Jun 28, 2026
Three professionals in traditional and modern attire collaborate on a project in a contemporary office setting. by Mikhail Nilov via Pexels
.
AI sovereignty: Collaborate globally, govern locally
.
.
Governments across the wealthy world are drafting national strategies for artificial intelligence, and nearly all of them approach sovereignty the same way: as something a country builds and buys.
Compute clusters, sovereign cloud, domestic energy, national champions, a venture fund to convert research into firms. The logic is coherent and, on its own terms, sound. Prosperity and security increasingly belong to nations that can build and govern AI rather than rent it.
The instinct to control the machine is not the error. The error is what the approach leaves out: the capacity to govern knowledge, which no amount of hardware supplies.
The strategies are written around an economic priority, and that priority is real – but it is a short-horizon reading of it. A country can attract the data centre and train the workforce and still find, a decade on, that the science it produces is analysed, owned and monetised elsewhere.
Knowledge governance is the capacity that protects the long return on exactly the economic bet these strategies are making. Leaving it out is not a competing vision of the economy; it is the part of the same vision that pays out later, and that the rush to stand up infrastructure now tends to discount.
Canada’s ‘AI for All’ strategy
Canada’s ‘AI for All’, launched in June 2026, is among the most sophisticated of these strategies, which is precisely why it shows the pattern cleanly.
It is candid about the country’s dependencies and serious about closing them.
It assigns universities four jobs, and all four point downstream: a literacy engine training a million students, colleges as applied-AI upskillers, institutions as nodes in workforce alliances aligned to industrial demand, and research universities as the origin point for AI-native companies fed by capital.
Each role is legitimate. None is the role on which sovereignty actually turns. Read together, they reduce the university to a pipeline – a supplier of talent and intellectual property to an economy that someone else governs.
This is not a Canadian failing. The OECD’s recent work on science and innovation describes member countries reorganising research policy around competitiveness, industrial strategy and national security, and within that shift valuing universities for two outputs only: the workers they train and the research they commercialise.
The pipeline view is becoming the default way the wealthy world understands what a university is for. What that view cannot see is the function on which sovereignty actually rests.
The function the strategies omit
Call it knowledge governance: the capacity to produce research, decide the terms on which it crosses borders, and capture the value it generates. It is the difference between a science system that controls its own knowledge cycle and one that merely feeds someone else’s.
A country that governs its science moves through the whole cycle – it collects the data, analyses it, publishes first, and captures the value. A country that cannot, becomes a supplier of raw material. It generates biodiversity records, genomic data, climate observations, and loses control over who analyses them, who publishes, and whether the findings ever serve local priorities.
As I argued in a technology profile on connectivity and digital sovereignty in the Global South, written for the International Science Council’s Centre for Science Futures, sovereign infrastructure determines whether institutions can conduct, analyse, publish and benefit from their own research, or whether they remain field stations generating data for processing elsewhere.
That was written about the Global South, but the mechanism is universal and it transfers directly to AI. Data is the bloodline of the system, and governance is decided less by who owns the hardware than by who controls the country’s data, identities and research environments – the platform that authenticates a researcher, the cloud that stores a dataset, the environment in which a collaboration takes place.
A recent review of 775 non-United States data centre projects found that US companies operate close to half of them when weighted by investment value.
The study concluded that building a data centre on home soil does not secure digital sovereignty if a foreign entity runs it: the operator’s nationality becomes a jurisdictional hook, letting its home government compel access to the data inside whatever country it sits in.
A nation can own the compute and still cede the science if its universities authenticate, store and collaborate on platforms configured and operated elsewhere.
Owning the machine is not the same as governing what is done on it. This is the gap in AI national strategies, and it is a strange one: a document can detail a university’s place in alliances and standards bodies while saying nothing about the systems its laboratories actually run on.
That silence is where sovereignty is conceded – the identity platform, the cloud tenancy, the data environment chosen years ago for convenience and never revisited, each one a governance decision made by default in favour of whichever vendor arrived first.
Here a distinction matters. The argument is not that universities should run national AI policy; they will not, and claiming otherwise would overclaim. The steering of AI – what gets funded, deployed, regulated, sold – runs through firms, ministries and capital, and the university sits at the periphery of that steering, despite the research and trained people it supplies to all of it.
Nor is the claim that universities will train frontier models; the capital required for state-of-the-art clusters has largely priced them out, and the foundational models of 2026 are built in private labs, not faculties.
But knowledge governance is a different function from strategic direction, and it is the one the strategies have left unassigned. No firm holds it. No ministry can manufacture it.
It lives, latent and unfunded, in the institutions that produce the open, public-interest knowledge a country cannot buy back once it has been ceded – the methods, the datasets, the trained researchers, the science that was never anyone’s product.
The pipeline view does not just undervalue the university; it leaves the governance function homeless.
Why this begins abroad
And here is the part the AI strategies invert most completely. Having decided that sovereignty is built by accumulating infrastructure inside one’s own borders, they treat international engagement as the channel through which dependency arrives – foreign cloud, foreign models, foreign recruitment – and route whatever ambition remains through trade missions and national firms. The university as an international actor in its own right disappears.
But the capacity these strategies want cannot be built behind a border. AI is possible at all only because of an open global knowledge system, in which researchers share methods and solve problems that belong to no single nation.
