Stanford HAI's 2026 AI Index Report reveals a severely underestimated trend: the competitive landscape of global top AI models is undergoing structural transformation. The report data shows the performance gap between China and the US on top-tier AI models has narrowed from 31.6% to just 2.7% — meaning that as of March 2026, China's top models have nearly caught up with the US's most advanced level.
This number overturns the outside perception of US absolute AI leadership. But the structural factors behind it deserve more attention: US private AI investment in 2025 reached $285.9 billion, while China's was only $12.4 billion — the former is 23× the latter. This magnitude of capital advantage did not translate into a corresponding technological monopoly, indicating that Chinese labs are achieving comparable results with fewer resources.
Behind the efficiency revolution is the maturity of the open-source ecosystem. In April 2026, four Chinese labs — DeepSeek V4, Kimi K2.6, GLM-5.1, and MiniMax M2.7 — released open-source coding models in quick succession within 12 days, each reaching Western-frontier-comparable levels on agent engineering tasks, while their inference cost was less than one-third of the latter. This performance/cost-ratio competition logic is changing the rules of the game — the bar for small teams to use cutting-edge AI is rapidly dropping, accelerating the penetration speed of AI applications.
Notably, in the same period, UC Berkeley RDI released research exposing systematic flaws in existing benchmark systems — 45 methods can score full marks on 13 mainstream leaderboards without solving any problem. The 2.7% performance gap thus needs a question mark: when model-capability differences enter the 2% range, benchmark error may already exceed the real gap, and the industry urgently needs a contamination-resistant real-task evaluation system to recalibrate perception.
Behind this David-vs-Goliath catch-up, the old logic of US-China AI competition is being rewritten. The model of capital and compute piling up is being diluted by efficiency and ecosystem advantages. The role of open-source models is no longer just catch-up tools, but a structural force reshaping the competitive landscape.