Guiyang already stores the country's data and powers it with some of China's cheapest, greenest computing power (算力). What it has struggled to do is sell the result. A city can host 26 data centers and still watch value leak away as raw capacity, not finished intelligence.
So the provincial capital of Guizhou has picked an unusual unit to build its next economy around: the token (词元) — the smallest slice a large model (大模型) reads and writes, and, the local government now argues, a thing you can meter, price and trade like electricity.
Key takeaways
- Guiyang and the Gui'an New Area published the China Data Valley · Token City three-year action plan (2026-2028) on 26 August 2026.
- The plan sets concrete 2026 targets: 10 high-quality datasets and more than 30 typical AI application scenarios, scaling to 30 datasets and 70+ scenarios by 2028.
- It pushes compute centers to shift from renting raw computing power (算力) to delivering callable "token production" services — "token factories."
- A dedicated token zone (词元专区) will open on the Guiyang Big Data Exchange, and authorities will issue "Guiyang-Gui'an token vouchers" (词元券) to cut enterprise costs.
- New intelligent compute centers must keep average PUE at or below 1.2 and green-power share above 80%.
- The framing treats tokens (词元) as a kind of "hard currency" of the intelligent economy, exportable along the Belt and Road via the Big Data Expo.
What a 'token city' actually means
In model terms, a token (词元) is the chunk a large model (大模型) breaks text, images and audio into before processing them. The more you generate, the more tokens you burn, and AI services already bill by them. Guiyang's bet is that if tokens are measurable, they can also be priced and exchanged — turning the output of a model into a commodity with a spot market.
The plan, issued by the General Office of the Guiyang Municipal Government together with the Gui'an New Area Office (document number 筑府办函〔2026〕36号), leans on the city's existing edge: it is the core of a national big-data pilot zone and already reports computing scale above 150 EFLOPS with intelligent-compute share over 98%.
Guiyang is not starting from zero. Over the past decade it attracted data centers — the local government counts 26 large-and-above facilities and claims the country's strongest domestic intelligent-compute capacity. The gap it now wants to close is value capture: hosting someone else's servers earns far less than selling the intelligence those servers produce. Pricing that intelligence by the token is the city's attempt to move up the stack without owning a frontier model, and to keep more of the economic return local.
The targets, year by year
The document lays out a three-step climb. In 2026 it aims to build 10 high-quality datasets and 30-plus AI scenarios, seed 5 OPC (One Person Company, 一人公司) communities and attract 10 token-related firms. By 2027 those cumulative figures rise to 20 datasets, 50-plus scenarios, 10 communities and 15 firms. By 2028 the goals are 30 datasets, 70-plus scenarios, 15 communities and 20 firms.
Those are planning targets, not appropriations — the document sets direction and assigns lead bureaus (the city big-data bureau and the Gui'an science-and-technology bureau) rather than earmarking a specific budget.
From raw compute to token factories
The most concrete shift is operational. The plan tells compute-center operators to move beyond "bare computing power (算力)" and toward standardized, directly callable token-production services — clusters of "token factories" (词元工厂). New centers face a green bar: average PUE no higher than 1.2 and green-electricity share above 80%, reflecting both cost and the region's hydro-rich profile.
On the supply side, it batches public-data authorization in fields like housing, healthcare, culture and tourism, and builds trusted data spaces so that compliant datasets can feed model pre-training, fine-tuning and agent (智能体) development.
The token-voucher twist
To pull demand, the plan promises "Guiyang-Gui'an token vouchers" (词元券), issued under a dedicated policy to lower what enterprises pay to call tokens. It also opens a token zone (词元专区) on the Guiyang Big Data Exchange where model-as-a-service (模型即服务, MaaS) and agent-as-a-service (智能体即服务, AaaS) products can be listed, and explores token-trading as a new transaction mode. That mirrors a broader Chinese pattern of using vouchers to seed early demand — here applied to model output rather than compute hours.
Honest limitations
This piece relies on the municipal government's published plan and provincial government summaries; I have not seen separate budget or implementation documents, and the numbered targets are the government's own and unverified by independent audit. The token (词元) "hard currency" framing is the plan's rhetoric, not an established market outcome. Whether a token exchange can gain real liquidity, and how token pricing interacts with data-property and cross-border rules, remains open. I have not assessed competitive responses from other AI clusters.
What readers can do now
- If you run an AI team, watch the Guiyang Big Data Exchange token zone (词元专区) for MaaS/AaaS listings that may price inference more transparently than opaque API bundles.
- If you are a startup, track the "token vouchers" (词元券) scheme as a potential subsidy on model calls, and the OPC (一人公司) community incentives for relocation.
- If you cover regional policy, compare Guiyang's token-metering approach with other cities' computing-power voucher models to see which demand-side lever proves stickier.