strategic policy
Artificial intelligence is becoming less weightless than its mythology suggests.
Energy policy and AI policy are therefore converging. That convergence remains underappreciated. Governments have spent the past several years debating model safety, copyright, competition and online harms. They are now being drawn into a less glamorous but ultimately decisive set of decisions: whether to prioritise power for data centres, how to price flexibility, when to allow behind-the-meter generation, and how to reconcile decarbonisation commitments with sharply rising digital demand.
This is not merely an engineering problem. It is a political one. When a hyperscale campus seeks a large grid connection, local communities may ask why scarce power and water should serve distant AI workloads rather than housing, public services or conventional industry. Regulators may worry about resilience. Utilities may welcome long-term demand but fear concentration risk. Environmental groups may question whether pledges on renewable matching adequately address local system stress. In short, AI expansion is being forced into the language of public utility politics.
The countries best positioned in this environment are not simply those with the biggest technology companies. They are those able to integrate industrial strategy, power-market design and planning reform. That may favour places with abundant low-carbon electricity, clear land-use rules and a state capable of moving faster on network build-out. It may also favour middle powers that cannot dominate frontier model research but can become attractive hosts for inference infrastructure, sovereign clouds or specialised industrial AI workloads.
Why packaging and the middle of the supply chain matter
Another surprise of the current moment is the strategic importance of activities once treated as secondary. Advanced packaging, substrate supply and high-bandwidth memory are no longer technical footnotes. They are leverage points. A state may subsidise fabrication, yet still find that useful output is constrained elsewhere in the chain. The result is a more granular politics of dependence. Instead of asking whether a country has a chip industry, policymakers must ask which parts of the post-fabrication ecosystem they control, and which single points of failure remain offshore.
This has implications for alliances. Semiconductor co-operation is becoming less about broad declarations of friendship and more about choreographing bottlenecks across trusted jurisdictions. The practical challenge is difficult. Every additional node in a friendly supply chain improves resilience in theory, but can increase cost and complexity in practice. The temptation, particularly under geopolitical pressure, is to overcorrect into duplication. The smarter approach is selective redundancy: enough spare capacity and diversification to absorb shocks, without assuming every country can or should replicate the full stack.
The cloud is becoming territorial
The early cloud market sold scale and abstraction. The emerging AI infrastructure market increasingly sells placement and assurance. Governments, hospitals, banks and defence ministries do not only want computational capacity; they want to know where it sits, who can access it, how it is governed and whether sanctions, export restrictions or extraterritorial legal claims could interrupt it. Data localisation debates anticipated this concern, but AI intensifies it because dependence is no longer only about data custody. It is also about sustained access to specialised processing.
The result is a more territorial cloud. Some states are pursuing sovereign or semi-sovereign arrangements. Others are encouraging domestic champions in hosting, networking and managed services around foreign hardware. Still others are accepting strategic dependence in exchange for speed, reasoning that trying to build a full domestic alternative would be ruinously expensive. None of these choices is costless. Full sovereignty can become a euphemism for thinner markets and weaker innovation; complete dependence can leave critical sectors exposed to external shocks.
The geography of compute is starting to matter as much as the quality of code.
Europe’s problem is not only regulation
The geography of compute is starting to matter as much as the quality of code.
Conventional commentary often caricatures Europe as rich in rules and poor in builders. There is some truth in the complaint that fragmentation, lengthy permitting and inconsistent capital markets have slowed continental scaling. Yet by 2026 Europe’s more interesting question is whether it can align three agendas that are usually discussed separately: AI governance, industrial competitiveness and energy-system modernisation. The Draghi report sharpened the competitiveness debate; the Chips Act and AI Act supplied parts of an institutional response. The missing element is execution across infrastructure.
Europe has assets that are often underrated. It has advanced industrial demand, leading research centres, sophisticated manufacturers, and in several markets an abundance of low-carbon generation relative to peers. It also has strong public legitimacy for safety and accountability frameworks. But these strengths can be neutralised if compute projects face interminable grid delays, if cross-border energy integration remains politically constrained, or if procurement systems fail to create reliable domestic demand for trustworthy AI services. The continent does not need autarky. It needs coherence.
Middle powers may gain leverage
A notable feature of the new compute map is that it creates opportunities for countries outside the traditional AI superpowers. Nations with cheap and reliable electricity, stable institutions, cool climates, subsea connectivity and pragmatic regulation can become disproportionately important. They may not produce the most celebrated models, but they can host inference clusters, backup capacity, specialist research environments or regulated-sector deployments. In doing so they acquire bargaining power in a system otherwise dominated by a few firms and a few chip-making regions.
