# AI sovereignty: five conditions for the power to decide

Author: Fernando Nieto Lobato
Original publication: 2026-09-16
Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-155-que-separa-a-los-paises
English URL: https://elcontemplador.github.io/estrategia-english/essays/155/
Status: Published translation

This is a translation of the original Spanish essay published on 16 September 2026. Its claims, examples and forecasts retain that historical context.

English publication: 2026-09-29

**Sovereignty in artificial intelligence has become a widespread political ambition. Almost every country uses the term in its reports, plans and policy projections. Yet our own experience leaves us with the impression that, all too often, those using it are not fully aware of what it actually means. That gap prompted us to address a key question briefly in this article: what does a country really need to be sovereign in AI?** The question came back to me with [Mistral's announcement of a €3 billion funding round on 8 September 2026](https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/). Good news for Europe, but, seen in perspective—as this article also attempts to do—it reveals the enormous scale of what is still missing.

Sovereignty, in its fullest sense, requires the ability to develop, use and adapt capabilities without being at the mercy of decisions made elsewhere. No country manufactures everything it needs within its own borders; even the United States depends on chips made in Taiwan. The question is what a country controls, what it has secured and what alternatives it retains.

With that in mind, we prefer to bring the debate down to five concrete conditions that the vast majority of countries—arguably all but two—remain far from meeting in full.

## 1. Electricity

Elon Musk put it this way in Davos in January 2026: [“I think the limiting factor for AI deployment is fundamentally electricity.”](https://www.weforum.org/podcasts/meet-the-leader/episodes/conversation-with-elon-musk-davos-2026/) His diagnosis, repeated since at numerous events and in interviews, points to a basic dependency: buying chips is of little use if you cannot power and cool them.

China has an extraordinary advantage here. In 2025, it generated around [10,575 terawatt-hours of electricity](https://www.stats.gov.cn/english/PressRelease/202602/t20260228_1962661.html), compared with approximately [4,430 in the United States](https://www.eia.gov/outlooks/steo/report/elec_coal_renew.php) and [2,810 in the EU](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Energy_production_and_imports).

Price matters too. According to the International Energy Agency, in 2025, energy-intensive industries in the European Union paid [roughly twice the US benchmark, based on Texas, and around 50% more than their Chinese counterparts](https://www.iea.org/reports/electricity-2026/prices). These are industrial electricity prices, not rates specific to data centres.

Europe has a valuable foundation: renewables and nuclear power supplied [just over 70% of the EU's electricity generation in 2025](https://ec.europa.eu/eurostat/en/web/products-eurostat-news/w/ddn-20260629-2). Turning that into a continuous, competitively priced supply, available where it is needed, is a separate task. Electricity sovereignty comes down to a working connection, a price and a date. In this area, China's advantage is currently considerable.

## 2. Compute

Compute makes it possible to train and run models: accelerators, memory, networks and data centres working together. It is highly concentrated. [Epoch AI estimated](https://epoch.ai/latest/introducing-the-ai-chip-owners-explorer) that, at the end of 2025, five US groups—Amazon, Google, Meta, Microsoft and Oracle—owned more than 70% of the world's capacity in leading AI chips. Customers in mainland China accounted for just over 5%, excluding possible smuggling and capacity rented abroad.

Europe's dependence is also evident in the cloud. In the first half of 2025, [Amazon, Microsoft and Google accounted for 70% of cloud infrastructure revenue in Europe; European providers accounted for 15%](https://www.srgresearch.com/articles/european-cloud-providers-local-market-share-now-holds-steady-at-15). Hosting data centres is obviously better than nothing, but it does not guarantee control over the companies operating them, or alternatives that could replace them.

Europe does retain one uniquely valuable strategic position: the Dutch company [ASML is the only manufacturer of extreme ultraviolet lithography systems](https://www.sec.gov/Archives/edgar/data/937966/000162828026011377/asml-2025xannualxreportx.htm), essential for making advanced chips. It is a formidable asset, but on its own it does not provide a European cloud of comparable scale or give the continent the real capacity to compete with the United States and China.

## 3. Laws and oversight

Personally, although I consider legislation highly relevant, I see it as secondary when the material and human capabilities that should underpin it are missing. It can protect rights and establish responsibilities; on its own, it can hardly turn a country into a technological power with practical autonomy.

