Chinese banks are quietly redefining what counts as collateral. On August 14, Guangdong launched the country's first "Token loan" — a credit product that uses a company's AI token usage and compute output as grounds for lending. Three banks signed on as initial pilots: Bank of China's Guangzhou branch, China CITIC Bank, and Guangzhou Bank. Three days later, Agricultural Bank of China rolled out its own Token loan plan at its Xuhui tech branch in Shanghai.
On the surface, this is a niche regional banking product. Look closer and it is a signal that the AI compute economy is crossing into mainstream finance: usage data is becoming a credit credential, and banks are positioning themselves as the intermediaries between capital and compute capacity.
The First Token-Backed Loan Products
Guangdong's Token loan targets companies across the compute supply, application, and service chain, covering businesses of every size. Instead of asking for physical collateral, it bases credit lines on four things: the enterprise's compute token output and consumption, the value of its compute service contracts, receivables generated by compute business, and token commission settlement volumes.
Bank of China's Guangzhou branch offers up to RMB 30 million per company, with terms up to three years. Its Haizhu sub-branch has already extended more than RMB 400 million in financing to compute-related companies, including over RMB 28 million in Token loan credit since July, brokered through government-backed matchmaking. In Shanghai, Agricultural Bank of China's new plan treats token demand as a key credit input and has put it to work in the "Yuanli Community" startup cluster, funding a local AI company's token procurement and daily operations.
How a Token Loan Actually Works
These loans do not run on token usage alone. Bankers are explicit that token consumption is a supplementary signal, never the only one. High token spend suggests a product has market pull — usage is an operating cost, and rising usage usually means revenue is following — but usage does not prove profitability. Token types vary widely in price, and traffic can be inflated with fake or "washed" calls.
So banks combine daily and monthly token consumption with order flow, cash flow, receivables, and revenue to judge a company's overall health. Agricultural Bank of China adds a full balance-sheet review: tax records, credit history, financials, contracts, and supply-chain position.
Why now? National data bureau figures show daily token calls in China passed 140 trillion by March 2026 — a thousand-fold jump from roughly 100 billion in early 2024, and up more than 40% in just three months since late 2025. When usage data reaches that scale, it becomes expensive to fake at scale, and banks can treat it as a rough proxy for real business.
The Real Shift: Compute Output as a Credit Signal
Think of this as the lending equivalent of the move from brick-and-mortar collateral to SaaS annual recurring revenue. For years, lenders extended credit against assets you could touch — real estate, equipment, inventory. The Token loan extends credit against a verified flow: how much compute a company actually consumes and how much it sells. This is a structural change for the light-asset AI industry, which has consistently struggled to borrow because it owns almost nothing physical to pledge.
Banks are effectively deciding that compute usage is a measurable, auditable value credential — and that an AI company's API bill can say more about its prospects than its office lease. That is why compute financing, which earlier looked like the domain of chip suppliers — NVIDIA’s 25% residual-value support already turned compute into an investable asset class — now has a banking channel of its own. It also extends the argument that token usage has outgrown technology and become a unit of economic value, a thread we traced in our look at Tokenpolitik and the global AI infrastructure race.
The shift is not confined to one product. Anhui province is explicitly encouraging "Token loans" and "model loans" plus pledges on data IP and algorithm copyrights. Wuhan announced a compute-loan facility up to RMB 10 million with up to 80% risk compensation. Region by region, the pattern is the same: compute capacity and usage data are being folded into the credit system.
Risks the Banks Are Already Worried About
The people building these products name three risks without prompting. First, data manipulation: token calls can be inflated with fake or no-op requests, and the industry has no unified standard for what counts as genuine. Second, volatility: token consumption is spiky and project-driven — usage collapses once a project ends, which can distort post-loan monitoring. Third, profitability: some companies burn impressive token volumes while earning little.
The proposed fix is to lean on the infrastructure layer. If banks share data with cloud providers and model platforms, real API behavior — call frequency, time-of-day patterns, input-output ratios — forms a fingerprint that is expensive to forge. When faking costs more than operating, the incentive to fake collapses. The obvious tension: token data standardization and authenticity verification still have no shared rules, which is why pilots favor teams with clean, traceable usage history, often spin-offs from big tech companies.
What This Means and What to Watch
For AI founders, the practical takeaway is that your compute bill is becoming a financial asset. Companies with consistent, traceable token usage — and the revenue to match it — will increasingly be able to borrow against it, at better rates and faster approval than the old asset-based system offered. That is a genuine unlock for light-asset startups that could previously show banks little more than a burn rate, especially as the industry has leaned on huge borrowing — nearly two trillion dollars of off-balance-sheet AI debt — to fund the build-out.
For the industry, watch three signals. First, data standardization: if the market agrees on how to define and verify genuine token usage, the products scale; if not, they stay niche. Second, cross-provider data sharing between banks and cloud or model vendors — the decisive enabler for fraud control. Third, how these loans survive the first down-cycle of compute prices; the products are being designed in a bull market for AI usage, and their first stress test is still ahead.
This is not a crypto story. It is a story about the compute economy graduating into the banking system — and about data replacing assets as the raw material of credit. Banks themselves describe the end state: shifting from pure money intermediaries toward "smart intermediaries" that connect financial resources to the compute ecosystem.
