For two decades, digital economics taught policymakers to think in terms of software scale, network effects and low marginal cost. That frame is now incomplete. The decisive bottlenecks in advanced artificial intelligence are increasingly physical: wafer capacity, lithography tools, memory bandwidth, advanced packaging, substation upgrades, water access and the lead times of transmission infrastructure. What appears, on the surface, to be a software revolution is maturing into an industrial system with unusually hard constraints.
This matters because the economics of AI are no longer governed chiefly by abstract innovation metrics. They are being shaped by a stack of material dependencies that sit awkwardly between industrial policy, energy planning and national security. In that stack, semiconductors are only the most visible layer. The more revealing development is that compute itself has begun to behave like a strategic input akin to energy, logistics or finance: tradable, concentration-prone, security-sensitive and unevenly distributed.
The result is a new geography of power. It is not defined simply by who can design frontier models, but by who can secure sustained access to the upstream systems that make those models possible. Mid-2026 looks less like a pure innovation race than a contest over corridors of production and control.
From digital abundance to industrial scarcity
The early internet economy rewarded firms that could distribute code at near-zero cost. AI, particularly at the frontier, reverses some of that logic. Training and serving powerful systems requires large quantities of specialised chips, dense networking, cooling, storage and uninterrupted electricity. The relevant scarcity is not just talent, nor even data, but the capacity to convert capital expenditure into usable compute at scale.
This shift helps explain why states that were once content to regulate the digital layer are now intervening in fabrication, packaging and energy systems. The United States used the CHIPS and Science Act to rebuild domestic capabilities while tightening export controls on advanced semiconductors and associated manufacturing equipment. The European Union, having spent years defining digital regulation through GDPR and now the AI Act, has moved with greater seriousness towards semiconductor resilience and strategic capacity. East Asian manufacturing hubs, already central to the chip economy, have become more explicitly geopolitical assets.
The scarcity that matters is no longer data alone, but the ability to turn electricity, silicon and capital equipment into reliable compute.
This is not a return to old-style autarky. The semiconductor supply chain remains profoundly international, with design, fabrication, materials, tools and packaging distributed across multiple jurisdictions. But it is a move away from the assumption that efficiency alone should govern allocation. Resilience, alliance structure and controllability now carry a strategic premium.
Compute as a factor of production
Economists have long treated land, labour and capital as classical inputs, with energy often analysed as a foundational enabler. Compute now warrants similar treatment. It is measurable, scarce at the frontier, and increasingly necessary to produce value across sectors from biomedicine to defence simulation to financial services. Unlike general-purpose cloud capacity, frontier compute depends on concentrated nodes of capability whose expansion cannot be improvised quickly.
That distinction matters. Not all computation is strategic. Routine enterprise workloads can be distributed across broad cloud infrastructure. Frontier AI workloads, by contrast, are unusually sensitive to the availability of leading-edge accelerators, high-bandwidth memory, low-latency interconnects and software stacks optimised for large-scale parallelism. They also depend on procurement arrangements and long planning horizons that favour actors with balance-sheet strength and state backing.
Once compute is viewed in this way, several policy moves become easier to interpret. Export controls are not merely trade restrictions; they are attempts to shape the global distribution of a strategic production input. Public subsidies for fabs are not simply job-creation schemes; they are efforts to secure optionality under conditions of geopolitical stress. Electricity market reforms increasingly intersect with technology strategy because marginal megawatts can determine where data centres and AI clusters are built.
The scarcity that matters is no longer data alone, but the ability to turn electricity, silicon and capital equipment into reliable compute.
The hidden layer: advanced packaging and memory
Public debate still over-indexes on chip design and leading-edge fabrication nodes. Yet by 2026 one of the more consequential constraints lies in advanced packaging, memory integration and the specialised manufacturing steps that allow performance gains even when simple transistor scaling slows. As workloads become more demanding, the ability to combine compute, memory and networking efficiently matters as much as the headline process node.