Oversight, we should remember, also requires capabilities of one's own. The AI Act [requires providers of general-purpose models posing systemic risk to conduct evaluations and adversarial testing](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-55): deliberate attempts to find failures and vulnerabilities. The [European AI Office can require access to models for evaluation and demand corrective measures](https://digital-strategy.ec.europa.eu/en/policies/enforcement-ai-act). Nor does the United States leave all scrutiny to companies: in May 2026, [CAISI, its centre for AI standards and innovation, announced agreements with Google DeepMind, Microsoft and xAI to conduct evaluations before deployment](https://content.govdelivery.com/accounts/USNIST/bulletins/415cadf).

This raises a decisive question: who carries out the tests, with what access, and what happens when a problem is found? A one-off evaluation does not guarantee lasting safety either. Staff must be capable of assessing manufacturers' claims, monitoring systems in operation and providing the grounds for demanding changes or limiting a use. That capacity for independent judgement is part of sovereignty.

Rules also gain force when they affect critical supplies. In January 2026, [Washington changed the conditions for authorising exports to China of accelerators such as the NVIDIA H200](https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china). Regulating a technology that others need confers a different kind of influence from regulating one you need to buy.

## 4. Models, people and data

Buying access to a model, running an open-weight model and developing your own provide different degrees of control. Having the weights—the learned parameters—allows you to operate the model and, depending on its licence, adapt it. It does not automatically provide the data and knowledge needed to reproduce its training.

Stanford's [AI Index 2026, with data updated in April](https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_1_research_development.pdf), attributes 59 notable models produced in 2025 to the United States, 35 to China and two to Europe. This is a selection of significant models, not a census or a quality score, but it is highly revealing of which territories alone possess real autonomy in this area. The [report also describes a substantial narrowing of the performance gap between the best US and Chinese models](https://hai.stanford.edu/ai-index/2026-ai-index-report). China is contesting that frontier despite restrictions on its access to chips. And, as we have learnt in recent weeks, [some of its new frontier models have already been trained directly on domestic Huawei chips](https://www.xataka.com/robotica-e-ia/startup-china-ia-que-lanzo-sorprendente-modelo-glm-5-2-acaba-alcanzar-otro-hito-crear-su-propio-centro-datos).

Mistral matters because it helps preserve Europe's development capability, although in recent months it appears to have shifted towards something closer to a European Palantir than a company that trains frontier models—an extremely difficult and costly undertaking, but also a very important one for autonomy. A single company, moreover, cannot replace an ecosystem of researchers, engineers and organisations able to use and improve what is produced. Training professionals who are then left without machines, projects or continuity is an incomplete policy.

## 5. Capital and speed

Mistral's funding round helps put the figures in perspective. Its €3 billion sits alongside [the $65 billion announced by Anthropic in May 2026](https://www.anthropic.com/news/series-h) and [the $122 billion in committed capital with which OpenAI closed its round in March](https://openai.com/index/accelerating-the-next-phase-ai/). Anthropic's figure includes $15 billion previously committed; OpenAI's commitments likewise do not amount to cash fully disbursed. The currencies differ, but the scale speaks for itself. What Europe celebrates as an exceptional transaction falls far short of the resources its leading competitors can mobilise.

There is also a revealing paradox: the very round that strengthens Mistral's position as Europe's leading champion also, to some extent, dilutes its European character. New investors include Samsung, NVIDIA, BlackRock and Andreessen Horowitz (a16z), alongside the Scaleup Europe Fund. In other words, even one of the most important assets in Europe's technological sovereignty needs funding and support from a capital and technology ecosystem that is deeply transatlantic and Asian.

The [OECD calculates that, in 2025, US companies attracted $194 billion in venture capital for AI, compared with $15.8 billion in the EU and $13.9 billion in China](https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/02/venture-capital-investments-in-artificial-intelligence-through-2025_3bcb227f/a13752f5-en.pdf). The United States received more than twelve times the EU's total. This indicator excludes important parts of AI financing, such as public investment and companies' internal expenditure—a caveat particularly worth remembering when interpreting the figures for China.

