AI models by country: who controls the frontier and who is fighting for sovereignty

You open a table of AI models by country and it looks like a sticker album: GPT here, Qwen there, Mistral in France, ALIA in Spain, Falcon in the UAE, Sarvam in India. If you read it like that, country against country, you miss the important part.

The question is not “which model does each country have”. The question is who controls the full chain.

A country can have a model and still depend on foreign chips, foreign cloud, foreign data, closed licenses, and APIs that can be switched off from another jurisdiction. Another country may not win any global benchmark and still be more sovereign for its language, public administration, or local industry.

This snapshot is closed on April 25, 2026. It does not try to list every existing model. It tries to organize the map: frontier models, open weights, languages, public models, media models, and infrastructure.

The thesis: having a model is not enough

AI sovereignty has several layers:

Model: training or adapting a useful family.
Data: local corpora, public data, sector data, and clear licenses.
Compute: GPUs, TPUs, NPUs, supercomputers, or capacity agreements.
Cloud: where the model is served and under which jurisdiction.
Deployment: API, product, enterprise integration, public services.
License: open weights, closed API, commercial use, auditability, and right to modify.

If one layer is missing, there is dependency. If three are missing, what you have is not sovereignty. It is access.

The quick map

Block	Relevant actors	Reading
Closed frontier	OpenAI, Anthropic, Google DeepMind, xAI, Meta	The U.S. still dominates the highest-capability product layer. DeepMind adds British scientific capacity, but inside Alphabet.
Chinese frontier	Qwen, DeepSeek, Kimi, ERNIE, Seed, GLM, Hunyuan/Hy3, MiniMax, Pangu	China is no longer “DeepSeek and Qwen”. It is a whole ecosystem of labs, apps, cloud, and partial hardware substitution.
Strong Europe	Mistral, FLUX, Stability AI, Pharia, ALIA, GPT-NL, Nordic and Polish projects	There is talent and there are good models. The bottleneck is compute, cloud, and industrial scale.
Language sovereignty	ALIA, Sarvam, tsuzumi, PLaMo, HyperCLOVA, SEA-LION, Falcon, Jais, LatamGPT	Many countries are not trying to win the global benchmark. They want their language and public sector not to depend on foreign APIs.
Media models	FLUX, Stable Diffusion, Veo, Imagen, Lyria, Seedance, Hailuo	Do not mix them with chatbots. They can be frontier in image, video, or audio without being general-purpose LLMs.

The U.S. and China are playing at another scale

The general frontier is still concentrated in two poles.

The U.S. has the full closed-product package: OpenAI with GPT-5.5, Anthropic with Claude Opus 4.7, Google with Gemini 3.1 Pro, xAI with Grok, and Meta combining closed product models with Llama as the open layer. The advantage is not only the model. It is distribution, clouds, capital, tools, integrations, and a massive user base.

China is the only alternative ecosystem with comparable width. Qwen, DeepSeek, Kimi, ERNIE, Seed, GLM, Hunyuan/Hy3, MiniMax, and Pangu cover reasoning, code, agents, vision, video, voice, consumer apps, national clouds, and enterprise deployment. Many of these models also publish weights or technical cards with enough detail for the rest of the world to test, tune, and deploy them.

The strategic difference is simple:

The U.S. maximizes closed product, global distribution, and monetization.
China maximizes ecosystem width, fast releases, open weights, and partial infrastructure substitution.

Open weights are now geopolitical

For years, open weights looked like a community issue. In 2026 they are also a policy of influence.

When DeepSeek, Qwen, Kimi, Mistral, Llama, Hy3, MiniMax, Sarvam, ALIA, or Falcon publish weights, they are not just giving away a model. They lower the entry barrier for universities, small companies, regional governments, and teams that cannot pay for or do not want to depend on a closed API.

That does not mean every open-weight model creates sovereignty. You still need to know whether the license allows commercial use, whether the model can run on your own infrastructure, whether there is enough data to adapt it, and whether inference cost is viable. But open weights change the power relation: they allow auditing, adaptation, and deployment.

That is why China is using openness so aggressively. It lets Chinese labs contest the developer infrastructure layer, the same way Android contested the mobile layer.

Europe has models, but not enough industrial muscle

Europe is not empty. France has Mistral, probably the European actor closest to a general frontier family. Germany is stronger in image and enterprise sovereignty: FLUX from Black Forest Labs, Pharia from Aleph Alpha, OpenGPT-X as a European language project. The United Kingdom contributes DeepMind as scientific power and Stability AI in creative models, although DeepMind is not full British corporate sovereignty.

Spain is a different case: ALIA/Salamandra is not trying to be “the European GPT”, but it matters for Spanish, Catalan/Valencian, Galician, and Basque. It is public infrastructure, not just product.

Europe’s problem is not lack of talent. It is industrial coordination.

There are researchers, universities, startups, regulation, and public programs. What is missing compared with the U.S. and China is an integrated block of compute, cloud, product, and deployment speed. EuroHPC and AI Factories help, but they do not instantly replace a hyperscaler.

Language sovereignty: the other frontier

Not every country needs to win Humanity’s Last Exam. Many need citizens to speak to public services in their own language, hospitals to process local documents, companies not to send sensitive data to foreign APIs, or teachers to use tools in languages other than English.

