By mid-2026, the orthodox story about sovereign compute still begins with wafer fabs, lithography tools and the fear of strategic dependence. Yet that framing now misses the more practical question facing ministries, cloud buyers, grid operators and laboratories: even when chips exist somewhere in the world, who can lawfully buy them, where can they be packaged, how quickly can a data centre connect to electricity, which basin can spare cooling water, and under whose jurisdiction does rented compute ultimately sit. The decisive bottleneck moved from transistor geometry to permissioning.
This matters because compute has become a layered system rather than a single industrial product. Design, fabrication, memory, advanced packaging, server integration, interconnects, power infrastructure, software stacks and tenancy rights sit in different places, under different legal regimes, with different time horizons for expansion. A country may subsidise fabrication and still remain short of usable compute if it lacks substation capacity, if export licences delay accelerator imports, or if domestic researchers depend on foreign cloud contracts governed by extraterritorial rules.
2019–2021: sovereignty begins as an industrial policy story
Before the present cycle of controls and shortages, semiconductor sovereignty was already re-entering policy. Pandemic disruption and vehicle-chip shortages made visible what supply-chain specialists had long argued: semiconductors are not a commodity market in the ordinary sense, but a tightly coupled hierarchy of niche dependencies. The OECD, CSET and national governments mapped an ecosystem in which no large economy was genuinely self-sufficient, even where it retained strength in design or equipment.
In this phase, the preferred remedies were familiar. Governments discussed domestic fabrication incentives, trusted supply, strategic stockpiles and resilience audits. The United States moved towards the CHIPS and Science Act. The European Union advanced what became the European Chips Act. Much of the commentary treated sovereignty as the restoration of manufacturing capacity, especially at advanced nodes. That was understandable, but it presumed that fabrication would remain the master bottleneck.
It did not. As demand for machine-learning training and inference surged, the operative scarcity shifted upward from wafers to deployable systems: accelerators in quantity, high-bandwidth memory, advanced packaging, suitable server racks, and above all places to put them with enough power and cooling.
October 2022: export controls redraw the map
The turning point came with the United States export controls announced in October 2022 and extended in subsequent updates, culminating in a broader architecture by late 2023. The policy aim was strategic denial of advanced computing capability to countries of concern, but the economic effect was wider. Export control policy did not merely restrict chips; it reorganised geography.
Supply chains became more conditional. Compliance teams became strategic actors. Firms and public institutions had to ask not simply whether a component could be procured, but whether end use, end user and destination would survive regulatory scrutiny. This encouraged a partitioning of markets, inventory and data-centre footprints. It also sharpened the distinction between nominal capacity and usable capacity: a chip not licensed, a package not cleared, or a cloud region not approved is not sovereign compute in any meaningful sense.
Export control policy did not merely restrict chips; it reorganised geography.
The wider lesson was institutional. Compute could now be curtailed without sabotaging a fab or blockading a port. It could be slowed by paperwork, by legal thresholds on interconnect performance, by reporting obligations, by restrictions on remote service, or by uncertainty over future rule changes. This was a subtler and, in some respects, more scalable form of power.
2023: the bottleneck moves to packaging and memory
The decisive bottleneck moved from transistor geometry to permissioning.
As generative AI demand accelerated, policymakers discovered another inconvenient truth. Sovereign compute is not secured once silicon leaves the wafer. Advanced packaging, chiplets, interposers and high-bandwidth memory became central constraints. NIST and industry analysts had long noted the strategic significance of packaging, but the policy community often treated it as an afterthought. By 2023 it was impossible to do so.
Leading accelerators depended on highly specialised packaging capacity and memory supply concentrated in a small number of firms and jurisdictions. For countries trying to build national compute capability, this created a mismatch between public rhetoric and physical reality. Subsidies announced for front-end manufacturing did not automatically produce near-term accelerator availability. Nor could every country credibly replicate the full stack. The result was a politics of queue position.
