Beijing Sets AI Inference Targets in New Token Economy Plan
The city is treating inference as industrial output that can be measured, routed and settled. That builds on an E-Town subsidy base already paying for compute, data and model consumption.
Beijing's new three-year action plan for the "token economy" grades AI inference infrastructure using operating measures that look more like factory KPIs than model-leaderboard scores. The September 18 plan includes token throughput, time to first token, cache-hit rates and power usage effectiveness, alongside token-quality evaluation, pricing, routing, metering and settlement.
The shift is easier to understand against Beijing E-Town's earlier build-out. In August, the district set a 2030 goal of more than 100,000P of compute, four 10,000-card token factories and a sixfold increase in token output per unit of electricity. E-Town also committed RMB100 million a year of compute vouchers and another RMB100 million of data vouchers, while offering qualifying companies subsidies of 30% of compute-rental costs up to RMB20 million.
Beijing's first token factory, launched in E-Town in June, has phase-one capacity of 1.4 trillion tokens a day and a stated long-term target of 10 trillion.
Beijing is standardizing output, not just hardware
A cluster's GPU count measures installed capacity. Token throughput measures what the facility produces. First-token latency, caching and PUE begin to describe the cost and service quality of turning accelerators and electricity into inference a customer can consume. Two facilities with similar hardware can therefore have very different economics if utilization, routing or cache reuse diverge.
An illustrative gross-revenue sensitivity shows the scale hidden inside the 1.4-trillion-token daily figure. DeepSeek currently charges $0.60 per million output tokens for V4.1 Flash off-peak and $1.20 at peak. If every token in Beijing's stated phase-one capacity were billable output at those rates and the factory ran at full utilization, the arithmetic would be about $840,000 to $1.68 million a day, or roughly $307 million to $613 million a year. At 50% utilization under the same all-output assumption, the annual range halves to roughly $153 million to $307 million.
The full-utilization range is best read as an illustrative gross-revenue ceiling at those reference prices, not a forecast. Beijing's factory can route different models; input and output tokens have different prices; discounts, caching, subsidies and actual utilization can change realized revenue sharply. The exercise turns a throughput target into a commercial scale that future utilization and realized-pricing disclosures can confirm or reject.
Subsidies still distort the clearing price
E-Town is subsidizing both compute supply and model consumption. Those incentives can raise utilization while lowering the price customers actually bear, so high token output by itself cannot establish an unsubsidized return on data-center capital.
Hardware suppliers have a visible demand tailwind because Beijing is funding inference infrastructure and usage. Operator economics are less settled. Realized token revenue after discounts and subsidies has to cover power, depreciation and the capital tied up in the factory.
Sources
- Beijing municipal September 18 token-economy action plan: https://jxj.beijing.gov.cn/zwgk/2024zcwj/202609/t20260918_4870474.html
- Beijing E-Town August token-economy measures: https://www.beijing.gov.cn/zhengce/zhengcefagui/202608/t20260806_4811112.html
- Beijing government, June report on the first token factory: https://english.beijing.gov.cn/beijinginfo/sci/latesttrends/202606/t20260612_4697373.html
- DeepSeek API, current V4.1 Flash pricing: https://api-docs.deepseek.com/quick_start/pricing/