China’s banks are using AI tokens to assess credit. Is it risky?

In 2026, a novel development emerged in China’s banking sector: banks began using firms’ token consumption to determine how much credit to extend. This development is not merely a hypothetical lending model. On 14 August, Guangzhou’s Haizhu district launched the “token loan”, the first financial product in Guangdong province dedicated to the token economy. Bank of China, China CITIC Bank and Bank of Guangzhou launched their respective products that day. 

From assets to token consumption

Bank of China’s Guangzhou branch has already approved credit facilities totalling approximately 28 million RMB (US$4.2 million). Earlier, in June, Agricultural Bank of China’s Shanghai Xuhui Technology sub-branch completed its first token-loan transaction. 

Wenzhou, Inner Mongolia and other localities subsequently followed suit. China Construction Bank, Bank of Jiangsu, Shanghai Rural Commercial Bank and Zhangjiagang Rural Commercial Bank have also introduced computing-related credit products, incorporating computing capacity and efficiency, token consumption and related indicators into credit assessments. 

In essence, token-based lending incorporates firms’ token consumption, the value of computing-service contracts, receivables from computing services, and the volume of token-related commission settlements into credit assessment. It represents a shift from assessing what assets firms own to assessing how much business activity they undertake. Legally, these products remain working-capital loans, provided primarily on an unsecured basis, against pledged accounts receivable, or through purchase-order financing arrangements. What has changed is the underwriting basis.

A simple analogy helps explain this lending logic. Suppose a bank is considering lending to a factory that has no premises to offer as collateral. If the factory’s monthly electricity consumption is stable and steadily increasing, can the bank infer that it is operating normally? Token-based lending follows a similar logic: artificial intelligence (AI) firms lack factory buildings and production lines, but the tokens they consume when accessing large language models provide, to some extent, an indication of business activity. 

Why token data may not be reliable for lending

According to the National Data Administration, China’s average daily token usage rose from 100 billion in early 2024 to 140 trillion in March 2026 — more than a thousandfold over two years. AI services increasingly position tokens as “units of measurement, settlement and statistical reporting”. For “electricity consumption” to become a reliable credit indicator, however, at least three conditions must hold: a neutral party must record meter readings, electricity prices must be stable, and all parties must share a common definition of a kilowatt-hour. Token-based lending currently falls short on all three conditions.

The use of operating data in place of collateral did not emerge in isolation. It is the latest development in the evolution of credit provision in China. The 1995 Guarantee Law established a legal basis for pledging intellectual property rights; the 2007 Property Law introduced pledges of accounts receivable; and the 2020 Civil Code expressly permitted pledges of “future accounts receivable”. 

Purchase-order financing, on which token-based lending relies, was not explicitly recognised until then. Three waves of innovation are particularly relevant: lending secured by intellectual property, which began with pilot programmes in 2008 and reached cumulative disbursements of 255.57 billion RMB (approx. US$36 billion) by 2024; bank–tax authority cooperation, launched in 2015, which had supported 45.2 million loans totalling 15.7 trillion RMB by the end of 2025; and data-asset financing, which began with a first transaction in Guiyang in 2016 but has remained confined to a credit-enhancement role. 

Of these, bank–tax authority cooperation is the only precedent to have achieved scale. Its success rests on a critical condition: the data source is neutral, authoritative and directly accessible. Government agencies generate tax data and exchange it directly with banks through dedicated data connections; they do not permit alternative transmission channels. This arrangement ensures that borrowers cannot manipulate the data. 

By contrast, token-based lending relies on data generated by computing-service platforms, which also receive the loan proceeds and administer local government subsidies that help companies pay for computing resources and AI models. These companies are neither neutral nor vested with statutory authority, and they have yet to establish direct data connections with banks. Viewed in historical perspective, token-based lending therefore appears closer to the trajectory of data-asset financing than to that of bank–tax authority cooperation. This raises questions about whether the data used for lending is sufficiently independent and reliable.

How token lending creates a self-reinforcing credit loop

Midstream cloud platforms occupy a pivotal position in the token-lending ecosystem. Banks lend to application-layer firms, with the proceeds generally earmarked for computing services. These firms use the loans to purchase tokens and computing services from midstream cloud platforms, which in turn buy or lease computing resources from upstream suppliers. Credit therefore flows from banks to application-layer firms and then upstream to cloud platforms and resource suppliers.

Yet midstream platforms also control key parts of the lending process. They issue token invoices and settlement records, refer borrowers to banks, receive loan-funded payments and administer subsidy-voucher redemption. As data providers, referral channels, loan beneficiaries and subsidy-redemption agents, they have incentives to present usage data in ways that support greater lending.

A feedback loop could emerge. When a platform cuts prices, customers may consume more tokens, boosting the usage figures used to assess their creditworthiness. Banks may then extend more credit, which borrowers use to purchase more services from the same platform. The platform thus benefits from both increased usage and the additional lending it helps facilitate.

