In short
Governments worldwide are funding "sovereign" AI: national compute, models and data that they control. Supporters argue it secures access, reflects local values and could be built to be efficient from the start. Critics argue full self-sufficiency is unaffordable, duplicates effort and adds data centres to strained grids. The sustainability outcome depends less on who owns the AI than on how it is built, sized and used.
Key points
- A 2026 index counted 184 sovereign AI projects across 67 countries.
- The UK, the EU and Switzerland are among those investing in national compute and models, at very different scales.
- The case for sovereign or public AI rests on security, equitable access, accountability and alignment with local law and values.
- The case against rests on cost, feasibility, duplication of energy-intensive training and the risk of political control.
- Small, local and open models offer a middle path that can serve sovereignty and efficiency together.
AI is increasingly treated as national infrastructure, like energy or telecoms. Many governments now want some AI capability under their own control, and they are spending heavily to get it. This raises a question that is often missed in the debate: is sovereign AI better or worse for the environment? This article sets out the arguments on both sides.
What is sovereign AI?
AI capability that a country, or a bloc of countries, controls. That can mean its own compute (data centres and supercomputers), its own models, its own data, its own skills, or some mix.
The idea has spread quickly. The Center for a New American Security's Sovereign AI Index counted 184 sovereign AI projects across 67 countries in data to June 2026.
A few examples, at very different scales:
| Who | What |
|---|---|
| United Kingdom | A Sovereign AI Unit with up to £500 million, launched in 2026, alongside national compute such as the Isambard-AI supercomputer and designated AI Growth Zones |
| European Union | InvestAI, announced in February 2025 to mobilise €200 billion for AI, including a €20 billion fund for AI "gigafactories"; selection of gigafactory sites was still under way in late 2026 |
| Switzerland | Apertus, a fully open multilingual language model released in September 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre, in 8 and 70 billion parameter versions, with weights, data and code published |
What is the case for sovereign or public AI?
Supporters argue that AI is too important to depend entirely on a few foreign companies. Their main arguments:
- Security and resilience. Services that governments, health systems and grids rely on should not depend on access that another country or company could restrict.
- Equitable access. Some argue for AI as a public utility, like water or rail, so that access does not depend on ability to pay. The Mozilla Foundation's 2024 Public AI paper argues for public alternatives alongside private AI, comparing them to public broadcasting and public transport.
- Law and values. A model trained on data that reflects local languages, law and norms may serve citizens better. Governments that build on open models trained mainly on other countries' data face the question of whose values those models reflect. The UK government, for example, has funded work on public-service tools building on open models such as Meta's Llama.
- Accountability. Public models can be inspected, audited and governed in public.
- Sensitive data. Some data, such as detailed electricity network or health data, cannot be sent to outside providers, so efficient local models are the only way to use AI on it.
A 2026 UK green AI sandpit of mathematicians, computer scientists and industry debated a version of this: a government-run national model, operated as a utility for equity of access, sovereignty and security, with mathematical efficiency as its competitive edge.
What is the case against?
Critics argue that full self-sufficiency is unaffordable, slow and, for most countries, impossible. Their main arguments:
- Cost and feasibility. The Tony Blair Institute's 2026 report argues that full self-sufficiency is too expensive, too slow and, for most countries, simply impossible, and frames sovereignty instead as agency over key choices. A 2026 Brookings report argued that full-stack AI sovereignty is structurally infeasible for almost any country.
- Duplication. Training a frontier model from scratch in every country repeats energy, water and hardware costs that have already been paid elsewhere.
- Political control. A state-run model raises concerns about government influence over what citizens can ask and are told.
- Data limits. A democratic government may not be able, or willing, to train on data that private companies use.
- Infrastructure and public opinion. New data centres need grid connections, water and planning consent, and face local opposition in many places.
- Betting on the wrong architecture. A national investment in a large central model assumes such models will still dominate in five or ten years, rather than smaller local models where data stays with the user.
What does sovereignty mean for energy and emissions?
It depends on what is built, where, and what it replaces. Sovereign AI can go either way:
- Worse: if it means more data centres on fossil-heavy grids, duplicated training runs and large general models used for tasks small ones could do.
