Part 1 of a three-part series on the sovereign AI trap.
There is a version of digital sovereignty worth defending. Public Digital’s forthcoming book Digital Sovereignty: The Power to Decide defines it as an organization making deliberate, informed choices that build its digital future by design.
That is a useful idea. It applies to a health ministry that cannot link hospital admissions data to its pharmacy logs, to a school administrator staring at 756 pages of vendor terms, to an NGO director reading the fine print on a free platform. Agency over digital choices is the problem statement of our era.
Sovereign AI is something else.
It is the idea that countries need their own domestically controlled large language models, trained on their own compute, hosted on their own sovereign clouds, rather than relying on global models from American and Chinese vendors.
It has been adopted as policy by Canada, India, the European Union, and as of 2025, Nigeria, Egypt, Kenya, and South Africa. The rhetoric borrows from the agency frame. Neither the economics nor the compute layer of sovereign AI survive contact with reality.
Sign Up Now to explore how sovereign AI is a trap
The vendor wrote the policy
The single loudest advocate for sovereign AI is not a head of state. It is Jensen Huang, the CEO of Nvidia, which sells the GPUs that every sovereign AI project needs to buy.
- In February 2024, Huang stood at the World Governments Summit in Dubai and told 4,000 delegates from 150 countries that “every country needs sovereign AI”.
- He told the UAE’s AI minister that codifying national language and culture into a domestic LLM is “not that costly, it is also not that hard.”
- He has repeated the line in Thailand, Japan, Denmark, Singapore, South Korea, France, and the UK.
The Financial Times reported that Nvidia is simultaneously encouraging countries to develop sovereign AI and supplying the chips those programs need. Huang has predicted the global installed base of data centers will double to $2 trillion over five years.
This is not a conspiracy. It is a sales channel. The framing is sovereignty. The product is GPUs. The buyer is the public purse.
Governments advocating sovereign AI
Africa’s four largest tech economies have each released or drafted sovereign AI policies since January 2025, and they are doing the development sector a favor by admitting in writing exactly what the dependency looks like. The Rest of World reporting by Ananya Bhattacharya collects the language.
- Kenya‘s policy names Microsoft, Google, Meta, and Nvidia as investors and acknowledges that its AI ecosystem heavily relies on funding and support from international organisations and private companies, which may limit the sustainability and autonomy of local AI initiatives.
- Nigeria‘s strategy lists Google, Microsoft, Meta, and Amazon as stakeholders in formulating the strategy itself, and frames its goal as reducing the cost and dependence on virtual AI environments.
- Egypt‘s draft policy aims to build more than 250 local AI companies and pivot away from importing AI solutions, with university programs co-designed with international firms.
- South Africa‘s draft policy identified foreign infrastructure as a data privacy and security risk and then had to be withdrawn in April 2025 because the AI tools used to help write it generated fake citations.
That last sentence is not editorial commentary. It is the most efficient summary of the sovereign AI debate currently in print. The country writing the policy could not produce the policy without using the foreign tools the policy is trying to escape, and the tools fabricated their own evidence.
The math is not close
The strongest version of the sovereign AI case is straightforward. Dependence on American and Chinese model providers is a real strategic risk, and building domestic capability is the only way to escape it. Now look at the weakest point.
The weakest point is scale.
Sam Altman confirmed that GPT-4 training cost “more than” $100 million. GPT-5 reportedly runs $500 million per training run. The four largest US hyperscalers are projected to spend $690 billion on capex in 2026 alone. Amazon alone has guided to $200 billion. Goldman Sachs projects cumulative hyperscaler capex from 2025 to 2027 at $1.15 trillion.
- EuroHPC’s roughly $2 billion total represents less than half of one percent of what four American firms will spend in two years.
- Canada’s $2 billion strategy and India’s $1.25 billion IndiaAI mission are rounding errors at hyperscaler scale.
A country that tries to compete head-to-head at the foundation model layer is committing public capital to a race where the favorites have a thousand-to-one head start and are widening the gap by 72 percent per year. Good luck with that.
The stack is the sovereignty question
Hannah Cooper Klein made this point on a Global Digital Health Network webinar I moderated recently. AI is not just a model. AI is compute, cloud, chips, data centers, energy, procurement power, cybersecurity, and governance.
The model layer is what gets demoed at conferences. The other seven layers determine whether the model layer can be used responsibly, sustainably, or at all.
Africans make up 18 percent of the world’s population and the continent has less than 1 percent of global data center capacity, according to the World Economic Forum. The top five African markets combined have less capacity than France had in 2024, McKinsey found. As of June 2025, high-income countries held 77 percent of global colocation data center capacity, while low-income countries held less than 0.1 percent.
A government that announces a sovereign LLM program is announcing a project that will run on foreign GPUs, in foreign data centers, on foreign cloud platforms, governed by foreign procurement terms it does not have the leverage to renegotiate. The sovereignty exists at the model fine-tuning layer. Every other layer points outward.
The familiar pattern
We have seen this before.
- Data localization was sold as sovereignty and delivered surveillance infrastructure.
- National digital ID programs were sold as inclusion and routinely became tools for exclusion.
- Free Basics was sold as connectivity and was a market capture play.
Sovereign AI fits the pattern. A politically attractive promise. A vendor pushing it hardest. A development community providing rhetorical cover. A bill that lands on national budgets that cannot afford it.
Sovereignty is the wrapper. Dependence is the contents.
If sovereign AI is the trap, what does the escape route look like? That is the subject of next week’s post on small language models and edge inference as the actual path out.