Risks of using token usage as datapoints

Token consumption, however, measures the cost of AI inputs rather than output. It tells a bank how much a firm has invested in those inputs, not whether it has generated a profit. Zeng Gang, deputy director of the National Institution for Finance and Development, makes a pertinent observation: establishing a statistical relationship between token consumption and credit risk requires sufficiently large samples and historical data spanning multiple cycles. At present, “the accumulated data are insufficient to support accurate pricing”, and token data serve “more as an initial threshold for customer screening than as a substantive dimension of quantitative credit assessment”.

These limitations create a few principal risks. The first is data distortion. Platforms produce token data unilaterally, without independent third-party record verification or auditing. Test calls and invalid requests can inflate measured consumption, while firms can use configuration switches to expand usage of the “thinking tokens” generated by reasoning models. Industry experts have also explicitly warned that firms can inflate token consumption through sham transactions or invalid calls, and that unified audit standards are lacking.

The second risk is a disconnect between usage and business performance. Higher token consumption does not imply profitability. AI projects often show sharp usage surges at launch, followed by declines as they mature, with fluctuations far greater than those in traditional firms’ revenues. The Agricultural Bank of China has already disclosed a “dynamic early-warning mechanism based on computing-resource consumption”. If a single month’s usage triggers an alert — or even a loan recall — firms may be required to repay precisely when they face their greatest liquidity constraints between projects. Conventional business lending relies on annual reports, which smooth fluctuations; token-based lending relies on monthly usage, which amplifies them.

The third risk is the depreciation of computing assets. Graphics processing units (GPUs) undergo rapid technological turnover, while computing-resource valuation, pledge registration and secondary-market transfers remain insufficiently standardised. In June 2026, Changjiang United Financial Leasing sued Sichuan Bingji Technology. Four enforcement proceedings involved a combined 560 million RMB and affected nearly 15 leasing institutions. This episode provided the most direct upstream warning, two months before the August launch of token-loan products.

Global parallels in AI and data-driven lending

Outside China, no direct equivalent to token-based lending exists, but two financing channels have similar structures while extending credit against different underlying assets or claims. 

The first is GPU-collateralised debt. In 2023, American cloud-computing company CoreWeave pioneered a debt-financing model secured by H100 GPUs. By June 30, 2026, it had accumulated approximately US$35 billion in debt. The rating rationale for its US$8.5 billion loan is instructive: Moody’s assigned an A3 rating not on the strength of CoreWeave’s standalone creditworthiness, but on Meta’s offtake agreement. As one analyst observed, “the contract is the real asset.” Nvidia has even provided Australian AI cloud-computing company Sharon AI with a six-year minimum-revenue guarantee totalling US$4.88 billion. Nevertheless, GPU residual values remain contentious: cloud rental rates for H100 GPUs fell by approximately 64% between 2023 and 2025.

The second channel is financing based on platform-recorded sales. Cloud-based e-commerce platform Shopify’s financing program, Shopify Capital, bases its lending decisions on platform gross merchandise volume data and collects repayments as a share of daily sales. It disbursed US$4.2 billion in 2025, with 91.9% of its portfolio performing at year-end. The model combines direct access to platform data with automatic repayment collection. Here, too, the platform supplies the data, but its conflict-of-interest structure differs from token-based lending. Shopify does not itself sell goods directly to borrowers, whereas computing-service platforms are both data providers and suppliers of the financed services. 

International regulatory signals also merit attention. In July 2026, the Bank of England noted that bond issuance by AI hyperscalers in the first half of 2026 had already exceeded the full-year total for 2025. The OECD estimated that private credit’s share of financing for AI investment had risen from 9% to 34% within a year.

Can token-based lending stand the test of time?

Taken together, token-based lending benefits midstream platforms, has mixed implications for application-layer firms, and supports upstream computing providers in the short term, although its medium-term effects remain uncertain. Its most valuable by-product is standardising token measurement and procurement contracts, while the biggest concern is the divergence between incentives to increase usage and the goal of improving efficiency. 

In March 2026, the National Data Administration formally adopted ciyuan (词元) as the Chinese term for “token”. In May, it convened a symposium on the token economy and explicitly incorporated the development of this economy into its work programme. The China Academy of Information and Communications Technology has released a Standards Framework for High-Quality Token Services and Specifications for Operational Management Capabilities. 

However, financial regulators have yet to recognise these as audit standards. At present, outstanding token loans account for about 0.01% of financing across the AI value chain, while publicly disclosed nationwide disbursements total no more than a few tens of millions of RMB. Their value lies in providing an initial set of labelled observations, not in the outstanding loan balance.

Returning to the electricity-consumption analogy, the viability of token-based lending depends, at the microeconomic level, on whether the “meter” is subject to manipulation and whether the “electricity price” is stable. At the macroeconomic level, it depends on whether a credit mechanism based on usage data can withstand a complete economic cycle.

Related: Why Chinese banks now act like local governments | Shadow banking under fire: AVIC Trust’s collapse