This possibility should not be romanticised. Becoming a compute host brings trade-offs. Large data centres can provoke local backlash over land, energy and tax arrangements. The economic spillovers may prove thinner than promised if the highest-value intellectual property and management functions remain abroad. Governments must therefore distinguish between being a landlord and being a participant. The former yields rents; the latter requires deliberate strategies in skills, applied research, supplier development and public-sector adoption.
Regulation still matters, but differently
None of this renders AI regulation obsolete. On the contrary, as compute becomes scarcer and more strategic, governance questions intensify. The EU AI Act, sectoral rules, public-procurement standards and technical frameworks such as NIST’s become mechanisms for shaping demand as much as constraining supply. They influence which systems are deployable in high-stakes settings, what documentation is required, and how liability may be apportioned across complex value chains.
Yet the centre of gravity is shifting. The first wave of AI policy focused on model behaviour and content outcomes. The next wave will increasingly address access conditions: who can obtain large-scale compute, under what verification requirements, with which reporting duties and for which classes of use. Export controls were the opening chapter. Domestic allocation debates may be the next. If compute is recognised as strategically scarce, states will be tempted to reserve, prioritise or audit it much more closely. That would mark a significant departure from the permissive assumptions of the early cloud era.
The environmental argument is changing shape
Environmental debate around AI has often centred on emissions from training runs. That remains relevant, but it is no longer sufficient. The more durable questions concern system integration. Can large computational loads help absorb renewable generation through flexibility, or do they worsen peak stress? Are operators investing in local grid reinforcement, storage and demand response, or merely claiming annual renewable matching? How should policymakers weigh the social value of AI-enabled productivity against the opportunity cost of scarce clean power?
In the next phase of AI competition, grid access may prove as strategic as talent.
These are not questions with a single ideological answer. In some regions, AI infrastructure could support decarbonisation by financing new generation and flexible assets. In others, it may entrench fossil back-up or delay the electrification of transport and heating. A serious environmental analysis must therefore move beyond generic claims about green data centres. The relevant unit is the regional power system and its constraints. Planetary AI, in this sense, is inseparable from planetary energy governance.
Security risks travel through infrastructure
As compute concentrates, systemic risk rises. A handful of facilities, network choke points or specialised suppliers can become targets for sabotage, espionage, cyberattack or coercive diplomacy. Security planners have long worried about undersea cables, satellite links and semiconductor fabs. AI extends the list to include model-serving clusters, packaging lines and utility interfaces. The threat surface expands because the economic and political value of uninterrupted inference is growing. It is one thing for a consumer service to go offline briefly; another for industrial control, logistics scheduling or public administration systems to lose access to core AI functions.
This concern strengthens the case for redundancy and for more explicit public-private planning. It also complicates the fashionable argument that every critical service should become more AI-dependent as quickly as possible. Efficiency gains are real, but resilience may worsen if institutions outsource too much cognitive and operational capacity to infrastructures they do not control. A sober strategy would pair adoption with fallback plans, local expertise and regular stress tests of both digital and electrical dependencies.
What the next phase of competition looks like
The next phase of AI competition is unlikely to be won by the country with the loudest rhetoric about innovation. It will favour those able to combine several unglamorous competencies: building transmission, financing generation, securing semiconductor supply, training technicians, accelerating permitting, managing water use, and designing credible governance for sensitive sectors. These are state-capacity questions as much as technology questions. They reward patience, administrative competence and institutional trust.
For businesses, the implication is equally stark. Competitive advantage will not derive solely from model prowess. It will also depend on long-dated access to compute, the ability to place workloads in politically and operationally suitable regions, and the discipline to design around energy and regulatory constraints. The firms that treat infrastructure as a strategic function, rather than a background utility, are likely to be better prepared for the coming decade.
In the next phase of AI competition, grid access may prove as strategic as talent.
A different way to see planetary AI
The phrase planetary AI can invite grand, even cosmic thinking. But its most immediate meaning may be more terrestrial. Artificial intelligence is being woven into the planet through concrete systems: mines and fabs, ports and substations, permitting offices and competition authorities. Its trajectory will be shaped not only by scientific breakthroughs but by planning hearings, transmission corridors and supply-chain diplomacy. The glamour lies at the model layer; the power increasingly lies beneath it.
That is the unexpected entry point for understanding AI in 2026. The decisive map is not only of who has the cleverest researchers or the largest datasets. It is of who can marshal physical capacity with political legitimacy. In the short run, this may slow some of the exuberance that defined the previous phase of the industry. In the longer run, it may produce a healthier realism. A technology that aspires to reorganise economies cannot remain detached from the material systems that sustain economies. The new cartography of compute is therefore not a side story. It is the story.
