Will Banks Lend to AI Companies With No Land, No Factory?

AI Was Item Eight of 19 Provisions

Reader, on September 28, China’s central bank, the People’s Bank of China, joined with the National Financial Regulatory Administration, the China Securities Regulatory Commission, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Finance, the Ministry of Commerce, and the Ministry of Culture and Tourism to release a document. It’s titled “Guiding Opinions on Financial Support for Expanding and Upgrading the Service Sector.”¹ Chinese media covering the document pulled the line “support AI+software through a variety of instruments including credit loans, bonds, and equity” as the headline.

I read this as a signal that China’s top leadership means business. The document states right in its opening sentence that it exists to carry out decisions made by the Party Central Committee and the State Council, and it bears the signatures of the central bank, financial regulators, and industrial and fiscal ministries alike.

But open the full text and the word “AI” doesn’t even appear in the title. The document lays out 19 measures in total.² AI and software make up item eight. The provision right next to it covers modern logistics and energy-saving services, and further down the list comes lodging and restaurants, elder care and childcare, cultural tourism, and even gaming, animation, and web novels. Read AI in isolation and you miss what this document is actually trying to change.

What the Document Set Out to Change in Its Opening Paragraph

The general statement spells out the document’s goal in a single sentence: overturn the notion that lending requires land and equipment as collateral, and solve the financing problem for asset-light companies. In the Chinese original, it says the goal is to shift away from the “asset-heavy, collateral-heavy” mindset and resolve the financing difficulties of “asset-light”³ enterprises.

AI and software companies are exactly this kind of asset-light enterprise. They have no factories, no land — what they have is code, people, and customer contracts. Lay out a collateral appraisal form at a bank counter, and there’s almost nothing to write in.

So the document lays out, one by one, what could substitute for collateral. It calls for reforming the intellectual property pledge system so that intangible assets can back loans, and for expanding unsecured credit loans to first-time borrowers. It also calls for building a risk-assessment and tiered-interest-rate system geared to small, asset-light companies that looks at a firm’s entire lifecycle. The list extends on to bond issuance, venture capital, listings and M&A, and even patient capital.

The “credit loans, bonds, and equity” mentioned in item eight weren’t a menu created specifically for AI. They were the menu already opened up to all asset-light companies, simply invoked again in the AI provision. Once you know that, the AI provision reads differently.

Read the AI Provision Line by Line and Two Kinds of Borrowers Emerge

Item eight’s list is quite specific. It promises to support intelligent programming tools, industrial software, foundational software ecosystems and open-source communities, and smart audiovisual systems. For new business models, it even names “Agent as a Service” and “Model as a Service”⁴ by name. It also calls for actively meeting the financing needs of service companies that help manufacturers go digital and intelligent, AI R&D, building high-quality industry-specific datasets, and expanding application scenarios.

The provision’s last sentence has a different texture. It calls for providing medium- and long-term financing to digital infrastructure such as data transmission and storage, and for strengthening financial support for the nationwide deployment of computing power, edge computing, and intelligent computing clouds.

So one provision actually contains two very different kinds of borrowers. Companies building coding tools or agents are the ones who couldn’t get loans for lack of collateral — the shift in collateral thinking that the general statement describes is aimed squarely at them. Data centers and GPU clouds, on the other hand, are asset-heavy. The machines themselves can serve as collateral, which makes this closer to the infrastructure lending banks already know how to handle. The catch is that the money stays locked up for a long time, and the value of those machines falls fast.

And AI doesn’t only sit on the borrower’s side. Item five tells banks to use artificial intelligence, big data, and the Internet of Things to build risk models suited to asset-light companies. In other words, banks are being told to use AI to find something — anything — to look at in place of collateral. The borrower is AI, and the tool for deciding whether to lend is AI too.

How Does a Document With No Dollar Figures Actually Move Anything?

Read the document start to finish and there’s no figure for how much is being funneled into AI. There’s no newly created, dedicated relending⁵ quota either. The policy tools cited in item eighteen are the service-consumption and elder-care relending facilities and interest subsidies that were already in operation. This is a document for which the word “declaration” is literally accurate.

That doesn’t mean it’s toothless, though. Item one tells nationwide banks to pick 3-5 priority directions within the service sector on their own and expand supply there, while regional banks pick 1-2 industries. The final item tells local financial-regulator offices to keep updating, together with the relevant industry authorities, a list of priority service industries and companies. Instead of a dollar figure, it hands banks homework.

It seems unlikely that AI would be left off the 3-5 directions banks choose — picking a field the central government has already named reduces how much explaining a bank has to do. That said, this isn’t an outcome the document mandates; it’s my own prediction of how banks will behave.

