Every few decades, governments rediscover the same idea: if a technology is strategic, the state should pay to own it at home. In the 2010s and early 2020s, the technology was semiconductors, and the result was the largest coordinated industrial subsidy push since the Cold War. In 2026, the technology is artificial intelligence, and the vocabulary has changed to "sovereign AI." The economics have not changed at all.
Strip away the language of sovereignty, autonomy, and national champions, and what remains is a familiar structure: capital grants, tax credits, state co-investment vehicles, and subsidized access to scarce inputs, all deployed to localize a capital-intensive technology stack inside national borders. That is the fab subsidy playbook, reprinted with a new cover. And it carries the same risks the fab race produced: escalating state outlays, chronic underdelivery against headline numbers, and a looming overhang of subscale, fragmented capacity.
This is not an argument against public investment in AI. It is an argument that policymakers are repeating a specific, recent, well-documented mistake, and that the data from the semiconductor precedent tells us how this movie ends.
The playbook is identical to chip subsidies
The most striking feature of sovereign AI programs is how little institutional innovation they contain. The instruments are the ones governments built for fabs: direct grants for facility construction, investment tax credits for equipment, state equity stakes in national champions, and preferential access to state-controlled resources, whether land, power, or compute.
The semiconductor precedent shows exactly where this instrument mix leads. OECD analysis of firms in its MAGIC database records a notable increase in government grants in 2023, reflecting the introduction by several governments of support schemes aimed at encouraging construction of new fabrication facilities; contract foundries received the most subsidies relative to their revenue, followed by integrated device manufacturers. In other words, once governments started competing to attract fabs, subsidy intensity climbed fastest exactly where capital intensity was highest, and support concentrated on a handful of anchor firms rather than lifting the broader ecosystem.
Sovereign AI is reproducing that concentration dynamic in real time. State money flows to national compute champions, flagship data center projects, and a small cohort of anointed model developers. The subsidy escalation curve that took semiconductors a decade to trace is being compressed into a much shorter window, because every government can see every other government's announcements, and no one wants to be the capital that underbid.
The money is already at chip-race scale
If sovereign AI were a modest complement to private investment, the subsidy-race framing would be overwrought. It is not modest. The commitments announced in the past eighteen months already rival, and in some cases exceed, the national semiconductor programs that defined the chip race.
The program-level figures speak for themselves. France has assembled commitments of roughly EUR 109 billion in AI investment pledges. Japan's government-backed national AI consortium will receive about 387.3 billion yen, roughly $2.4 billion, in subsidies this fiscal year and approximately 1 trillion yen, around $6.1 billion, over five years. The United Kingdom has launched a GBP 500 million national unit to back domestic AI companies, alongside a GBP 282 million programme to fund shared strategic AI assets such as high-value datasets. Canada has committed approximately $890 million to build a national AI supercomputing system under its Sovereign AI Compute Strategy.
For comparison, the US CHIPS and Science Act, the single largest semiconductor intervention in the West, allocated $52.7 billion, and took years of political wrangling to pass. Sovereign AI programs have reached comparable aggregate scale in roughly two years, with far less scrutiny, because "AI leadership" currently functions as a political trump card that closes debate rather than opening it. Speed of escalation is itself a warning sign; subsidy races accelerate when governments respond to each other's announcements rather than to underlying economics.
Sovereignty is being conflated with self-sufficiency
The deeper design flaw is conceptual. Sovereignty, properly defined, is about strategic control: the ability to keep operating, to govern AI according to national values, and to avoid single points of foreign dependency. Self-sufficiency is something else entirely: owning every layer of the stack, from silicon to applications, inside national borders. Most sovereign AI programs claim to pursue the first while budgeting for the second.
This is not a fringe critique. The World Economic Forum's January 2026 analysis of AI sovereignty finds that several economies have attempted to compete by owning the entire AI value chain, from raw materials to AI-based applications, and concludes that based on investment patterns, "AI sovereignty" has been conflated with self-sufficiency; it argues explicitly that this is not the only path to AI competitiveness.
