Baidu says its new token exchange handles more than 700 billion tokens a day, but its user agreement excludes breach compensation for supplier-side failures. On one DeepSeek listing, one supplier recorded 99.7% of five suppliers' calls, on CTOL's arithmetic.
Baidu's new token exchange checks which AI model a buyer gets, but its user agreement excludes breach compensation for supplier-side failures. On one DeepSeek listing, one supplier recorded 99.7% of the calls logged for five suppliers, on CTOL's arithmetic.
Baidu presented 词元所 (Ciyuansuo) at Baidu World on 12 November. Robin Li called it an exchange where enterprises and individuals can find, buy and use trusted model tokens, and said it had been live for about a month. Baidu reports more than 100 supply resources, over 50 million calls a day and more than 700 billion tokens a day. These usage figures are Baidu's own.
Li's pitch is trust. He said token relay stations are not responsible for model quality, stability or response speed, and that even a model's authenticity cannot be verified. But Ciyuansuo's user agreement, updated on 11 November, limits the platform to access, authentication, request forwarding, billing and marketplace display.
CTOL's view: treat the ratings as a way to narrow supplier choice and keep a backup route.
Baidu tests models but does not promise service
Baidu describes two mechanisms. Verification (验真) compares model fingerprints and runs capability, performance and availability tests. Credit assessment (征信) uses spot checks, active probes, full-link call monitoring and supplier ratings, per the mechanism report.
The rating manual defines five grades, from 优信 to 冻结, and says key results are recorded on blockchain. It also says a rating is not an absolute guarantee of real-world performance or continuous availability.
The user agreement is explicit. Section 12.1 excludes platform breach compensation for supplier-side faults, network fluctuations, model degradation or a supplier withdrawing a model. Section 17.3 caps total platform compensation at the service fees a user paid in the prior 12 months. Under section 10.4, billing for failed or timed-out requests follows the supplier's billing status and the platform's published rules.
One DeepSeek listing runs through one supplier
Ciyuansuo's DeepSeek-V4.1-Flash page for Guangzhou lists five suppliers and their seven-day call counts. 博川科技 shows 328 million calls. The other four total 839,500. On CTOL's arithmetic, 328 million of 328.84 million is about 99.7%.
博川科技 also posts the lowest price of the five, at RMB0.74 per million input tokens and RMB2.96 per million output tokens. The others range from RMB0.80 and RMB3.20 to RMB1.50 and RMB6.00. CTOL infers that price may play a part, but the page gives no reason for buyers' choices.
These are Baidu's own counters for one model, one region and one seven-day window. They cover only that slice of the exchange. They do show that five listed suppliers can coexist with one carrying nearly all the recorded calls.
Traffic alone cannot value the exchange
Dividing 700 billion tokens by 50 million calls gives 14,000 tokens per call, on CTOL's arithmetic. That holds only if both totals sit exactly at their quoted thresholds. Both are "more than" figures, so the true average is unknown.
Ciyuansuo publishes supplier prices, but The pages read give no Baidu platform fee and no split of paid third-party use. OpenRouter, a Western model marketplace, shows one fee design: it charges 5.5% on Standard and 8% on Business credit purchases, never on individual requests. That is a comparison, not an estimate of Baidu's take.
The next check is Baidu's third-quarter results. Search listings place them on 16 November. That quarter ended on 30 September, before most or all of the exchange's first month, so the results can at best show whether Baidu discloses paid third-party volume and fee receipts. Until then, the evidence favours treating the exchange as supplier selection, with service risk left to buyers.