Guangdong's AI-token venue launches with 12 suppliers and a first state-owned buyer

Against year-end goals of 100 suppliers and 50 buyers, the launch-day count is 12 signed and 25 demand-side parties. The reports read give no price, volume or contract value.

Guangdong launched the 湾数·元启 platform on 11 November. Twelve service providers signed at the event, against a year-end target of 100. The Southern Net launch report says a registration-to-API-call workflow already works. CTOL's view: treat the venue as launched but unproven until settled purchases appear. Here a token is the unit AI models use to measure what they process, not a cryptocurrency.

Guangzhou Data Exchange runs the Guangdong Token Trading and Service Center, unveiled on 30 July. Southern Net calls the platform China's first provincial-level token trading platform. Hubei's token center launched on 20 September with more than 200 model services and 326 ecosystem partners, and no reported volume.

Signing, intent and paying are three different stages

Event coverage counted 12 first-batch supplier signings, 34 suppliers with intent and 25 demand-side parties. Southern Net gives year-end goals of 100 providers, 50 demand-side users and 200 listed products. On CTOL's arithmetic, signings are 12% of the provider target, intent is 34% and the 25 demand-side parties are 50% of the buyer target. Intent is not registration or payment.

Guangdong United Electronic Services signed as the first state-owned token procurement party. Liang Yitao, general manager of its Big Data Research Institute, called it a zero-to-one step for standardized on-exchange purchasing by state-owned enterprises. No contract value or term was given.

A usage record may be the real product

A one-year Guangdong pilot in Guangzhou's Pazhou zone names a problem: providers count usage differently, so the data do not compare. Its API gateway collects usage without new development by firms. The provincial bureau intends verified usage vouchers to speed subsidy claims and support usage-based credit loans. These are intended uses in a separate pilot, with payouts still unreported.

A bank in Wuhan has already extended a RMB 3m credit line to an AI information-services startup, with the first RMB 2m to be spent on tokens, local news reported. That is another province, but it shows a lender attaching credit to token purchases.

The National Data Administration's June plan says in Article 18 to explore token trading. Article 20 encourages government departments and state-owned enterprises to budget for data purchasing. Its wording is encouragement about data purchasing, not tokens.

A token price can hide the real bill

Wang Jiandong of the State Information Center wrote that the same text yields different token counts across models. One vendor's tokenizer change raised the count for the same text by about 30%.

The DeepSeek off-peak flash tariff charges US$0.15 per million uncached input tokens, US$0.003 for cached input and US$0.60 for output. On CTOL's arithmetic, a job of 800 million input and 200 million output tokens costs US$240 with no caching. With 80% of input cached it costs US$145.92, which is 39.2% less. Both figures use DeepSeek's tariff, so they illustrate method only.

Base rates argue for waiting on volumes

Shandong's compute platform sells compute, not tokens. It reported 114 transaction contracts and over RMB 400m of transaction value. The reports read give no Guangdong transaction value.

Researchers Fu Xiwen and Zheng Lei found that most of 20 first-wave data exchanges set up in 2014 to 2017 had stopped operating by September 2025, judged by website activity. They also found that volume targets were often met by moving existing deals between state-owned firms onto the exchange. Those were data exchanges, a different kind of venue.

The platform's own targets fall due on 31 December 2026. CTOL's test is whether the operator then discloses settled purchases by buyers beyond state-owned firms, with verified API delivery, next to the counts of providers, buyers and listings. Until then the evidence favours treating the venue as launched but unproven.