In practice, sovereignty now lives in the queue. Who gets the scarce packaging slot, the scarce memory allocation, the scarce systems integrator attention, the scarce air-freight capacity, the scarce data-hall fit-out team? These are not glamorous questions. They are nevertheless the questions that decide whether a new national compute initiative arrives in twelve months or thirty.
2023–2024: electricity becomes industrial policy by other means
The next choke point was power. Data-centre demand had been growing for years, but the concentration of AI workloads changed both scale and timing. High-density racks, larger clusters and greater cooling requirements collided with transmission constraints and local permitting delays. The International Energy Agency’s work on data centres and networks has documented the rising significance of electricity supply in digital infrastructure planning. By 2024, many advanced economies faced a familiar contradiction: they wanted more domestic compute, but their grids could not connect it quickly.
This changed the meaning of semiconductor sovereignty. A jurisdiction with robust chip policy but weak grid expansion could still lose out to another with faster interconnection and more permissive planning law. For operators, the effective scarcity was often not the accelerator itself but the energised building. For states, this transformed substation approvals, transformer procurement and transmission upgrades into strategic technology questions.
The issue was not simply megawatts in the abstract. It was time-to-power. A delayed grid connection can destroy the economics of a compute cluster long before any hardware becomes obsolete. It can also force reliance on foreign cloud regions that are politically uncomfortable but immediately available. In that sense, energy policy became an instrument of technological dependence as much as of decarbonisation or security.
2024: water and climate enter the sovereignty debate
As AI facilities expanded, scrutiny widened from power to water and climate resilience. Research on the water footprint of AI workloads, though still developing, helped push a once-marginal concern into mainstream infrastructure discussion. At the same time, increasingly severe heat and drought sharpened awareness that cooling assumptions cannot be treated as politically neutral engineering details.
A sovereign compute strategy that ignores hydrology is no strategy at all. Water rights, discharge permits, ambient temperature and local opposition now shape the siting of data centres and therefore the geography of available compute. In hotter regions, cooling costs and resilience burdens rise; in water-stressed regions, political legitimacy can evaporate quickly if public services appear to be sacrificed for server halls.
Climate therefore introduced a second-order sovereignty problem. Even where governments can finance compute expansion, they may not be able to maintain social licence for it, particularly when agriculture, households and industry face competing claims. The chokepoint is not only physical scarcity but administrative legitimacy.
In practice, sovereignty now lives in the queue.
Export control policy did not merely restrict chips; it reorganised geography.
2024–2025: cloud procurement exposes the legal gap
Many countries responded to hardware scarcity by renting compute rather than owning it. This was rational. It reduced capital expenditure, shortened deployment times and gave researchers access to frontier systems despite domestic shortages. But it also revealed a legal gap between access and control.
Rented compute may satisfy immediate capability needs while leaving unresolved questions about jurisdiction, service continuity, auditability and sanctions exposure. If the account, region, software layer or support channel remains ultimately subject to foreign law, then national capability exists only on sufferance. For universities and start-ups that may be acceptable. For defence, public administration, critical science and strategic industry, it is less comfortable.
The distinction here is subtle. Dependence is not binary. A country may be highly capable in chip design and still dependent for burst capacity, model training or emergency procurement. Conversely, a country with little indigenous chip manufacturing may achieve meaningful autonomy if it controls domestic data-centre capacity, power, trusted procurement and legal governance over sensitive workloads. This is why the older fab-centric discourse no longer captures the whole problem.
2025: planning law overtakes prestige policy
By 2025, the practical winners were often not those with the loudest sovereignty slogans but those able to compress the lead time from policy announcement to operational facility. Planning law, environmental review, land assembly and local grid coordination proved more decisive than many headline industrial packages.
This favoured states with either centralised administrative capacity or well-rehearsed infrastructure procedures. It disadvantaged places where semiconductor and AI policy sat in one ministry while power, water, land use and competition review sat in several others with no mechanism for rapid arbitration. In such systems, compute capacity could be pledged quickly but built slowly.
The implication is unromantic but important. Sovereignty in compute now resembles airport capacity or electricity transmission more than classic manufacturing policy. It requires competence in permits, sequencing, public consultation, construction logistics and network planning. Grand strategy is necessary; municipal process is decisive.