This is a subtle but important economic shift. It means the strategic map of AI does not align neatly with the map of wafer fabs alone. Jurisdictions with expertise in packaging, test, substrate supply or power electronics can exert leverage disproportionate to their visibility in public discourse. It also means policy designed around a narrow understanding of semiconductor sovereignty may miss critical bottlenecks.
For investors and ministers alike, this complicates the standard question of where value accrues. The answer is no longer confined to model developers or foundries. It extends across a wider ecosystem of materials suppliers, equipment makers, utilities, engineering contractors and logistics providers. Exo-economics, properly understood, is about those outer layers where political economy and technical architecture meet.
Energy becomes destiny again
If semiconductors are the machinery of AI, electricity is the metabolism. The expansion of data centres, AI training clusters and inference-heavy consumer services has renewed an old truth: abundant, reliable and reasonably priced power is a precondition for industrial growth. The novelty is that advanced digital sectors, once imagined as relatively light-footed, now compete directly for grid capacity, transmission upgrades and firm generation.
In several economies, planning disputes around data centres increasingly resemble those once associated with heavy industry. Questions of land use, water intensity, backup generation, cooling technology and local network constraints have become politically salient. Grid connection queues, transformer availability and regional power pricing are no longer back-office issues. They are determinants of national competitiveness.
This creates a policy tension. Governments want the productivity gains promised by AI while also pursuing decarbonisation, affordability and energy security. The interaction is not straightforward. Some jurisdictions can pair data-centre expansion with renewables, storage and transmission investment. Others face trade-offs between industrial demand growth and already strained grids. The practical consequence is that compute capacity will cluster where energy systems can absorb it, not simply where software talent is concentrated.
In the next phase of AI competition, grid connection queues may matter almost as much as research papers.
Export controls and the politics of denial
The tightening of export controls by the United States and coordination with allies has changed the operating environment for advanced chips and AI-related equipment. The logic is clear enough: restrict access to the hardware and tools most relevant to frontier capability, while preserving advantage among a narrower set of trusted partners. What is less clear is the long-term equilibrium this produces.
Controls can slow diffusion, raise costs and force redesigns of supply chains. They can also stimulate substitution, stockpiling and indigenous capability efforts in targeted states. In other words, denial is rarely static. It reshapes incentives throughout the system. The recent emphasis on AI diffusion frameworks underlines that governments are thinking not only about military end uses, but about the geography of general-purpose compute itself.
Export controls have not deglobalised technology; they have reorganised it into narrower, more strategic corridors.
There is a danger, however, in assuming that every chokepoint remains durable. Industrial policy can broaden capacity over time, and strategic rivals can redirect resources towards the layers where catch-up is still feasible. The most plausible outcome is not full decoupling but a segmented market: high-trust compute blocs at the frontier, wider commercial ecosystems at lower performance tiers, and persistent contestation in the middle.
Europe's dilemma: regulation without enough infrastructure
Europe enters this phase with strengths in research, industrial engineering and standards-setting, but with evident weaknesses in scale infrastructure for frontier compute. The bloc has helped define the governance vocabulary of digital technology, first through privacy regulation and now through the AI Act. Yet rule-making does not itself generate fabs, transmission lines or large AI clusters.
The deeper dilemma is institutional. Europe is adept at market design and competition policy, but less comfortable with the speed and concentration that frontier compute appears to require. A fragmented energy landscape, variable permitting regimes and uneven capital-market depth make coordination difficult. Even where the policy ambition is present, execution tends to be slower than in the United States or parts of Asia.
That does not condemn Europe to irrelevance. It suggests a different strategic path: leveraging strengths in specialised semiconductors, power systems, industrial AI, scientific computing and trusted governance. But the notion that Europe can remain strategically autonomous in AI while importing most of the critical hardware and depending on external cloud concentration is becoming harder to sustain.
Middle powers and the rise of compute diplomacy
One of the more surprising consequences of this new geography is the enhanced significance of middle powers. Countries that possess reliable power, political stability, engineering talent, submarine cable connectivity or niche semiconductor capabilities can bargain more effectively than before. They may not host the entire stack, but they can become indispensable to parts of it.
