THE SCENARIO: A regional power with a growing technology sector announces a national artificial intelligence strategy. The plan is ambitious: AI research centres, public-sector deployment programmes, and a domestic AI industry targeting global markets. Within six months, the strategy collides with an uncomfortable reality. Every component required to build and run frontier AI models — the advanced semiconductor chips, the fabrication facilities, the cloud computing infrastructure, the specialised software frameworks — is controlled by entities in three countries. The nation’s AI strategy is not a strategy at all. It is a procurement wish list directed at a supply chain it cannot influence, priced in a currency it does not control, and governed by export control regimes in which it has no vote.
When geopolitical tensions rise and chip export restrictions tighten, the nation’s AI ambitions are not defeated — they are simply rendered impossible. The data centres stand half-empty. The research teams lack the compute to train competitive models. The AI strategy, announced with great fanfare, becomes a document of aspirational dependence. The nation discovers that in the age of artificial intelligence, sovereignty is measured in floating-point operations per second — and those operations run on someone else’s hardware.
The Question
Can any nation claim technological sovereignty when the physical substrate of artificial intelligence — semiconductor fabrication, advanced compute, and the software stack that orchestrates them — is concentrated in the hands of three countries and a handful of corporations?
Part One: The Compute Supply Chain and the Illusion of AI Independence
Compute is the unexamined dependency at the foundation of every national AI strategy. While policy discourse focuses on data sovereignty, algorithmic fairness, and model governance — all important questions — the physical reality that makes AI possible at all receives remarkably little attention. That reality is stark: the global supply chain for advanced AI compute is one of the most concentrated industrial structures in the modern economy, and the concentration is increasing, not diminishing.
The chain begins with semiconductor design. The intellectual property for advanced chip architectures — the blueprints that determine how efficiently a processor can perform the matrix multiplications that underpin modern AI — is overwhelmingly held by a small number of firms. The design software used to create these blueprints is produced by an even smaller group. The electronic design automation tools are dominated by American companies. The instruction set architectures that define how software talks to hardware are controlled by a duopoly. Before a single transistor is fabricated, the design layer is already a dependency chain reaching back to a concentrated set of foreign entities.
Fabrication is the most visible choke point. The advanced manufacturing processes required to produce cutting-edge AI chips — currently at three nanometres and below — are commercially available from exactly one company: Taiwan Semiconductor Manufacturing Company. Samsung offers a competing node, but at lower volumes and with a narrower customer base. Intel is investing heavily but remains years behind in volume production for external customers. The entire world’s supply of the most advanced AI training chips passes through a single geography — Taiwan — where geopolitical risk is among the highest on the planet. Every nation training frontier AI models is dependent on the continued operation of fabrication facilities in a location that military planners consider a potential flashpoint.
The chip design and fabrication choke points are compounded by the equipment layer. The machines that make advanced chips — extreme ultraviolet lithography systems — are produced by a single company: ASML, headquartered in the Netherlands. These machines are among the most complex devices ever manufactured, incorporating components from thousands of suppliers across multiple countries. ASML itself is subject to export control regimes that restrict where its most advanced systems can be shipped. The equipment that fabricates the chips is itself a sovereignty dependency wrapped in geopolitical restrictions.
Above the hardware sits the cloud infrastructure layer. The hyperscale data centres that provide the compute capacity for training frontier models are operated by three American companies and one Chinese firm. These data centres represent cumulative capital investments measured in the hundreds of billions of dollars. The economies of scale are so extreme that building competitive alternatives requires state-level investment sustained over a decade or more. Even nations with significant financial resources find that the gap between their domestic cloud capacity and hyperscale infrastructure is measured in orders of magnitude, not percentages.
The software stack adds another dependency layer. The frameworks used to train AI models — the libraries that translate mathematical operations into efficient GPU instructions — are dominated by CUDA, NVIDIA’s proprietary parallel computing platform. While open alternatives exist, the performance gap, ecosystem maturity, and developer familiarity with CUDA create a switching cost that makes migration economically prohibitive for most organisations. The software that makes AI hardware useful is itself a proprietary dependency controlled by the same company that dominates the hardware market.
