In early September 2026, Xataka highlighted three events that together paint a picture not just of technological progress, but of deep structural shifts in the AI industry. South Korea is accelerating the creation of sovereign infrastructure with free citizen access to national models, OpenAI releases GPT-6 Astra with record results on the Korean CSAT exam and claims about AGI, and Nvidia acquires Hugging Face for 12,9 billion dollars. These steps are not random: they reflect competition between national strategies and control over the open ecosystem.
Korea's 'Independent AI Foundation Model' program and the 'AI for All' project involve investments of over a trillion dollars in data centers and the selection of domestic models — A.X K2 from SK Telecom (688 billion parameters) and K-EXAONE from LG. The government requires that at least 50% of the service operate on Korean models, and citizens receive unlimited free access. This is not just import substitution: experiments show that such models handle the Korean language and local tasks better than GPT-5.6, while remaining open for fine-tuning. However, the question of scaling remains open — whether 8,4 GW of capacity by 2029 will be enough to compete with American clusters.
GPT-6 Astra from OpenAI demonstrates a different vector. The model was the first to score a perfect 450 on the full Korean CSAT 2026, using only 357 thousand tokens versus 429 thousand for its predecessor. On FrontierMath Tier 4 — 98%, ARC-AGI-3 — 99,9%, ExploitBench — 100%. Astra handles long computer tasks faster than a human: finding a job in 2 minutes 51 second instead of five hours. OpenAI and Jensen Huang directly speak of an 'era of AGI,' citing training on over 100 thousand Grace Blackwell systems. However, the evaluation methodology remains partially closed: it is unclear how well the results generalize beyond benchmarks and how robust the model is to 'going off the rails' in real-world scenarios.
Nvidia's purchase of Hugging Face for 12,9 billion dollars closes the triangle. The platform, which hosts millions of models, including Korean A.X K2 and K-EXAONE, is now under the control of the chip maker. Nvidia promises to maintain openness and not require its own accelerators, but it is clear that control over the repository strengthens the company's position in the open-weight model ecosystem. This is a direct response to sovereign initiatives: if countries like Korea want independence, Nvidia through Hugging Face gains leverage over the distribution and fine-tuning of models.
Comparing approaches reveals key divergences. The Korean strategy bets on localization of data, tokenizer, and infrastructure, reducing dependence on OpenAI and Google. Astra, meanwhile, shows that frontier models still lead in complex agentic tasks and mathematics. The Nvidia-Hugging Face deal could accelerate the spread of open models, but simultaneously concentrates influence in one player. For researchers, this means new opportunities to experiment with Korean weights, but also the risk that key tools fall under the influence of a single company.
What remains unresolved: how sustainable Korean models are when scaled to Astra's level, and whether Nvidia will truly maintain platform neutrality after the deal closes in 2027. Independent tests on real Korean tasks and checks of Astra's behavior in long offline sessions will be decisive. Future work will likely compare the efficiency of national clusters with American ones under conditions of limited data access.
These events show that the AI field is dividing into national infrastructure projects and control over open repositories — and it is here that it will be decided who sets the rules of the game in the coming years.