Where Resolve Meets Arithmetic

I’ve already run the numbers on the infrastructure-side borrower in this newsletter once before. In the September 14 issue, following Morgan Stanley’s analysis, I found that Chinese AI servers generate high margins on a per-unit basis, but Alibaba Cloud’s company-wide profit margin stayed at just 11-12%. China’s server capex was estimated at roughly three times that of an equivalent U.S. configuration. Of the three ways to monetize the same machines, selling tokens through MaaS was the most profitable — but only when inference share and utilization held up.

The guiding opinions call for providing medium- and long-term funding to exactly this segment. From a bank’s perspective, there is a machine that can serve as collateral. But that machine’s value could depreciate faster than the loan matures. Resolve can be put into a document, but repayment has to come out of the cloud business’s profit margin.

The asset-light side has the opposite problem. Banks need to be able to read the cash flow of companies building agents or coding tools, and the methods the document offers for that are intellectual property pledges, shared credit information, and AI risk models. All three point in the right direction, but none of them is yet a yardstick a bank’s loan officer can actually use today.

Oswarld’s Lens

These days I’m traveling to various sites advising on and reviewing AI data center projects. When a review request comes in, before looking at site area or GPU count, I check whether there’s a confirmed tenant. After that, I check whether the lending syndicate is actually moving on the strength of that tenant’s credit. I laid this out in a September 19 ZDNet column.⁶ In data center lending, what a bank actually trusts isn’t the building — it’s the credit of the tenant who’s going to move in.

Read through that lens, this guiding opinion looks like China trying to slot state resolve into the spot where tenant credit should be. Getting onto the list and being named a priority direction gives banks a justification to act. But justification can’t substitute for repayment capacity. That’s why I think whether this document succeeded should be judged not by outstanding loan balances, but by delinquency rates a few years from now.

What I’m watching more closely is the asset-light half. If small companies building agents and coding tools can tap bank money without selling equity, that’s a change that will outlast tens of thousands of GPUs. It’s the process by which a yardstick emerges for what, exactly, banks look at when they decide to lend to an AI company.

---

Closing

China has indeed made up its mind. But what it’s resolved to change is less an AI budget than the habits at the bank counter. A bank that learns to lend without collateral will lend the same way to the next industry that isn’t AI.

---

💬 If your company or an AI startup you know has tried to get a bank loan, what was the first thing the loan officer asked?

📨 Share this piece with a colleague who runs or invests in an AI startup. What Chinese banks are looking for instead of collateral is worth discussing together.

---

References & Further Reading

How a GPU Business Earns — and Recoups Its Money — Depends on the Sales Model · Issue 251 · INLEVEL9You can buy GPU servers and lease them out, rent them and sublease them, or charge per-token through a model API. The three approaches differ in how much money you spend upfront and how the revenue comes back.inlevel9.com

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

• Guiding opinions (指导意见, zhidao yijian): Not a law, but a policy document through which a ministry gives financial institutions direction and tasks. It often sets policy direction and implementation frameworks rather than specific dollar figures.
• Asset-light (轻资产, qing zichan): Companies with few tangible assets. They’re the most disadvantaged under collateral-based lending.
• Relending (再贷款, zai daikuan): A tool through which the central bank supplies low-interest funds for a specific purpose to commercial banks.

Footnotes

1. People’s Bank of China and 7 other ministries, full text of “Guiding Opinions on Financial Support for Expanding and Upgrading the Service Sector” (dated September 25, 2026; released September 28). Published on Sina Finance, https://finance.sina.com.cn/china/2026-09-28/doc-initkpex8571863.shtml ↩

2. 21st Century Business Herald, a Chinese financial newspaper, “Eight Ministries Roll Out 19 Financial Measures to Support Service Sector Expansion and Upgrading,” September 29, 2026. https://m.21jingji.com/article/20260929/herald/381087064846ed775f83377d47d351a6.html ↩

3. A Chinese policy term for companies with a low share of tangible assets like factories, land, and equipment, which instead run on intangible assets such as technology, talent, and brand. ↩

4. A business model where, instead of deploying the model directly, a company charges based on API calls. Agent as a Service is the same model applied to selling agents that perform tasks. ↩

5. A policy tool in which the central bank relends funds at low interest rates to banks that have expanded lending in a specific sector. ↩

6. Kwangseob Ahn, “The Essence of AI Data Centers… A Complex Four-Variable Function,” ZDNet Korea, September 19, 2026. https://zdnet.co.kr/view/?no=20260919140654 ↩