This has given rise to what might be called compute diplomacy. Governments are negotiating over data-centre investment, energy offtake, cross-border power interconnection, chip packaging capacity, research collaboration and trusted cloud zones. These arrangements are not as visible as naval treaties or trade pacts, but they are becoming part of the architecture of alignment.
Such diplomacy is likely to matter most for states that cannot hope to dominate frontier model development outright. Their opportunity lies in positioning themselves inside the trusted corridors through which compute, talent and capital flow. The economic rewards may be significant, but so are the strategic choices: neutrality becomes harder when infrastructure itself acquires security meaning.
Corporate strategy under sovereign pressure
For multinational technology firms, the era of treating geography as a tax variable or compliance issue is over. Where capacity is built, which customers are served, how chips are sourced, and what workloads are permitted now sit under growing sovereign scrutiny. Boards must think less like global optimisers and more like managers of politically exposed infrastructure.
This changes investment logic. Redundancy, once criticised as inefficient, becomes a form of insurance. Supplier concentration carries greater geopolitical risk. Long-term power procurement becomes central to competitive planning. Legal teams, export-control specialists and grid engineers gain influence relative to the purely financial view of expansion.
In the next phase of AI competition, grid connection queues may matter almost as much as research papers.
There is also a subtle shift in business models. Access to premium compute is increasingly shaped by long-term contractual relationships rather than open-market availability. That favours larger incumbents and state-linked actors, potentially entrenching concentration. Competition authorities may eventually need to consider whether the structure of compute markets creates downstream distortions in AI innovation itself.
The measurement problem
One reason this transformation remains underappreciated is that statistical systems are poorly equipped to capture it. National accounts can describe investment in manufacturing or electricity infrastructure, but they struggle to represent the strategic quality of compute access. Patent counts and venture rounds tell only part of the story. A country may appear digitally sophisticated while lacking the practical ability to train or deploy advanced systems at scale.
More useful indicators would track the full chain: advanced chip imports and domestic output; exposure to foreign lithography and packaging; grid headroom for large data centres; transmission build-out timelines; concentration in cloud and accelerator procurement; and the legal restrictions governing cross-border access to high-end compute. These are not yet standard macroeconomic indicators, but they increasingly shape productivity potential.
For the field of exo-economics, this is the central analytical task: to measure the outer infrastructure that determines what societies can actually do with digital technology. The visible layer of applications is only the end of the pipe.
What sovereignty now means
Digital sovereignty used to imply control over data, platforms and regulatory standards. By 2026, that definition is too narrow. Sovereignty in the AI age also entails some combination of domestic capacity, allied access and emergency fallback across semiconductors, energy and network infrastructure. Few countries can achieve full self-sufficiency; the realistic objective is managed dependence.
Managed dependence requires clarity about which dependencies are tolerable, which are not, and under what conditions they become dangerous. It also requires institutions that can connect ministries which historically operated in silos: industry, energy, trade, finance, research and defence. The countries best positioned for the next decade may be less those with the loudest AI rhetoric than those with the most coherent cross-sector planning.
There is a broader lesson here. The digital economy did not escape material constraints; it merely obscured them for a time. AI has brought them back into view with unusual force. Silicon, copper, cooling and power have returned to the centre of strategic analysis.
The coming settlement
The likely settlement of the next few years is neither borderless globalism nor complete techno-national separation. It is a layered order in which frontier compute is concentrated within a limited set of trusted networks, while lower-end digital capacity remains relatively global. States will continue to subsidise domestic capability, scrutinise foreign dependencies and align infrastructure with security priorities.
This will have costs. Duplication can be inefficient, controls can distort markets, and tighter blocs may reduce the pace of some forms of diffusion. But the old settlement, built on the assumption that advanced digital infrastructure could remain strategically neutral, has already weakened. Compute has become too consequential for that fiction to hold.
The unexpected entry point into AI economics, then, is not the model leaderboard or the latest consumer application. It is the substation, the packaging plant, the equipment licence and the transmission corridor. Those prosaic assets now shape the possibilities of the most celebrated technology of the age. The map of the digital economy is being redrawn in steel, silicon and law.
