Banks Are Now Lending Against AI Tokens

August 14. Haizhu District. Token Loan.

Put those three together, and one thing is clear: banks now use token consumption as the basis for lending. Bank of China’s Guangzhou branch launched this product with four core underwriting metrics – token consumption, value of computing service contracts, accounts receivable, and token commission settlement. Single‑borrower cap: 30 million yuan. Term: three years. The pilot has approved 28 million yuan, with 8 million actually drawn.

This is not high finance. The logic behind the Token Loan is identical to the storefront cash‑flow loans of the e‑commerce era or the order‑based loans of food‑delivery platforms: whatever orders you process in a day, that’s your credit line. Only this time, the "orders" are not GMV – they are tokens.

Token is not cryptocurrency. It is byte count – the smallest semantic unit into which a large language model chops text. You type "hello"; the model consumes a few bytes and outputs a few lines. Those bytes are tokens. AI companies sell those bytes. Whoever sells more, in the bank’s eyes, "has cash flow."

At the start of 2024, China’s daily token calls stood at roughly 100 billion. By the end of 2025, that figure had hit 100 trillion. By March 2026, it reached 140 trillion – a thousand‑fold increase in two years. If you are a bank credit officer facing an AI company with no factory, no equipment, just servers – how do you judge its ability to repay? Open its operations dashboard, glance at the token‑consumption curve, then log into its platform for a live demonstration. "Click here, we’ll watch it run," the examiner said at the press briefing. Verifying the token‑consumption ledger and doing on‑site platform displays are the core verification methods.

Haizhu District simultaneously rolled out the "Eight Measures for the Token Economy," which for the first time in a policy document included the phrase "financing based on token revenue rights." It also offers interest subsidies for AI firms taking loans through this channel – up to 2 million yuan per firm per year. Revenue rights that can be financed: that is a clear signal that future cash flows tied to tokens are being treated as pledgeable assets. From pledge to securitisation is only one step away.

Across the ocean, a different calculus is at work.

Nvidia has brought in BlackRock, Blackstone, Goldman Sachs – six top‑tier financial institutions – with plans to mobilise half a trillion dollars to package GPUs and data centres into investable infrastructure assets. The operational route: put GPUs into SPVs (special purpose vehicles), sign long‑term leases, generate cash flows, then slice and sell those cash flows to institutional investors. Jensen Huang’s narrative: compute is "fungible" and "transferable." If Company A goes bust, you lease to B; if the GPT workload dries up, you run Claude; the CUDA ecosystem keeps upgrading. So GPUs are not three‑year‑depreciation‑to‑zero servers; they are infrastructure on par with electricity and telecommunications. Infrastructure is not valued by three‑year residual value; it is valued by twenty‑year discounted cash flows.

But Michael Burry, the prototype of The Big Short, warned publicly that this is structurally indistinguishable from the moves before the 2008 subprime crisis. Underlying assets were packaged layer upon layer; risks drowned in the structure; leverage ignored in the euphoria. Bloomberg strategists put it more bluntly: half a trillion dollars of long‑term capital flooding into the compute industry will breed overcapacity and distort real supply‑demand signals. GPUs iterate every two years – so‑called "long‑term leases" get obsoleted by newer models by year three. The locked‑in customers run away; the cash‑flow model fails. CUDA can lock in developers, but it cannot lock in clients switching to cheaper inference chips.

Two paths are converging toward each other.

The Token Loan enters from the downstream, through operating data – using cash flow as a substitute for collateral, solving the financing deadlock of small and medium AI firms that have "business but no assets." Nvidia enters from the upstream, through hardware – packaging compute itself as a tradable asset, tapping Wall Street’s trillion‑dollar long‑term capital.

They are converging in the same direction: turning "compute" from a technological variable into a financial variable.

The difference is speed. Nvidia’s plan is measured in trillions; the Token Loan is currently in the tens of millions. But that line in Haizhu’s policy document – "financing based on token revenue rights" – logically connects the two endpoints. Once the revenue rights generated by token consumption are standardised, valuated, and pledgeable, the next step is to package them into ABS. By then, "compute ABS" and "token revenue trusts" will appear on the financial markets of both countries under different names, but they will be the same thing at heart.

Three risks stand out.

First, token consumption is not revenue. Two AI firms may consume the same number of tokens; one may achieve 40% gross margin through an efficient model, while the other barely survives below cost. Worse: test calls, internal debugging, free promotions, even malicious volume inflation – all can show a rising token curve on the dashboard. For now, banks rely only on self‑reported ledgers plus live demos. There is no unified, real‑time, tamper‑proof third‑party token‑attestation platform. This is not a technical challenge; it is a governance vacuum.

Second, the AI sector changes its weather every three months. The top ten token‑consuming firms today could lose their cost advantage next year to a more efficient model, or have their business boundaries clipped by a new data‑compliance policy. The traditional banking skill – extrapolating the future from historical data – faces a fundamental failure in this industry. Dynamic monitoring and periodic re‑valuation are written into the product manual, but whether post‑loan management can outrun the speed of industry change remains an open question.

Third – and the only one that could create systemic risk – the Token Loan itself is on‑balance‑sheet credit, constrained by capital adequacy ratios; it cannot cause a major disaster. But once "financing based on token revenue rights" moves into practice, non‑bank institutions, asset‑management products, and even retail investors can enter through structured products, stretching and obscuring the risk chain. If the underlying token economy suffers a systemic correction – commercialisation of large models falls short of expectations, oversupply crashes token prices – the entire chain will collapse from the far end upward. One clean lesson from 2008: over‑financialising a nascent asset amplifies its inherent volatility and manufactures credit bubbles detached from the real economy.

We are not there yet. The Token Loan remains a tens‑of‑millions‑yuan, risk‑manageable, helpful credit innovation for asset‑light AI firms. But the direction is drawn: from Token Loan to revenue‑right financing to securitisation – on every step of that path, financial institutions are already moving.

Tokens are for counting bytes, not for counting money. But when an industry counts 140 trillion bytes every day, finance will inevitably follow – and count them again, in its own way.

This fire of Token Loans needs to be stoked – and watched.