AI Companies’ Token Spending Gives Rise to Bank “Token Loans”
Bank of China, Industrial and Commercial Bank of China, Bank of Chengdu, and other banks have launched loan products related to computing power and Tokens, using data such as Token consumption, procurement contracts, business orders, revenue, and cash flow to assess credit for AI companies. Some companies have obtained loans ranging from RMB 1 million to RMB 10 million, although actual amounts are mostly several million yuan, and borrowers typically already have established business operations and historical transaction records. The article notes that Token consumption does not directly reflect business quality, while verifying the authenticity of transaction records could also become a new risk-control challenge for banks.
Why read this
AI score 69
Drawing on implementation cases from several banks, the article outlines the conditions under which Token usage data can be incorporated into credit models, the types of companies suited to such loans, and potential verification risks.
Topics
Body · AI translation
AI Companies Burn Tokens, Giving Rise to a “Lending Business” Xiaguang AI Lab November 19, 2026, 06:35:51
On the one hand, venture capital is becoming increasingly concentrated in the hands of a few giants. According to IT Juzi, in the third quarter of 2026, 43 mega-deals worth more than RMB 1 billion each—just 1.2% of all deals—accounted for 45.8% of total funding. Meanwhile, 881 small financing rounds worth less than RMB 10 million each contributed only 0.4% of the total. The median amount raised per deal remained RMB 20 million, unchanged from last year. There is indeed more money, but most of it is flowing to industry leaders; the vast majority of small and medium-sized AI companies are not getting any more than before.
On the other hand, AI companies’ funding needs are themselves changing. Unlike traditional internet companies, AI companies spend the bulk of their money on continuous, frequent purchases of computing power. Behind this strong demand for computing resources, the financing difficulties faced by AI businesses are becoming increasingly pronounced. Filling these short-term working-capital gaps by giving up equity is too costly, while obtaining conventional bank loans is difficult because these companies are asset-light, have weak financial statements and fragmented cash flows, and cannot offer factories, equipment or similar assets as collateral. As a result, some AI companies have begun considering a different approach: Since they burn Tokens every month, could the Tokens they burn be converted into credit recognized by banks?
“Token loans” emerged in response. As the name suggests, companies present banks with information such as their monthly Token-usage records, computing-power procurement contracts and business orders in exchange for loans that cover short-term funding gaps.
At present, Bank of China, Industrial and Commercial Bank of China, Bank of Chengdu and others have all launched related products of their own—including BOC Computing Power Token Loan, ICBC Intelligent Computing Connect and Computing Power Loan—and there are already real-world cases of these products being issued.
From burning Tokens to borrowing against Tokens, a business that originally involved only settlements with model providers has sprouted a financial branch.
From Burning Tokens to Borrowing Against Tokens
One VC investor said that Token usage has now become an important dimension of due diligence.
“More and more AI companies are now delivering products in the form of Agent Harnesses, and many of the Tokens are actually consumed through real calls made by end users. To some extent, this data can therefore serve as a reference for product activity and depth of use.”
Another investor at a US-dollar fund also said that when evaluating projects, the fund looks at a company’s Token consumption, although the data itself leaves considerable room for interpretation.
For the same amount of Token consumption, computing costs may vary greatly across different models and tasks. Even with the same volume of business, companies may differ completely in their efficiency in model selection, inference optimization and computing-resource scheduling.
The figure can therefore reflect the scale of business activity, but it does not necessarily reflect operating quality, much less directly equate to commercial value.
Sometimes, greater Token consumption may instead indicate that a company still has room to improve its cost control and technical efficiency.
Investors therefore care more about how much commercial value these computing investments ultimately create, including metrics such as genuine paying demand, revenue growth, gross margin and customer retention.
“For some Agent Harness companies in particular, gross margin reflects the combined outcome of subscription allowance pricing, average Token usage per user and the cost of Token calls. The higher the gross margin, the more value the Harness layer provides,” the VC investor said.
To put it plainly: How many Tokens are burned indicates whether users are showing up; how high the gross margin is indicates whether the business is viable.
VCs therefore never assess Tokens as an absolute figure. Tokens generated alongside orders and customer payments are fundamentally different, in an evaluation, from Tokens accumulated through free allowances.
Banks, however, apply a different logic when examining this Token bill.
Data show that at the end of the second quarter of 2026, commercial banks’ net interest margin stood at 1.41%. It had declined continuously after falling below 2% in 2022 and did not record its first quarter-on-quarter increase until this year. In Q1 this year, commercial banks’ net profit fell 3.73% year on year.
Year-on-year growth in commercial banks’ first-quarter profits, Q1 2022–Q1 2026. Data sources: Ping An Securities and TF Securities
It is becoming increasingly difficult for banks to make money, and the areas in which they can extend loans are gradually shrinking.
Businesses are reluctant to invest and residents are reluctant to spend—in the first half of this year, RMB-denominated loans issued to the real economy increased by RMB 10.76 trillion, RMB 1.98 trillion less than in the same period last year.