This concatenation of dependencies — design tools, chip architectures, fabrication processes, manufacturing equipment, cloud infrastructure, and software frameworks — creates a compute sovereignty deficit that renders most national AI strategies aspirational rather than operational. A nation can announce AI ambitions, fund research programmes, and train talent — but when it needs to actually run the computations that train a frontier model, it must ask permission from a supply chain it does not control. Permission is a poor foundation for sovereignty.
The compute dependency operates through three mechanisms that reinforce each other. The first is the innovation flywheel. Companies that control the fabrication capacity capture the revenue from advanced chip sales, which funds the next generation of research and development, which extends their technological lead, which increases the dependency of everyone else. This flywheel has been spinning for decades in the semiconductor industry and shows no sign of deceleration. Catching up is not a matter of matching current capability; it requires matching the rate of improvement while the leader continues to advance.
The second mechanism is the export control ratchet. Nations that dominate the compute supply chain use export controls as instruments of foreign policy, restricting which countries can access which levels of compute capability. These controls are dynamic — they tighten in response to geopolitical events and technological developments. A nation that designs its AI strategy around a particular level of compute access may find that access restricted before the strategy can be implemented. The export control regime introduces a political risk premium into every compute-dependent national strategy.
The third mechanism is the standards lock-in effect. The interfaces, protocols, and software ecosystems that dominate AI compute become de facto standards through sheer market presence. Applications are written to these interfaces. Developers are trained on these tools. Research is conducted on these platforms. Even if alternative hardware becomes available, the ecosystem switching cost — retraining developers, rewriting software, revalidating research — creates an inertial barrier that preserves the incumbent’s dominance long after technical alternatives exist.
Permission is a poor foundation for sovereignty. Yet permission is precisely what the global compute supply chain demands from every nation that seeks to train frontier AI models.
The Seven-Layer Stack Audit: Compute Sovereignty
Applying the Seven-Layer Sovereignty Stack to compute infrastructure reveals dependency at every level. At the physical infrastructure layer, the question is whether a nation possesses domestic data centre capacity sufficient for sovereign AI workloads. For the vast majority of nations, the answer is no. Hyperscale compute requires power infrastructure measured in gigawatts, cooling systems of industrial scale, and network connectivity at terabit speeds. Building this infrastructure requires not just capital but regulatory environments, construction supply chains, and operational expertise that take years to develop.
The network layer determines how data moves between compute nodes during distributed training. The interconnects that link thousands of GPUs into a coherent training cluster — technologies like InfiniBand and NVLink — are proprietary technologies controlled by a small set of vendors. The network fabric that makes distributed training possible is itself a dependency. Without sovereign access to high-bandwidth, low-latency interconnects, distributed training at frontier scale becomes impossible regardless of how many individual chips are available.
The data layer in compute sovereignty concerns the training data that flows through the hardware. While data sovereignty receives more policy attention than compute sovereignty, the two are linked: sovereign data processed on foreign-owned compute infrastructure is not truly sovereign. The compute provider has physical access to the data as it moves through memory, storage, and network interfaces. Privacy-preserving computation techniques can mitigate but not eliminate this exposure. Sovereign data demands sovereign compute.
The application layer encompasses the AI frameworks, model architectures, and training pipelines that run on the hardware. The dominance of a small set of frameworks — PyTorch and TensorFlow, with CUDA as the underlying execution engine — means that even if alternative hardware becomes available, the software ecosystem may not support it efficiently. The application layer dependency is subtle but profound: the tools that make compute useful are themselves non-sovereign.
The governance layer determines who decides how compute resources are allocated, which workloads receive priority, and under what conditions access can be modified or revoked. When compute infrastructure is cloud-based and foreign-operated, governance is exercised through terms of service, acceptable use policies, and credit limits — all determined unilaterally by the provider. A nation that relies on foreign cloud compute for its AI strategy has outsourced the governance of its most strategically significant computational resource.
The interoperability layer in compute sovereignty concerns how domestic compute infrastructure connects to international research networks, model repositories, and collaborative platforms. Isolation is not the objective; sovereign participation is. But participation requires standards compliance, and the standards are set by the same ecosystem that controls the hardware. The interoperability question is whether a nation can connect on terms that preserve sovereignty rather than surrender it.