2025–2026: middle powers discover the politics of selective sufficiency
Most countries cannot replicate the entire semiconductor stack, nor should they try. By mid-2026, a more realistic doctrine had emerged among middle powers: selective sufficiency. The aim is not autarky but the ability to guarantee certain classes of compute for priority uses under stress. That may mean domestic data-centre reserves for public research, assured packaging access through alliance frameworks, power prioritisation rules, or procurement contracts that preserve operational continuity during export-control shocks.
This approach reflects the economics of the sector. The semiconductor supply chain remains globally specialised because it is difficult and expensive to duplicate. The question, then, is not whether interdependence can be abolished, but whether it can be structured. Selective sufficiency accepts dependence in low-risk layers while hardening the layers whose interruption would impair national decision-making.
That also changes what counts as success. A country need not lead at 2-nanometre logic to improve sovereignty if it can secure mature-node control electronics, robust networking gear, trusted packaging channels and predictable access to accelerator clusters for public-interest workloads. The geopolitical conversation has been slow to absorb this, preferring symbols of prestige over the dull mechanics of continuity.
In practice, sovereignty now lives in the queue.
2026: the queue becomes a policy arena
By mid-2026, queue management itself had become a strategic function. Governments increasingly sought visibility into where domestic researchers, hospitals, laboratories and industrial users stood in relation to available compute pools. Some built public compute programmes; others leaned on national champions or university consortia. But the common problem was allocation under scarcity.
The policy challenge is familiar from energy crises and pandemic logistics. When supply is constrained, market allocation alone may conflict with strategic priorities. A lucrative commercial customer can outbid a public laboratory; a foreign hyperscale tenant can occupy the power capacity a domestic science cluster needed. Sovereignty therefore turns into a governance problem: which uses are privileged, by what criteria, under whose authority, and with what transparency.
This is a difficult terrain for liberal economies because arbitrary intervention can deter investment. Yet complete passivity can leave critical functions exposed. The likely direction is not command-and-control but conditionality: public subsidy tied to domestic capacity commitments, emergency reservation clauses, location requirements for sensitive workloads, and stricter reporting on concentration risks.
What this means for semiconductors themselves
None of this makes fabs irrelevant. Fabrication capacity remains strategically vital, and leadership in process technology still confers formidable advantages. The point is narrower: fabrication is no longer sufficient as the organising metaphor for compute sovereignty. A state can pour money into silicon and still discover that its AI and high-performance-computing ambitions are stalled by packaging, transformers, water permits or foreign tenancy law.
The semiconductor debate is therefore broadening from manufacturing sovereignty to operational sovereignty. That widens the cast of actors. Utility regulators, environmental agencies, customs officials, procurement lawyers and local planners now shape national compute outcomes alongside chip designers and trade negotiators. Their decisions may look parochial. In aggregate, they determine strategic capability.
The uncomfortable conclusion is that the world did not enter an age of sovereign compute by making the supply chain simpler. It did so by multiplying the number of points at which compute can be delayed, denied, rerouted or conditioned. The glamorous choke point was the fab. The governing choke points, increasingly, are the permissions around it.
The post-fab view of sovereignty
This yields a distinct thesis for mid-2026. The geopolitics of compute is no longer best understood as a race to make the most advanced chip. It is better understood as a contest to control the permissions that turn components into usable, governable capacity. Those permissions include export licences, packaging access, energy interconnection, water allocation, planning approval and cloud jurisdiction.
That is a less cinematic story than a battle over lithography machines. It is also a more accurate one. States will continue to subsidise fabs and court chipmakers, because prestige and resilience both matter. But the countries that convert semiconductor policy into durable capability are likely to be those that master the mundane choreography connecting silicon to service.
The decisive bottleneck moved from transistor geometry to permissioning. That shift does not end chip politics. It domesticates it, dragging grand technological ambition into the slower, often contested machinery of administration. For sovereign compute, that machinery is now where power resides.