That is where models that may not win global headlines still reduce real dependency:

Language area	Models or initiatives	Why it matters
Spanish and co-official languages	ALIA/Salamandra	Spain covers Spanish, Catalan/Valencian, Galician, and Basque through public infrastructure.
Indian languages	Sarvam, Indus, BharatGen, Krutrim	India has 22 official languages. Voice, low cost, and public services matter more than the global benchmark.
Japanese	tsuzumi, PLaMo, Fugaku-LLM	Japan prioritizes enterprise use, efficiency, local deployment, and national supercomputing.
Korean	EXAONE, HyperCLOVA X, Solar	Korea combines industrial groups, national models, and local cloud.
Arabic	Falcon-H1 Arabic, Jais, ALLaM, Fanar	The Arab world is investing heavily in local models, state capital, and government deployment.
Southeast Asia	SEA-LION, Typhoon, Sahabat-AI, VinAI	The region is working more on language adaptation than on pure frontier competition.
Latin America	Sabiá, LatamGPT	Culturally important, but the compute and foundational-weight gap remains large.

Language sovereignty is less glamorous than a huge model, but probably more important for public impact.

Do not mix LLMs with media models

A common mistake is measuring everything as if it were a chatbot. FLUX, Stable Diffusion, Stable Audio, Stable Virtual Camera, Veo, Imagen, Lyria, Seedance, and Hailuo should not live in the same mental table as GPT, Claude, or Qwen.

They are another frontier.

Germany looks weaker if you only look at general LLMs. With FLUX, it becomes a global image pole. The U.K. does not have a sovereign chatbot champion at Mistral’s level, but Stability AI still matters in image and audio. China is not only competing in reasoning: ByteDance and MiniMax are pushing video and voice very fast.

The right question is not “which country has the best model”. It is “in which modality does each country control real capability”.

Model origin: training is not the same as adapting

To read the map correctly, separate four cases.

Type	Examples	Reading
Own foundation family	GPT, Claude, Gemini, Qwen, DeepSeek, Kimi, ERNIE, Seed, GLM, Hunyuan, MiniMax, Pangu, Mistral, Cohere, Jamba	This is strong industrial capacity: data, compute, training, evaluation, and product.
Own sovereign model, not necessarily frontier	Sarvam, ALIA, Pharia, EXAONE, Solar, tsuzumi, PLaMo, Fugaku-LLM, Falcon-H1, Jais	They do not always compete with GPT or Gemini, but they reduce language and deployment dependency.
Specialized media model	FLUX, Stable Diffusion, Stable Audio, Veo, Imagen, Seedance, Hailuo	Creative frontier. Measure it by image, video, or audio, not by chat.
Adaptation over open bases	Typhoon, OpenThaiGPT, Sahabat-AI, GEITje, part of TAIDE, several Nordic models	Useful for language and local data, but not the same as training a foundation family from scratch.

This distinction prevents inflated headlines. A local fine-tune can be very useful, but it does not mean the country controls the frontier.

Quick regional cards

Region	Strengths	Gaps
North America	General frontier, agents, code, cloud, product, monetization	Private concentration and NVIDIA/TSMC supply-chain exposure.
China	Full ecosystem: models, apps, cloud, open weights, video, voice, Ascend infrastructure	Advanced-chip restrictions and uneven transparency.
Europe	Science, open models, language sovereignty, regulation, EuroHPC	Fewer native hyperscalers and less industrial speed.
Japan/Korea/Taiwan	National languages, electronics, memory, industry, supercomputing	Smaller global open ecosystem than China or the U.S.
India/Southeast Asia	Language diversity, voice, public programs, fast adoption	Compute gap and heavy dependency on foreign GPU/cloud.
Middle East	Capital, energy, Arabic, state deployment	Imported accelerators and foreign alliances.
Latin America/Africa	Clear public need, languages, cultural data	Lack of compute, foundation weights, and sustained funding.

The matrix that actually matters

If I had to evaluate a country’s AI sovereignty, I would not start with the benchmark. I would start with this matrix:

Layer	Uncomfortable question
Compute	Can it train or serve relevant models without asking permission abroad?
Data	Does it have local, clean, legal, useful corpora?
Talent	Does it train from scratch or only fine-tune other people’s models?
Cloud	Where do the data live and under which jurisdiction?
Product	Does it reach companies, citizens, and public administration?
License	Can it audit, modify, and deploy without closed contractual dependency?

With that matrix, China looks stronger than many countries with flashy models. Spain looks stronger than benchmarks alone would suggest, because ALIA attacks language and public infrastructure. Latin America looks weaker than it should given its size and cultural importance, because it still lacks compute and regional continuity.

Conclusion

In April 2026, the geography of AI is a pyramid.

At the top are the U.S. and China. The U.S. keeps the strongest closed frontier, the best global distribution, and the clouds that package AI as product. China has built the most complete alternative ecosystem: many labs, aggressive open weights, massive apps, national cloud, and partial hardware substitution through Ascend.

Europe has science and models, but is still fighting for scale. Mistral, FLUX, Stability AI, Pharia, ALIA, GPT-NL, and Nordic and Polish projects show real technical capacity. The European problem is turning that capacity into sustained infrastructure.

India, Japan, South Korea, and the Middle East show another reading: sovereignty does not always mean beating GPT or Gemini. Sometimes it means controlling language, voice, public services, sensitive data, local industry, and deployment cost.

Latin America and Africa remain the most exposed to external dependency. They have population, languages, public needs, and cultural data, but not yet comparable foundation-model infrastructure. The challenge there is not only technical. It is funding, energy, public procurement, talent, and regional cooperation.

The thesis is simple: having a model is only the first layer. Sovereignty appears when you can sustain it, audit it, adapt it, deploy it, and update it with controlled infrastructure and a license compatible with your interests.