The jurisdictional layer is the most legally complex. Compute infrastructure spans multiple jurisdictions: the chip fabrication occurs in one country, the cloud data centre in another, the software licensing in a third, and the end-user application in a fourth. When a dispute arises, which law applies and which court decides is often specified in contracts that the user nation had little power to negotiate. The jurisdictional layer of compute sovereignty is a patchwork of foreign legal regimes, any one of which can disrupt the compute supply chain through legal action, regulatory change, or sanctions enforcement.
Part Two: The Sovereignty Test Matrix — Compute
The five-domain Sovereignty Test Matrix applied to compute infrastructure produces sobering scores across every dimension. Political sovereignty in compute is measured by a nation’s ability to train and deploy AI models without dependence on foreign government authorisation through export controls. For nations outside the small group that produces advanced semiconductors, the political sovereignty score is effectively one out of five — every frontier training run requires chips that are subject to foreign export licensing. The political dimension of compute sovereignty is not about technical capability; it is about who holds the veto power over national AI development.
Economic sovereignty in compute is measured by the proportion of AI compute expenditure that remains within the national economy versus flowing to foreign hardware vendors and cloud providers. For nations without domestic semiconductor industries or hyperscale cloud providers, essentially one hundred percent of frontier AI compute expenditure leaves the national economy. This is not merely a trade deficit question; it is an economic dependency that creates a persistent outflow of national wealth to a small set of foreign corporations, with no pathway to domestic value capture within the compute supply chain.
Cultural sovereignty in compute concerns whether a nation’s AI models reflect its own cultural values, linguistic patterns, and knowledge systems — which depends on the ability to train models on domestic infrastructure. When training must occur on foreign cloud platforms, the training process is subject to the content policies, safety filters, and acceptable use restrictions of the foreign provider. A nation that wants to train models reflecting its cultural heritage may find that the available compute infrastructure imposes content restrictions that exclude that heritage. Cultural sovereignty in AI requires compute sovereignty as its enabling condition.
Intellectual sovereignty in compute is perhaps the most compromised dimension. The knowledge required to design, fabricate, and operate advanced AI chips is concentrated in a small number of firms and research institutions in a small number of countries. The fabrication process knowledge — the accumulated know-how of producing chips at three nanometres and below — represents decades of proprietary research and development. Replicating this knowledge base requires not just financial investment but the cultivation of entire scientific and engineering ecosystems. The intellectual sovereignty score for compute is near zero for all but a handful of nations.
Technological sovereignty in compute is measured by the ability to independently operate the full compute stack — from chip fabrication through cloud orchestration — within national borders and under national control. No nation currently achieves full technological sovereignty in compute. Even the nations that dominate parts of the supply chain — the United States with design, Taiwan with fabrication, the Netherlands with equipment — are dependent on each other for the complete stack. For all other nations, technological sovereignty in compute is not a near-term objective but a multi-decade strategic programme requiring sustained investment and international cooperation.
Part Three: The Red Flag Checklist — Compute Sovereignty
The compute sovereignty deficit manifests through eight red flags that indicate structural dependency on foreign-controlled compute infrastructure. The first red flag is the absence of domestic advanced semiconductor fabrication capability. A nation that cannot produce its own AI chips at competitive process nodes is dependent on foreign fabrication for every frontier training run. This red flag applies to more than one hundred and ninety of the world’s roughly one hundred and ninety-five countries.
The second red flag is reliance on foreign cloud providers for AI training workloads. When national AI research, government model development, and strategic AI applications run on hyperscale platforms operated by foreign corporations, the compute governance — including data access, model storage, and training prioritisation — is exercised by entities beyond national jurisdiction. The cloud provider’s acceptable use policy becomes, in effect, a regulatory instrument governing national AI development.
The third red flag is exposure to foreign export controls on semiconductor technology. Nations whose AI strategies depend on access to chips subject to foreign export licensing are building their technological future on a permission-based foundation. Export controls are dynamic policy instruments that can change with minimal notice in response to geopolitical developments. A strategy premised on continued access is a strategy premised on continued geopolitical alignment — a condition no sovereign nation should accept as a permanent constraint.
