Beijing (北京) published a 2026–2028 AI-token-economy plan covering inference-chip development, model–chip optimization, and evaluation standards for token-production infrastructure.

Beijing’s municipal economic and information technology bureau and development and reform commission issued the Beijing Action Plan for Accelerating Token-Economy Development, 2026–2028 (北京市加快词元经济发展的行动方案（2026—2028年）). The document covers both token-production infrastructure and downstream AI-agent applications.

For token factories, the plan calls for graded evaluation standards covering the number of supported models, token throughput, first-token latency, cache-hit rate, and power usage effectiveness (PUE). It also calls for better provision of land, energy-consumption allocations, and communications networks, with support for token factories demonstrating industry leadership; no numerical qualification thresholds are supplied.

Technical priorities include foundation-model iteration, model–chip adaptation and optimization, inference-specific chips, model lightweighting, and edge deployment. The plan also calls for developing and accessing AI-computing resources to improve token-production capability, but the supplied excerpts contain no committed accelerator purchases, deployment volumes, or funding totals.

On the application side, it supports general-purpose and specialized agents for individuals and institutions, including information, business, and daily-life services, plus manufacturing, research, biomedicine, and education. Its software priorities include safe and controllable technology stacks, memory management, permission controls, efficient inter-component communication, and the development of agent operating systems and new software-tool businesses.