Money pays for hiring, buys compute time and funds attempts that fail. It then has to be turned into operational capability. [In this interview, which we previously recommended, Pep Martorell drew a revealing comparison](https://www.youtube.com/watch?v=vgC4W6H_INg&t=5997s): the time Elon Musk took to build Colossus against the time Europe was spending, and is still spending, deciding where to locate its gigafactories.

According to [NVIDIA, xAI brought the first Colossus online in 122 days during 2024, with 100,000 Hopper accelerators](https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus): roughly four months. It had an advantage: [it repurposed a former Electrolux factory in Memphis](https://content.govdelivery.com/accounts/TNMEMPHIS/bulletins/3d43c58).

The Commission [announced InvestAI in February 2025, with €20 billion earmarked for gigafactories](https://digital-strategy.ec.europa.eu/en/news/eu-launches-investai-initiative-mobilise-eu200-billion-investment-artificial-intelligence). The [call opened in July 2026; selection is expected in early 2027, with operations beginning within the following eighteen months](https://www.eurohpc-ju.europa.eu/eurohpc-joint-undertaking-launches-ai-gigafactories-call-2026-07-30_en). That is approximately two years between the announcement and the anticipated selection. This is the timetable for gigafactories, not for all of Europe's existing supercomputing capacity, but it reveals a profound disconnect from the reality of today's relentlessly accelerating environment.

Europe, at least—particularly when compared with individual countries that are not superpowers yet set themselves maximalist ambitions for AI autonomy—has industrial, scientific and energy foundations on which to build, alongside enormous gaps in platforms, financing and the production of notable models. For many countries, securing critical applications, retaining their own teams and keeping alternatives available will be more viable than attempting to reproduce the entire chain.

So the next time any government promises sovereign AI, it is worth asking for concrete answers: how much electricity has it secured, what machines can it use, who knows how to operate them and oversee the systems, how will the next stage be financed, and when will it be working? That is where we begin to discover how much real decision-making power lies behind the words.

[Fernando Nieto Lobato](https://www.fernandonieto.es/)

*Director of Digital Innovation at [Institución Educativa ALEPH](https://institucioneducativaaleph.com/) and editor of the estrategIA newsletter*

*Translation note: Musk's quotation is translated from the Spanish article. Its exact wording has not been independently checked against the original English recording.*

[![Infographic on AI sovereignty, its five operational conditions, the different positions of the United States, China and Europe, and a practical path for other countries. Complete English text follows.](https://elcontemplador.github.io/estrategia-english/assets/images/7d65d9ff-eeff-4128-ba0d-224d043bef49_967x1536.png)](https://substackcdn.com/image/fetch/$s_!qFm6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d65d9ff-eeff-4128-ba0d-224d043bef49_967x1536.png)

## English text alternative for the infographic

**AI sovereignty: what separates rhetoric from capability.**

No country manufactures everything. What matters is what it controls, what it has secured and what alternatives it retains.

The opening illustration is labelled ‘Decisions that build the future’; ‘Infrastructure, talent, rules, capacity’; and ‘More capability, more autonomy, more opportunities’.

### Five conditions that must work together

1. **Usable energy.** Chips need electricity and cooling: a continuous supply, a competitive price and a connection where they are needed.
2. **Hosting is not control.** Having data centres does not guarantee control over their operators or alternatives that could replace them.
3. **Regulation requires the ability to evaluate.** Rules need teams with access and independent judgement to assess manufacturers’ claims, monitor systems and provide grounds for corrections. The clipboard is labelled ‘Evaluation’.
4. **Models and talent of one’s own.** Buying access, operating open-weight models and developing one’s own offer different degrees of control. Talent needs resources and continuity. The books are labelled ‘Research’, ‘Talent’ and ‘Innovation’.
5. **Investing and putting capabilities into operation.** Capital must be turned into people, compute and operational infrastructure. Timing also determines real capability.

### An unequal map of power

| Region | Position described in the graphic |
| --- | --- |
| United States | Concentrates compute and capital, and maintains substantial model-development capability. |
| China | Has a major electricity advantage and contests the frontier of AI models despite restrictions on its access to chips. |
| Europe | Has ASML and scientific and energy foundations, but remains far behind in platforms, financing and models. |

### A viable path for many countries

Secure critical applications, retain teams of their own and keep alternatives available, rather than attempting to reproduce the entire chain.

*estrategIA. This is a translation of the original infographic, retaining the article’s September 2026 context.*