Moreover, traditional core lending businesses are also slowing.
Data from the National Financial Regulatory Administration show that at the end of the second quarter of 2026, the outstanding balance of inclusive loans to micro and small enterprises was RMB 38.9 trillion, up 8% year on year, with the growth rate down 4.3 percentage points from the same period last year.
Yet most of these companies lack the common assets held by traditional businesses.
In the past, when making corporate loans, banks first asked three questions: Do you have a factory? Do you have equipment? What do your financial statements look like? Traditional businesses could answer these questions—the equipment and inventory were right there, visible and tangible.
At the same time, the nature of entrepreneurship has also changed. Vibe Coding has lowered technical barriers, enabling more people to build products themselves, while OPCs have emerged one after another. They have users and orders but no physical assets, making them virtually invisible within the traditional credit system.
A deeper change lies in the cost structure. In the internet era, infrastructure was virtually free, with the cost spread so thinly across each user that companies could acquire users first and think about monetization later.
But the AI era is different. Model calls, Token consumption, and computing-power procurement all cost money from the very first day a product launches; infrastructure is not cheap. Every additional user adds to the bill, Token costs grow linearly, and economies of scale are not particularly evident.
This has created an awkward situation: the better the business performs, the tighter its finances become.
And those bills are growing ever larger.
At the beginning of 2024, average daily Token usage nationwide was about 100 billion; by March 2026, that figure had surpassed 140 trillion.
It was in this gap that Token transaction volume came into view. This data sitting in API backends has also become a third dimension beyond financial statements and collateral.
But banks do not look at Tokens alone. Loans ultimately have to be repaid: whether a company has stable revenue, whether orders can be converted into cash receipts, and whether cash flow can cover principal and interest remain hard requirements.
For companies themselves, Tokens were originally merely an expense on the cost ledger—a “liability.” Now that Token usage is being priced, verified, and incorporated into credit assessments, it can also be read as a credit credential, an “asset” that can be financialized.
How are credit lines worth several million yuan approved?
First, consider some figures on Token loans that have already been issued:
Jiaxing Rongyi Cultural Creativity, an AI short-form comic drama producer, received RMB 1 million from Bank of Jiaxing to pay for Token purchases;
Zhejiang Cheng'an Big Data, a smart-city company, received RMB 2 million from Agricultural Bank of China's Wenzhou Lucheng Sub-branch to fill a funding gap incurred while building an intelligent computing center;
Lightbox Power, an AI short-form comic drama production contractor, received RMB 5 million from Bank of China's Guangzhou Branch to purchase computing power and expand production;
A Chongqing company that operates a cloud-computing digital platform received RMB 5 million from ICBC's Chongqing Free Trade Zone Branch to purchase computing resources and develop its cloud platform;
Xuntu Technology, which develops large financial models, received RMB 10 million from Bank of China's Hongkou Sub-branch for large-model R&D, Token purchases, and laboratory construction...
Xuntu Technology focuses on intelligent investment research collaboration systems for the financial sector. Image source: Xuntu Technology's official website
It is not difficult to see that banks across the country are actively advancing pilot programs. At this stage, however, lending limits remain relatively conservative. Some products advertise a maximum of RMB 30 million per borrower, but the amounts actually disbursed are generally only several million yuan.
These figures demonstrate that the loans are being implemented, but the process is actually more noteworthy: how exactly does a monthly Token bill become money that a bank can lend?
“Since February this year, the company has shifted its business from live-action short dramas to AI short dramas and comic dramas, and it now burns through about 15 billion Tokens per month,” said Liang Jian, CMO of Lightbox Power.
But that figure is still far from sufficient for the company's current stage of development.
He ran the numbers: producing a premium short drama for TikTok costs RMB 1,800–2,000 per minute. Because the cost of repeated generation attempts is relatively high, Token consumption accounts for about 12% of revenue.
Lightbox Power filming short-drama content. Image source: Lightbox Power's official website
That sounds like a fairly normal level, but the key point is that once an order is received, the money for those Tokens has to be paid upfront. For a company, that is no small sum.
Faced with this upfront funding pressure, Lightbox Power began seeking external capital in the market.
Liang said the company had approached many investment institutions over the past year or so and had undergone due diligence, but the process often stalled at the investment-decision stage. That was because investors placed great emphasis on financial statement data, including historical revenue flows and profitability metrics.
In 2025, Lightbox Power recorded just over RMB 20 million in revenue, which was not particularly high.
With the equity-financing route temporarily blocked, the company turned its attention to banks. But without collateral, it would have been difficult to secure even RMB 2 million or RMB 3 million through conventional lending channels.
Just as the company was mired in Token anxiety, matters took a turn for the better.
In April this year, the government of Guangzhou's Haizhu District organized a meeting with banks and investment institutions in the district. As an industry benchmark company, Lightbox Power received government endorsement and successfully entered the banks' whitelist—financing that had previously been unavailable became accessible, while companies that could previously borrow only a little could now borrow more.
