NextFin News — Inside a Huangpu data center, GPU capacity is no longer sold mainly as whole servers. It is portioned into tokens—units of language-model output—and delivered through APIs. The operator is a district-owned digital subsidiary of Guangzhou’s Science City group. Its platform, marketed as a “model supermarket,” aggregates several large models on compute the group controls and advertises lower token prices together with standardized billing and local data handling.
Early customers include state firms, software developers and short-drama studios. Usage from the drama segment has been described as rising steadily day by day. The commercial proposition is straightforward: buy inference the way a factory buys electricity—on demand, metered, and without assembling a private cluster. Free or discounted allowances help new users start; volume pricing takes over once consumption grows.
A parallel experiment is under way across the city in Haizhu. In mid-August three banks—Bank of China, CITIC Bank and Guangzhou Bank—introduced lending products that treat token consumption as a material input to credit decisions. Facilities can be sized partly against contracted or observed token use, platform records and related receivables rather than solely against factories or equipment. One branch has reported pilot approvals of roughly 28 million yuan, with a portion already drawn. The products remain small, yet they mark an attempt to underwrite companies whose primary operating statistic is how many tokens they burn.
These two moves belong to a wider pattern. In recent months local investment vehicles and data platforms in several other Chinese cities have announced token trading centers or model aggregation services. After years of relying on land development and heavy infrastructure, many municipal groups are looking for lighter, more recurring lines of business. Packaging compute into standardized tokens is one of the openings they see.
The economic logic is simple on paper. Heterogeneous servers and multiple models are hard for a small firm to navigate. A single gateway that offers unified APIs, transparent metering, compliance checks and one invoice lowers the barrier. If the gateway is backed by a state platform, it can also attach policy incentives, park tenancy benefits and assurances that certain data stay inside approved boundaries. Those extras are the main claimed advantages over pure public-cloud resale.
Whether the advantages are durable is less clear. If the service reduces to discounting someone else’s models, margins will be thin and competition intense. Sustainable returns are more likely to come from higher-value layers—domain-specific agents, secure data environments, custom integration and transaction fees on a marketplace of applications. Operators themselves describe the present phase as ecosystem building; profitability is expected to follow density of users and repeat purchases rather than early mark-ups.
A provincial token trading and service center has been set up under local data authorities, linked to the Guangzhou Data Exchange. Its stated role is closer to rule-making and matching than to direct selling: defining metering standards, reviewing compliance and giving buyers a common reference for comparing offers. Such an institution could reduce the friction that currently makes every AI purchase a bespoke negotiation. It could also remain largely ceremonial if suppliers and customers continue to deal bilaterally.
For banks the experiment is equally provisional. Token burn can serve as a real-time signal of activity, but only if the underlying platform data are reliable and the borrower’s revenue ultimately converts that activity into cash. Early loan books are modest. Scaling them will require longer performance histories and clearer recovery paths when usage collapses.
The larger industrial question is whether tokens can become as legible and financeable as more familiar utilities. Metered supply, unified interfaces, compliance wrappers and even credit products keyed to consumption are all being built. The test ahead is commercial rather than technical: whether enterprises continue to buy tokens after introductory discounts fade, and whether the resulting cash flows can support both the platforms that sell them and the lenders now writing the first facilities against them. Guangzhou’s district platforms and banks are among the first to run that test in public. Others are watching the results.
Explore more exclusive insights at nextfin.ai.
