Beijing’s Economic-Technological Development Area says six bank branches have approved nearly CNY 2 billion of credit for AI companies under its first batch of “Token loans.” The district announcement does not say how much cash borrowers have drawn.
That omission defines the story. Token consumption has entered the credit file, but the public evidence does not show that it predicts repayment. The products are an underwriting experiment for companies with few hard assets. Tokens are neither collateral nor revenue.
A credit line is not cash
The Beijing programme spans the local branches of Agricultural Bank of China, China CITIC Bank, Industrial Bank, Bank of Beijing, China Minsheng Bank and Hua Xia Bank. The district calls the nearly CNY 2 billion figure total credit support, or approved facilities, for several companies. It reports no disbursement, utilisation or loan balance.
Two examples are more concrete. CITIC gave autonomous-driving platform developer Guoqi Zhikong a CNY 30 million unsecured line after considering token use and technical strength. Industrial Bank increased IT-services provider Shenzhou Guangda’s facility to CNY 30 million after combining token use with its team, technology and growth prospects. The announcement again describes available credit, not cash advanced.
Guangdong offers a useful check. The provincial financial authority, citing the People’s Bank of China Guangdong branch, says Bank of China had approved more than CNY 28 million of Token-loan credit since July. Xinhua reported that the facility covered six small firms. An earlier Bank of China Guangzhou disclosure, reported by China Securities Journal, put approvals at CNY 28 million for five borrowers on August 14. Only three had drawn CNY 8 million, while CNY 20 million awaited supporting purchase contracts. The later six-firm count appears to reflect an additional borrower, not a contradiction.
The comparison is deliberately uneven. It shows why approved capacity cannot be reported as lending delivered. Beijing’s facility total is about 250 times Guangdong’s disclosed cash draw, but the numerator and denominator sit at different stages of the credit process.
What token data changes
Guangzhou’s Haizhu district gives the clearest account of the mechanism. Its official summary says Bank of China can set a facility from a borrower’s contract value or token-consumption quota. The bank calls consumption one core sizing input. It also considers token output, computing-service contracts, receivables and token commission settlements. The maximum facility is CNY 30 million for three years.
Those inputs sit inside familiar structures. Bank of China offers unsecured credit, receivables pledges and order finance, with guarantees or conventional collateral available to enlarge a limit. New companies need evidence of continuous operations from a predecessor or a guarantee before applying against orders. Beijing lenders likewise combine token data with the team, technical barriers, intellectual property and business prospects.
Token use changes the information set. The disclosures do not show that it replaces ordinary credit judgment.
The gaps are material. No lender reviewed has published a scoring weight, a token-to-limit conversion, an interest-rate adjustment or a monitoring trigger tied to usage. Bank of China told China Securities Journal that it currently checks borrower-supplied platform settlement lists and watches the borrower log in during an on-site review. Direct bank access to platform data is a future goal.
Xinhua says banks cross-check token data against contracts, receivables, operations, financial statements and credit records. It also quotes lenders saying there is no direct rule that more tokens produce more credit. Publicly, token burn can affect screening and facility size. Its effect on approval odds, price and continuing monitoring remains undisclosed.
The metric can be manufactured
Token burn measures an input. It does not identify who paid for the output, whether a related party generated the traffic, whether the work produced revenue or whether the borrower used an efficient model. A company can raise consumption with testing, repeated calls or verbose workflows while cash generation stands still.
The policy context adds noise. Haizhu separately offers digital-marketing and digital-media companies support based on token-consumption expense, capped at CNY 2 million per company each year. Subsidised usage may be economically rational. It is still a weaker signal of unsubsidised customer demand.
Cailian Press reported the same vulnerability after interviewing bank-credit specialists. Short-term fake orders or test calls could inflate consumption, they said, while the market lacks a uniform, auditable token-record standard. None of the product disclosures reviewed names controls for related-party traffic, subsidised calls, abnormal prompt patterns or rapid usage reversals.
Keep the signal, demote the headline
Token data can still help. A verified time series can show whether an AI service is active before a young company has mature accounts. Contracts, third-party platform logs, customer concentration, receivable collection and token cost per unit of revenue can distinguish productive usage from expensive motion.
The direction of efficiency matters too. A borrower that cuts tokens per completed task may be improving its margins. A credit model that rewards burn without measuring output would penalise the better operator. Banks usually prefer borrowers to waste less of a purchased input.
The next disclosures should report facility utilisation, loan pricing, delinquency and losses by origination vintage. They should also explain who verifies platform data and how usage is normalised across models and tasks. Until those numbers arrive, China’s Token loans show that banks are willing to look beyond buildings and balance sheets. They do not show that token burn is a proven credit signal.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.