- Better: if it means right-sized, specialised models, siting data centres where clean power is in surplus, reusing open models instead of retraining from scratch, and setting public standards for measurement and reporting.
Siting matters. The International Energy Agency notes that data-centre demand is concentrated in a few regions and that grid constraints are a growing issue. Places with surplus renewable generation, which today sometimes has to be curtailed, are natural candidates. A public programme can also require energy reporting and efficiency from the start, in a way that is harder to impose on foreign providers.
Does open-source change the picture?
Yes. Open models let countries adapt rather than rebuild. Open-weight models from many countries are now widely used. The ATOM Project reported that Chinese open models overtook American ones in cumulative downloads on Hugging Face in mid-2025.
Fine-tuning an existing open model for a language, sector or legal system uses a small fraction of the energy of training a new one. Switzerland's Apertus shows another route: a fully open model, built in public, that others can reuse. Both reduce the need for every country to pay the full training cost.
Is there a middle path?
Many argue for sovereignty over the parts that matter, and interdependence for the rest. In practice:
- Control the data and the deployment, even if the base model comes from elsewhere.
- Prefer small, local and specialised models, which can run on national or on-premises infrastructure and keep sensitive data in place. See AI at the edge.
- Share compute and open models with allies rather than duplicating them.
- Build efficiency, measurement and clean power into public AI from the outset.
Some go further and ask whether distributed inference, spread across many smaller providers, could stop any single company or government controlling how AI is used. That is an open research question, with its own efficiency trade-offs.
What to do next
- If you shape public AI programmes, require energy measurement and reporting from the start: see green AI policy.
- Prefer adapting open models to training new ones, where the use case allows.
- Site compute where clean power is available: see hosting and data centres.
- Weigh edge and local deployment for sensitive data: see edge or cloud?.
- Read why green AI matters now for the energy context.
Frequently asked questions
What is sovereign AI?
AI capability that a country, or a group of countries, controls: its own compute, models, data or skills, so that it does not depend entirely on foreign companies or governments.
Is sovereign AI greener?
Not automatically. A national model trained from scratch duplicates energy that others have already spent. It can be greener if it is smaller, specialised, runs on clean power and replaces less efficient use, but that has to be designed in.
What is public AI?
The idea that some AI should be provided as public infrastructure, like public broadcasting or public transport, alongside private services, with public accountability and access.
Sources
- Sovereign AI Index, Center for a New American Security (2026)
- Sovereign AI Unit, UK Government, Department for Science, Innovation and Technology (2026)
- AI Opportunities Action Plan: one year on, UK Government (2026)
- EU launches InvestAI initiative to mobilise €200 billion of investment in artificial intelligence, European Commission (2025)
- Call for tenders: selection of AI gigafactory consortia, EuroHPC Joint Undertaking (2026)
- Apertus: a fully open, transparent, multilingual language model, EPFL, ETH Zurich and CSCS (2025)
- Public AI: making AI work for everyone, by everyone, Mozilla Foundation (2024)
- Sovereignty in the age of AI: strategic choices, structural dependencies, Tony Blair Institute for Global Change (2026)
- Is AI sovereignty possible? Balancing autonomy and interdependence, Brookings Institution (2026)
- UK's best AI engineers can apply now to build tech for public services in £1 million fellowship, UK Government (2025)
- The ATOM Project, ATOM Project (2026)
- Energy and AI, International Energy Agency (2025)
Related
How to write a green AI policy
A step-by-step guide to writing an organisation's green AI policy, with a template outline covering use, models, measurement, buying, claims and review.
Green AI hosting and data centres
How to choose where AI runs: grid carbon intensity, PUE and WUE, renewable energy claims, water use and siting, with questions to ask hosts.
Why green AI matters now
AI is driving rapid growth in data-centre electricity demand, straining grids and raising costs. The evidence on scale, energy, cost and policy in 2026.
AI at the edge: greener, or just less visible?
On-device and edge AI can save network and data-centre energy, but spreads computation across billions of devices nobody measures. What we know and don't.