China narrows AI gap with open models, but US maintains compute advantage

Chinese AI models are closing the performance gap with American frontier systems despite facing significant restrictions on access to advanced chips. Moonshot's Kimi K3 model, released in July, has substantially narrowed the gap with top models from Anthropic and OpenAI in benchmarks, even surpassing them in some areas. Chinese open-source models — which can be downloaded, modified and self-hosted — are currently the world's most capable, and many experts consider China ahead in robotics. While Chinese models lag behind the frontier, they have emerged as cheaper and highly capable alternatives to American options, with growing adoption on American soil and in developing economies in Africa. Analysts warn that if Chinese AI becomes the default for developing countries, those countries may align themselves politically with Beijing and give Chinese AI companies significant market opportunities. However, computing remains a critical challenge for China — US export controls have severely limited Chinese AI companies' access to the most advanced chips, affecting their ability to train larger and more capable models. Nevertheless, China is working to overcome computing constraints, with companies accused of accessing advanced computing abroad and its domestic chip industry gaining ground, though still at a significant distance from America.