Leaders in Artificial Intelligence (AI), such as OpenAI and DeepMind, view themselves as being in a race to build artificial general intelligence (AGI), a model capable of performing any intellectual task that a human can. At the same time, the US and Chinese governments view the AI race as a national-security priority that demands substantial investments, reminiscent of the Manhattan Project.

In both cases, AI is seen as a new form of "hard power," accessible only to superpowers with vast computational resources and the means to leverage them for economic and military dominance.

But this view is incomplete, and increasingly outdated. Since the Chinese developer DeepSeek launched its lower-cost, competitively performing model earlier this year, we have been in a new era. No longer is the ability to build cutting-edge AI tools confined to a few tech giants. Multiple high-performing models have emerged around the world, showing that AI's true potential lies in its ability to extend soft power.

The era of "bigger-is-better" models came to an end in 2024. Since then, model superiority has not been determined solely by scale (based on ever more data and computing power). DeepSeek proved not only that top-tier models can be built without enormous capital, but also that introducing advanced development techniques can radically accelerate AI progress globally. Dubbed the "Robin Hood of AI democratisation," its decision to go open-source sparked a wave of innovation.

The OpenAI monopoly (or oligopoly of a few companies) of only a few months ago has given way to a multipolar and highly competitive landscape. Alibaba (Qwen) and Moonshot AI (Kimi) in China have also since released powerful open-source models, Sakana AI (my own company) in Japan has open-sourced AI innovations, and the US giant Meta is investing heavily in its open-source Llama program, aggressively recruiting AI talent from other industry leaders.

Boasting state-of-the-art model performance is no longer sufficient to meet the needs of industrial applications. Consider AI chatbots. They can give "70-point" answers to general questions, but they cannot achieve the "99-point" precision or reliability needed for most real-world tasks, from loan evaluations to production scheduling that heavily rely on the collective know-how shared among the experts. The old framework in which foundation models were considered in isolation from specific applications has reached its limits.

Real-world AI is required to handle interdependent tasks, ambiguous procedures, conditional logic, and exception cases, all of which involve messy variables that demand tightly integrated systems. Accordingly, model developers should take more responsibility for the design of specific applications, and app developers should engage more deeply with the foundational technology.

Such integration matters for the future of geopolitics no less than it does for business. This is reflected in the concept of "sovereign AI," which calls for reducing one's dependence on foreign technology suppliers in the name of national AI autonomy.