The fourth red flag is dependence on proprietary software frameworks with no sovereign alternatives. When the tools required to train AI models are controlled by foreign corporations, the ability to develop sovereign AI is contingent on continued access to those tools under terms set unilaterally by the provider. The absence of viable sovereign alternatives to the dominant AI software stack creates a dependency that is as binding as hardware lock-in.
The fifth red flag is the absence of domestic talent capable of operating the full compute stack. Compute sovereignty requires a workforce that understands semiconductor physics, chip design, fabrication processes, data centre engineering, distributed systems, and AI framework optimisation. Nations without educational pipelines producing this multidisciplinary talent are structurally incapable of achieving compute sovereignty regardless of their financial investment.
The sixth red flag is the dependence on foreign-manufactured fabrication equipment. Even if a nation were to invest in domestic chip fabrication, the machines required to build advanced fabrication facilities are produced by a single foreign company and are subject to their own export controls. The equipment dependency creates a recursive sovereignty problem: achieving chip sovereignty requires equipment sovereignty, which requires chip sovereignty. Breaking this recursion requires international cooperation or a multi-generational industrial programme.
The seventh red flag is the absence of domestic power and cooling infrastructure sufficient for hyperscale AI compute. Frontier AI training consumes electricity at the scale of small power plants. The data centres that house training clusters require cooling systems of industrial magnitude. Nations without the energy infrastructure to support sovereign compute are, regardless of their other capabilities, unable to operate independent AI training at frontier scale.
The eighth red flag is the lack of a documented compute sovereignty strategy with measurable milestones and allocated funding. The absence of a strategy is itself a red flag. Nations that have not formally assessed their compute sovereignty deficit, established reduction targets, and allocated multi-year funding are, by default, accepting perpetual compute dependency as a structural condition of their AI development. When three or more of these red flags apply — and most nations trigger seven or all eight — the compute sovereignty deficit is critical and demands strategic response.
A strategy premised on continued access to foreign-controlled compute is a strategy premised on continued geopolitical alignment — a condition no sovereign nation should accept as a permanent structural constraint.
Part Four: The Phased Implementation Framework — Compute Sovereignty
Reclaiming compute sovereignty is the most capital-intensive sovereignty programme any nation can undertake. Unlike data sovereignty, which can be advanced through legislative and regulatory action, compute sovereignty requires physical infrastructure, industrial capability, and scientific ecosystems that take decades to build. The phased implementation framework acknowledges this reality and sequences interventions from the most achievable to the most ambitious, delivering incremental sovereignty gains at each phase while building toward long-term independence.
The assessment phase, spanning weeks one through four, establishes the compute sovereignty baseline. This phase requires a comprehensive audit of national AI compute consumption: what hardware is being used, where it is physically located, who operates it, what software stack runs on it, and under what legal and contractual terms access is provided. The audit must map every significant AI training workload to its compute provenance — the chain of dependencies that made that training run possible. The output is a Compute Sovereignty Deficit Report that quantifies the dependency gap across all seven layers of the stack and all five domains of the sovereignty matrix.
The assessment phase also identifies the most immediate sovereignty risks. Which critical AI workloads are running entirely on foreign-controlled infrastructure? Which government applications depend on cloud platforms subject to foreign export controls? Which research programmes would be interrupted if chip access were restricted tomorrow? These risk-mapped dependencies become the priority targets for the strategic planning phase.
The strategic planning phase, spanning months two and three, develops a multi-track compute sovereignty roadmap. The first track addresses immediate risk mitigation: migrating the most sovereignty-sensitive AI workloads to the most sovereign compute infrastructure currently available, even if that infrastructure is less performant. The principle is that a sovereign model running on last-generation hardware is preferable to a frontier model running on foreign-controlled infrastructure for applications involving national security, critical infrastructure, or citizen data.
The second track addresses domestic compute capacity building. This involves investment in national data centre infrastructure, procurement of AI hardware at the highest performance tier available without export control restrictions, and development of the power and cooling infrastructure required to support sovereign compute. The strategic planning phase must produce detailed specifications for the sovereign compute architecture: the number of data centres, their locations, their power requirements, their network connectivity, and their cybersecurity architecture.