Knowledge governance worth the name, therefore, does not begin with domestic control and reluctantly admit collaboration; it begins with collaboration and builds control on top of it. International engagement is not a risk to sovereign AI. It is the precondition.
The instruments already exist, and they are institutional and international at once. The hardest gap for most countries is compute itself – the frontier clusters are scarce, expensive and, as the US restrictions on advanced chips to China have made plain, subject to control by whoever holds the supply.
But a nation that cannot build its own cluster is not therefore shut out, because the networks are what make scarce compute reachable and keep the data that runs on it under local terms.
RedCLARA connects the national research and education networks of Latin America, linking their universities to GÉANT in Europe, Internet2 in the United States, and partner networks in Africa – the UbuntuNet Alliance and WACREN among them.
Through that infrastructure it provides federated identity, dedicated high-speed circuits between laboratories, and secure environments for large-scale data exchange. AfricaConnect ties the continent’s regional networks into the same global fabric.
The logic is identical in each case: institutions that could never individually afford intercontinental cables or computing clusters pool their demand, build jointly, and meet dominant providers from greater collective strength – owning capacity rather than renting it.
Europe’s digital sovereignty
These networks are how nationally funded supercomputers are reached and allocated across institutions that could never each own one, and how the datasets those machines run on stay under terms a country sets rather than rents. They are the layer that makes sovereign compute usable, and the reason a country short on compute is not thereby short on sovereignty.
The Global North understands this perfectly when its own autonomy is at stake.
GÉANT, the body that operates Europe’s pan-continental research and education network, defines one of its strategic pillars as maintaining control over intercontinental connectivity in support of European digital sovereignty, and European policy analysts treat research-network infrastructure as a deliberate instrument for reducing dependence on US and Chinese suppliers.
Europe has gone further than rhetoric: EuroHPC pools national money into shared supercomputers, and the European Open Science Cloud builds the federated data layer to match. Most national AI strategies have not followed, even as their own universities sit as nodes in the same federated system and could be resourced to make it carry sovereign AI.
The point holds with most force where resources are scarce. For a wealthy country, these networks amplify a system that already works.
For an under-resourced one, they are the precondition for engaging the global system at all – a university without the capital to build its own compute can still work at the frontier as a node in a shared network, and cannot if it stands outside one.
That a university cannot afford a frontier cluster is the case for federation, not against it: pooled demand is how institutions reach compute none of them could buy alone, and the network is what keeps their data under local terms while they reach it.
So the instruction reverses depending on where you stand. The wealthy country is told international engagement is a dependency to manage; the under-resourced one finds it is the only road to the capacity in question.
For these countries the strategies are not merely incomplete but backwards: the collaboration they treat as a vulnerability is the single route to the sovereignty they say they want, because there is no domestic substitute for it to fall back on.
The harder truth
More collaboration is not automatically better.
The same OECD work that records the securitisation of science also records its cost: international collaboration has lost momentum after three decades of growth, and a chilling effect now pushes institutions to avoid flagged partnerships on thin guidance and researchers to steer clear of important but high-risk fields.
A blanket application of research security measures, the OECD warns, threatens the quality, productivity and integrity of the national research system. Sovereignty pursued through walls has well-documented failure modes.
The remedy is to build the instrument well, not to set it down.
The security-first case deserves a straight answer. That case runs as follows: open standards are how sensitive work bleeds to adversaries, so walls are a necessary quarantine. The answer is that federation governs one dimension of the problem and not the whole of it.
Interoperability and control are not opposites. Federated infrastructure is what makes selective control possible – data held in local custody rather than on a foreign vendor’s servers, access governed institution by institution, sensitive environments segmented from open ones on shared foundations.
It does not, on its own, address the parts of research security that have nothing to do with where data sits: researcher vetting, dual-use fields with direct military application, intellectual property that leaves through a person rather than a server. Those need their own instruments, and federation is no substitute for them.
But on the dimension it does govern, the logic holds: a country that runs its science on systems it does not control has no quarantine to offer; it has already exported the thing it means to protect. The choice is not between openness and security. It is between governing the terms of exposure and not knowing what they are.
That distinction sets the two kinds of sovereignty apart. One breaks interoperability and isolates. The other leaves the shared foundations intact and governs what is built on them – global connectivity on locally governed terms, which is the footing sovereign AI actually requires. The remedy is not less internationalisation but internationalisation built for equity rather than extraction.
That is the version of sovereignty these strategies keep missing. They locate it in infrastructure they can announce and capital they can attract, and treat universities as suppliers and international engagement as a threat to manage.
Buying technology is the visible move. The decisive work is institutional and, paradoxically, external: universities capable enough, connected enough and trusted enough to operate at the frontier on terms they help set.
Sovereignty in a domain with little respect for borders is not won by building higher ones. It is won by science systems that can collaborate globally while governing locally – a capacity that does not begin at home, and that lives, in every country now drafting one of these strategies, in the institutions those strategies have reduced to a pipeline.
Carlos Vargas is the founder of Societas Partnerships, a higher education advisory firm based in Panama City, and the author of a technology profile on connectivity and digital sovereignty in the Global South written for the International Science Council’s Centre for Science Futures. He previously spent 14 years in senior internationalisation roles at the University of Toronto, Carleton University and the University of Calgary in Canada.
.
.
