Banks Are Turning AI Compute Bills Into Working Capital
NextFin News — For years, artificial-intelligence startups in China raised money the same way consumer internet firms did: sell equity to venture funds. That path has narrowed. Large checks cluster around a handful of winners, while most smaller rounds stay modest.
At the same time, the cost structure of AI does not wait for a fundraising cycle. Tokens, API calls, and cloud capacity are paid continuously. Growth often means a larger bill before it means a larger bank balance.
Into that gap, commercial banks have introduced a new label: loans tied to token and computing spend. Bank of China, Industrial and Commercial Bank of China, and regional lenders have rolled out products that treat usage records, purchase contracts, and customer orders as evidence of real activity.
Disbursements reported so far are mostly in the low millions of yuan, not transformative balance-sheet events. The signal matters more than the size. A cost line that once lived only in a model vendor’s invoice is being read as a credit input.
Venture investors do look at token volume, especially when end users drive agent-style products. Volume can hint at engagement. It does not, by itself, prove a business. The same usage can mask inefficient routing, generous free tiers, or thin willingness to pay. Investors still care more about paid demand, gross margin, and retention than about raw consumption.
Banks start from a different problem. Net interest margins have been tight, loan growth to the real economy has cooled, and many conventional borrowers are cautious. AI-related firms look like a growth sector, yet they fail the old corporate checklist.
There is often no factory, no heavy equipment, and little inventory to pledge. Teams, code, and data do not sit comfortably in a collateral file. One-person or very small product companies can show users and orders while remaining almost invisible to classic credit models.
AI also reverses a familiar internet pattern. Online software once scaled on thin infrastructure cost per user. Model inference scales closer to linear cost. More customers can mean more immediate cash outflow. Firms that are “doing well” on usage can feel poorer on cash. That is a working-capital problem. Solving it by selling more equity is expensive. Ignoring it stalls production.

Reported cases span AI short-drama studios, smart-city vendors, and financial-model developers. Typical uses are paying for tokens, storage, or capacity expansion. One AI drama producer that shifted from live-action work described monthly token use in the tens of billions of units and token costs around low double digits of revenue on premium overseas jobs—costs that must be paid up front when an order lands.
Equity talks stalled on thin historical revenue. Ordinary unsecured credit was hard without collateral. After local government introductions put the firm on a bank shortlist, a dedicated “compute token loan” from a major bank’s Guangzhou branch delivered a few million yuan within weeks of product launch, at a rate slightly below ordinary credit lines and with a repayment schedule lighter in the early periods.
Bankers do not underwrite a single usage chart. They pair API or console data with annual purchase contracts, downstream supply agreements, monthly receivables, and payment confirmations. Token spend rising alongside receivables is treated more like a full production line than spend rising alone.
History still gates access. Cold-start companies with little track record remain largely outside the pilot lists. The product helps firms that are already moving, not those that have not yet begun.
Participants describe the innovation as useful, not revolutionary. Finance has long built credit around invoices, orders, intellectual property, or specialized equipment. Token-linked lending is another adaptation to a sector’s dominant cost.
As the industry matures, many of these borrowers will be expected to graduate into ordinary working-capital lines. Intermediaries that help package files for banks will appear, as they do in every credit market. Once usage affects limits, some borrowers will try to make usage look stronger than the underlying business. Then the same data that unlocked credit becomes something banks must verify, not merely observe.
National daily token calls have grown from roughly 100 billion early in 2024 to more than a trillion by early 2026, according to figures cited in industry coverage. That scale makes the bill hard to ignore.
Whether loans tied to it become routine depends on a simpler test: whether token spend converts into cash collections reliable enough to service debt. Banks are not buying the romance of AI.
They are testing whether the meter on the model can stand in, carefully, for the warehouse receipt they used to hold.
更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App
内容来自 钛媒体:引领未来商业与生活新知,版权归原作者所有。
阅读原文 ↗