The third track addresses international cooperation. No nation can achieve full compute sovereignty in isolation — the semiconductor supply chain is too complex, too capital-intensive, and too globally distributed for autarky to be feasible. The strategic planning phase must identify coalition partners — nations facing similar compute sovereignty challenges — and develop cooperative frameworks for shared fabrication facilities, pooled research and development, and mutual recognition of sovereignty standards. Compute sovereignty in practice will be achieved through sovereign alliances, not sovereign isolation.
The implementation phase, spanning months four through twelve, executes the first wave of sovereignty investments. Domestic data centres come online. Sovereign AI workloads migrate from foreign cloud platforms to national infrastructure. Training programmes produce the first cohorts of compute infrastructure engineers. Procurement agreements with chip vendors secure multi-year hardware supply with sovereignty-compatible terms — including physical delivery of hardware to domestic data centres, audit rights, and contractual commitments to continued supply regardless of export control changes.
The implementation phase also includes the development of sovereign software alternatives to the dominant AI frameworks. This does not require building a replacement for CUDA from scratch — an undertaking of prohibitive complexity — but rather investing in the open-source alternatives that already exist, contributing to their development, and ensuring they are sufficiently performant for sovereign workloads. The software sovereignty track is a contribution strategy, not a replacement strategy: building capability within the open ecosystem rather than attempting to replicate proprietary stacks.
The institutionalisation phase, spanning months thirteen through eighteen, embeds compute sovereignty as a permanent national capability. A National Compute Sovereignty Authority is established with responsibility for ongoing assessment, investment planning, and international coordination. A Compute Sovereignty Fund provides multi-year, multi-administration funding that insulates the programme from political cycles. Educational pipelines from secondary through postgraduate levels build the talent base for long-term sovereignty. And a Compute Sovereignty Index, published annually, provides public accountability for progress against sovereignty targets.
The phased framework acknowledges a fundamental truth about compute sovereignty: it cannot be achieved quickly, but it can be achieved through sustained, strategically sequenced investment over a decade or more. The nations that begin now will have sovereign compute infrastructure in the early twenty-thirties. The nations that delay will find themselves further behind, their AI strategies increasingly dependent on a supply chain that is growing more concentrated, not less. The question is not whether compute sovereignty is difficult — it is difficult. The question is whether the alternative — perpetual compute dependency — is acceptable for any nation that takes its sovereignty seriously.
The Question Revisited
Can any nation claim technological sovereignty when the physical substrate of artificial intelligence — semiconductor fabrication, advanced compute, and the software stack — is concentrated in the hands of three countries and a handful of corporations? The answer, examined through the TEE Method framework, is that technological sovereignty in compute exists on a spectrum, not as a binary condition. No nation achieves absolute compute sovereignty. But the nations that are furthest along — those that have invested in domestic fabrication, built sovereign cloud infrastructure, developed domestic talent pipelines, and formed international coalitions for shared capability — exercise a qualitatively different level of agency over their AI futures than nations that have accepted compute dependency as a permanent structural condition.
The TEE Method provides the analytical discipline for navigating this spectrum. Test the compute stack at every layer against sovereignty criteria — not against an absolute standard of independence, but against the standard of whether dependency creates unacceptable strategic vulnerability. Evaluate the findings against the five-domain matrix, identifying which sovereignty dimensions are most compromised and which interventions would deliver the greatest sovereignty gain for the investment required. Evolve the architecture through sustained, sequenced investment, building domestic capability while engaging internationally to avoid the isolation that makes absolute sovereignty impossible.
Compute is not just another technology input. It is the industrial foundation of the AI era, as fundamental to the twenty-first century as steel was to the nineteenth and oil to the twentieth. The nations that control their compute control their AI. The nations that control their AI control the cognitive infrastructure of their economies, their governments, and their societies. Compute sovereignty is not a technology question. It is the central strategic question of the coming decade.
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework, the Seven-Layer Sovereignty Stack Audit methodology, and the Sovereignty Test Matrix, see tonishatagoe.com.