The semiconductor race made the same conflation, and it was the single most expensive mistake of that era. No country, including the United States, achieved chip self-sufficiency, because the value chain is irreducibly global: lithography from the Netherlands, design tools from the US, advanced packaging from Asia. AI is even less separable. Frontier models, accelerator hardware, training data, and talent flow across borders by default. A national program that prices full-stack ownership is buying an outcome that the structure of the industry does not permit at any price a mid-sized economy can pay.
The precedent shows underdelivery, not capability
Suppose we set aside design questions and ask a simpler one: when governments announce these numbers, what actually gets delivered? The chip race provides a clean natural experiment, and the answer is uncomfortable.
The European Chips Act, adopted in 2023, was intended to mobilize public support comparable to the US program, yet as of early 2025 only EUR 13.75 billion in state aid had been approved, while the US Department of Commerce had awarded $33.7 billion in grants plus $5.5 billion in loans; independent analysis concludes the act has underdelivered both in the scale of funding mobilized and in the strategic coordination of its deployment, aggravated by what the OECD describes as an ongoing subsidy race.
Note what this comparison shows. Even the better-performing program, the US one, delivered awards well below its headline over its first years, and the EU program delivered roughly a third of its American counterpart despite similar ambitions. Headline numbers in subsidy races are announcements, not appropriations; they are discounted by permitting delays, state-aid procedures, co-funding shortfalls, and shifting political priorities.
Sovereign AI inherits every one of these frictions, plus new ones. Compute infrastructure depends on grid connections with multi-year queues, on power prices that several European governments cannot control, and on GPU supply governed by one dominant vendor's allocation decisions. When national AI strategies are audited in 2029, expect the same pattern the chip race produced: delivered capability at a fraction of announced ambition, with the gap widest in the economies that announced loudest.
Fragmentation destroys the returns
The final problem is arithmetic. AI infrastructure economics are brutally scale-dependent: utilization rates, energy contracts, hardware refresh cycles, and model training runs all reward concentration. A subsidy race pushes in exactly the opposite direction, toward dozens of parallel national stacks, each built for political visibility rather than economic efficiency.
The concentration data makes the problem concrete. The WEF's cumulative investment estimates, spanning electricity capacity, silicon processing, chip manufacturing, data centres, foundation model training, and application development, show spending overwhelmingly dominated by a small set of economies, with a long tail of countries grouped into "rest of world". The national programs multiplying across that long tail are subscale by orders of magnitude relative to the leaders. A $900 million national compute program is a rounding error against private hyperscaler capex, which is measured in the hundreds of billions annually.
Subscale capacity is not merely inefficient; it is stranded-asset risk. GPU fleets depreciate fast, and a national supercomputer that runs below utilization because the domestic ecosystem cannot fill it is a monument, not an asset. The chip race left behind exactly this residue: announced fabs delayed, downsized, or quietly cancelled once subsidy terms met market reality. Sovereign AI's equivalent will be half-utilized national clusters, still drawing subsidized power, defended in budget hearings as strategic.
What a smarter sovereignty would buy
None of this means governments should stand aside. It means the objective function is wrong. The lesson of the chip race is that indispensability beats self-sufficiency: economies won durable leverage by dominating specific chokepoints, not by replicating the whole stack. The Netherlands did not build a national semiconductor industry; it built a chokepoint in lithography.
Applied to AI, that logic points away from national full-stack replication and toward three cheaper, higher-leverage plays. First, specialize: fund the layers where the economy has genuine comparative advantage, whether sovereign datasets in regulated domains, domain-specific models, or energy-advantaged compute for others to rent. Second, pool: shared compute across allied economies captures the scale economics that national fragmentation destroys. Third, secure optionality rather than ownership: contractual guarantees, multi-vendor strategies, and trusted-partner agreements deliver most of what sovereignty actually requires at a fraction of the capital cost.
The subsidy race will not stop because this argument is right. Races have their own momentum, and no elected government wants to explain why it declined to fund the defining technology of the decade. But investors, corporate strategists, and the analysts who will eventually audit these programs should be clear-eyed now: sovereign AI, as currently constructed, is the fab race with better branding. The precedent is not encouraging, and this time we cannot claim we did not have the data.