Inside a branch in Zhangjiagang, a credit officer has a problem that used to be solved by a land certificate. Two borrowers are asking for money. One cleans medical records for hospital clients. The other builds industrial drones that schedule themselves at fire scenes. Both have orders. Neither owns a factory worth pledging.
The old answer was no. The new answer, in a growing number of Chinese cities, is a bill.
Not a bill for servers. A bill for tokens — the units a model is billed in every time it answers — which the bank now reads as evidence that a business is actually running.
Key takeaways
- Bank of China's Guangzhou branch lends up to RMB 30 million (≈US$4.2m) for three years against token consumption, with RMB 28 million (≈US$3.9m) in trial credit (Guangzhou Daily).
- Haizhu district launched Guangdong's first token-economy loan on 14 August 2026, alongside eight token-economy measures with per-firm caps from RMB 1.5m to RMB 5m (Guangzhou municipal government, 7 September 2026).
- No Chinese bank in these schemes takes tokens as security; the token bill is one input to a credit score that still includes receivables, contracts and patents.
- In Suzhou, two firms were approved RMB 6 million (≈US$845,000) in combined credit under the country's first "IP plus compute-token" combined loan (Modern Express, 1 October 2026).
- China processed more than 140 trillion tokens a day by March 2026, against about 100 billion in early 2024 (National Data Administration, 24 March 2026).
The collateral problem
China's AI sector is unusually asset-light at exactly the moment it needs capital. A model developer's balance sheet is salaries, servers it may rent, and a licence; the durable value sits in a checkpoint and a customer list. Banks that lend against plant and land have little to work with.
The central bank has turned that into a policy problem. At the 2026 Financial Street Forum on 19 October 2026, the People's Bank of China said it would "further improve the financing support system for AI enterprises", focused on core technology and AI Plus (人工智能+) applications, and noted it had already issued a policy document on AI Plus finance (中国人民银行, 19 October 2026).
District governments moved first, and they moved on a specific metric. On 14 August 2026, Haizhu district in Guangzhou released 词元八条 — eight measures on the token economy (词元经济) — alongside a loan product the district described as the first of its kind in Guangdong (Guangzhou municipal government, 7 September 2026). Bank of China, China CITIC Bank and Bank of Guangzhou each brought a version.
What the bank actually reads
The clearest published terms are Bank of China's: its Guangzhou branch lends against a compute contract or a token-consumption quota, for up to three years and RMB 30 million (≈US$4.2m / HK$33m), with RMB 28 million (≈US$3.9m) in trial credit by late October (Guangzhou Daily, 22 October 2026).
Read the guarantee language. The facility can be secured by credit, receivables, order financing, or a mix with guarantees and mortgages. Tokens are nowhere on that list. What a token bill does is feed a score.
Bank of Suzhou's 苏心智算Token scheme is more explicit, converting four kinds of evidence into what it calls credit references:
- Compute purchases and token consumption, including API calls and settlement flows.
- Compute-service contracts, read as evidence of committed spend.
- Business orders, which show contracted demand.
- Intellectual property, valued separately and then bundled with the rest.
The Zhangjiagang loan bundled all four, earning the country's-first label: two borrowers were approved RMB 6 million (≈US$845,000 / HK$6.6m) in combined credit (Modern Express, 1 October 2026).
The unit of account has moved twice
Haizhu did not invent this from nothing. A year earlier the same district ran a traffic loan (流量贷) that underwrote digital-marketing firms on platform traffic, platform credentials and customer lists; it generated more than RMB 1 billion (≈US$141m) of new credit across 90 firm-visits (Guangzhou Daily, 22 October 2026).
Moving the anchor from traffic to tokens is a smaller step than it looks. Both measure volume of activity routed through a platform, and neither is tightly connected to whether the borrower earns a margin. A firm can burn tokens on retries, on evaluation runs, or on customers who never renew.
Token consumption is an activity indicator (活动指标), not a profitability indicator: a borrower with a large bill and thin margins looks busier than one with a small bill and good margins. Credit officers know this, which is why the two published versions keep receivables and orders inside the model.
The part nobody has priced
Three unknowns matter most in a downturn.
- Portfolio size. The disclosed figures are a district pilot of RMB 28 million and two Suzhou borrowers at RMB 6 million. Neither is a book.
- Default experience. No bank has published delinquency or loss data for token-linked lending; the products are months old.
- Verification. Token data comes from the borrower's own platform account and from compute vendors. Both can be inflated by calls made to raise a number rather than to serve a customer, and no published standard for verifying them exists yet.
The central bank asked for a financing support system, not token-based underwriting, and has said nothing about token data as a supervisory metric.
Honest limitations
The two products described here are district-level pilots with disclosed terms and no disclosed performance; the RMB 28 million figure is trial credit approved, not credit drawn. "First in the province" and "first in the country" are descriptions used by the district government and the lender, not third-party findings. The Suzhou borrowers are named but their financial statements are not public, so the quality of that underwriting cannot be judged. Nor is it clear whether the subsidy measures and the loans reinforce each other or simply overlap; the token volume cited above is a national total, not a lender's book.
What readers can do now
- Ask any Chinese AI supplier which credit line it used and what security was pledged; a token-linked facility usually means no hard collateral exists.
- Request the compute contract rather than the token count when testing a supplier's capacity claims — contracts are enforceable, consumption dashboards are not.
- Watch for the first disclosed non-performing loan figure on token-linked credit; until one appears, treat these products as marketing wrapped around a small pilot.