Stanford University's Human-Centered AI Institute (HAI) today put out the official Chinese version of its 2026 AI Index Report. The report says artificial intelligence is no longer just a race over raw model capability. It has pushed into broader, system-level shifts in infrastructure, organizational processes, and institutional frameworks.
Right now, the world’s top models are getting close to human-level performance on the SWE-bench software engineering benchmark, hitting nearly 100% accuracy. And agents are improving fast in real computer environments: task success rates climbed from 12% to 66%. That points to a shift from “generating information” to “executing tasks.”
The performance gap between the best US and Chinese models has shrunk to 2.7%, and the leading models on both sides have repeatedly traded the top spot. The report also points to physical infrastructure as a main constraint on AI development. Global AI data center electricity capacity has reached 29.6 gigawatts. Chips, energy, and water resources are turning into hard strategic bottlenecks for AI progress.