The best AI models created outside the US and China are released approximately one year later than US models of the same level, according to a Village.dev study built on the LLM Atlas capability scale for 857 models. The gap has neither grown nor shrunk since 2024: "sovereign AI" is not catching up to the frontier but is settling at the level of the previous model generation. For anyone choosing models outside the two main markets, this means an adjustment of approximately one year for freshness and quality.
What happened
Village.dev published a study based on its proprietary LLM Atlas capability scale, which evaluated 857 models. The methodology measures lag using the time-to-parity principle: it records how much later a model is released that has reached the level of an already released model, so the lag is calculated in months rather than in disputed absolute scores. The leader of the scale remains GPT-6 Astra Pro with a score of 170.2; the best French model, Mistral Large 4, scored 159.7, South Korea's Motif-3-Beta — 153.9, Canada's Cohere North Mini Code — 139.7, the UAE's K2 Think V2 — 135.3, and India's Sarvam 30B — 124.1. The best models from labs outside the US and China appear on average 9.9 months later than US models that have reached the same level; the minimum gap since 2024 has not fallen below 7.3 months, while China lags by an average of 5.2 months. In terms of release volume in 2026, Atlas evaluated 113 models from the US, 79 from China, and about 20 models from all other countries combined. The study was published in the Village.dev blog and is being discussed in the Hacker News community.
Context
"Sovereign AI" here refers to models from labs and companies in individual countries outside the US and China, and their stakes differ significantly. Mistral focused on the open weights of Mistral Large 4, thereby reducing the gap to approximately 8 months; South Korea compressed the lag from 14.9 to 7.3 months; Cohere shifted to the enterprise segment with the North platform and the acquisition of Aleph Alpha in April 2026; the UAE is investing in computing infrastructure in the Stargate UAE project with OpenAI, while India has moved to training models from scratch — Sarvam 30B and Sarvam 105B — on state GPUs. One average lag metric masks these differences: the general annual corridor is combined with very different ways of closing it. The scale of the base outside the two leaders is also important: the narrowness of this set turns the conversation about equal competition with the frontier into a conversation about a second tier that consistently supplies models of one-year-old level.
Why this matters for the industry
For the industry, the study provides a quantitative picture of AI geopolitics and confirms that the second tier is not a temporary phenomenon but a stable layer of the market. When choosing a vendor and making purchases outside the US, it is reasonable to account for a lag of about 9–12 months in freshness and quality, or to accept dependence on the frontier led by GPT-6 Astra Pro. While the gap is stable, a two-tier market is strengthening: frontier APIs for complex tasks, and open and sovereign models for workloads with data and cost constraints. In such a configuration, multimodal routing and fallback schemes become a standard product pattern, while open weights reduce switching costs for self-hosting. If the lag persists over a couple of years, competition within the second tier will finally shift from capabilities to availability, licenses, multilingualism, and inference infrastructure.
Why this matters for users
For readers, this is a guide to understanding which models are actually available outside the US and China and how far behind they are: when choosing a local or "sovereign" model, one should account for an approximately one-year gap in freshness and quality. For typical tasks — extraction, classification, RAG, and medium-complexity agentic pipelines — the difference from the leader on the Atlas scale is comparable to the cost of one model release, so a working product can be built on second-tier models. This choice provides a significant advantage in data control and predictable cost, with the trade-off being a lag in frontier capabilities.
What is still unknown / limitations
Conclusions should be read with caveats. The number of models evaluated by Atlas is a product of labs' release policies and marketing: the statistics mix variants, fine-tunes, and models trained from scratch, so release volume cannot be interpreted as a measure of capabilities, and the thesis of a "near-duopoly" is a judgment about distribution, not science. Benchmarks are mostly in English, which may underestimate models optimized for Korean, Arabic, or Indian languages. Statistics for individual countries are fragile: conclusions about India are based on only two models. The study measures benchmark scores, not production readiness: it contains no data on latency, inference cost, API availability, or reliability. Future developments depend on whether external teams will verify the Atlas methodology on public benchmarks; if language bias is confirmed, the annual corridor for Korea, the UAE, and India may shrink, and with continued open releases, it may fall below the historical minimum.
Sources
- Outside the US and China, the world is building last year's AI — Village.dev
- AI models built outside the US and China are nearly a year behind — Hacker News discussion
Author
Look at AI, editorial team