In August, Bank of China's Guangzhou Branch launched the “BOC Computing Power Token Loan” in Zhuhai District. As one of the first companies approved, Lightbox Power obtained a RMB 5 million credit line, signed the agreement, and received the funds that same month.
Less than a month elapsed between the product's launch and the loan disbursement. What exactly did the bank examine before approving the money?
“The bank certainly did not look only at a Token consumption figure,” Liang said. Tokens are comparable to electricity bills in the industrial era: in the computing-power economy, they are a cost indicator. An electricity bill means nothing in isolation. Only when electricity expenses and accounts receivable are both rising does it show that the production line is operating at full capacity.
Therefore, in addition to real-time Token consumption data from the backend, the bank also monitors annual framework procurement contracts with partners, supply contracts with downstream platforms, monthly accounts-receivable reconciliation statements, confirmation emails for financial transactions, and other materials.
Pricing, of course, also changes accordingly.
Liang noted that conventional unsecured business loans carry an interest rate of about 3.5%, while Token loans offer a more favorable rate of around 3%–3.1%.
Guanghe Dongli’s Token loan has a three-year term and is repaid in six installments, with principal repaid every six months for the first five installments. Compared with the equal principal-and-interest repayment method used for traditional loans, this somewhat eases the repayment burden in the early stages.
After the funds arrived, the company poured all of them into production: its team grew from just over 30 people in May to around 140, its premises expanded to four times their previous size, and its monthly Token consumption rose from 10 billion in August to 15 billion now.
However, the fact that the loan was approved smoothly does not mean the threshold was low.
“A company has to reach a certain stage of development and become an industry benchmark before it can qualify,” Liang Jian said.
Banks still essentially manage risk based on a company’s historical data, such as Token usage records and revenue flows, before deciding whether to lend and how much. Without such data, technology companies still in the cold-start phase are currently unlikely to qualify for the candidate list.
This is also the harsh reality at present: Token loans may not go to the startups most in need of money, but rather to those that are already up and running.
When Token usage records start being “dressed up”
“Token loans are an innovation by Chinese banks and a positive experiment that has effectively helped some emerging technology companies grow from small to large and from large to strong,” Liang Jian said when discussing the future of Token loans. “In the AIGC industry, Token loans may become a standard loan product for genuinely high-quality companies.”
This view aligns with that of a VC investor: Tokens are a major component of AI companies’ costs and are also a relatively standardized product, giving financial institutions effective means of oversight and allowing greater flexibility in lending and the use of funds.
For companies, such loans can support them until their products reach customers and begin generating revenue, and they may eventually become one of the common forms of borrowing.
Another US dollar fund investor, however, held a starkly different view. He said that many AI companies previously relied on equity financing, but equity financing takes longer and is also relatively costly. For companies that have already reached a certain scale of operations and have predictable revenue and cash flow, supplementing working capital through bank loans is a fairly natural choice.
But that does not mean it will develop into a widely adopted lending model.
“Historically, financial institutions have often introduced specialized credit products for emerging industries, such as loans based on intellectual property, orders, accounts receivable, or specific equipment assets. To some extent, Token loans follow a similar logic: they are financial products tailored to a particular stage in an industry’s development. From a bank’s perspective, the fundamental logic of lending has not changed. Ultimately, it still comes down to whether a company has a stable source of repayment, reliable creditworthiness, and appropriate risk-control measures.”
To put it in Liang Jian’s words: it cannot be called a groundbreaking, epoch-making breakthrough, but it can be described as an institutional innovation.
Under this logic, Token loans may play a role at a certain stage and among certain types of AI companies. But as the industry develops, they will ultimately be absorbed into more conventional corporate credit and financing systems.
However they evolve, once Token loans move toward the mainstream, they will inevitably encounter a familiar figure in bank lending: intermediaries.
Intermediaries may be involved in any bank lending business, because companies cannot possibly know the approval preferences of every bank.
Most intermediaries merely help companies organize their materials and reconcile their transaction records. After all, banks still fundamentally make their assessments based on past financial statements and detailed accounts receivable, both current and projected.
However, once Token usage records begin to affect credit limits, some people will want to make those records look a little better.
At that point, Token data may go from being evidence to being evidence that needs verification. Whether and how this evidence should be verified is work that both banks and companies will need to undertake.
There is currently no consensus on these future directions and risks, but one thing is certain: the value of Tokens has fundamentally changed.
They are no longer merely a line item in a cost ledger, but a field in a credit-assessment model.
This shift matters more than the size of any individual loan. It means that, for the first time, burning Tokens and lending against Tokens have become two businesses that connect: the former turns computing power into content, while the latter turns transaction records into cash.
As for how far this chain can go, it will depend on the most basic consideration of all—whether companies can turn their Token investment into real business and direct capital toward growth that creates genuine value.
AI Trends