Singapore banks and the AI bubble: credit exposure

Through which credit routes do Singapore banks meet the AI bubble?

Four routes carry the exposure. Direct loans finance data centre owners and operators. Corporate loans finance technology firms that sell or buy AI services. Wealth lending finances clients who hold AI shares as collateral. Indirect links run through customers whose own income depends on AI demand, such as construction firms and equipment suppliers. Each route has its own risk and its own indicator.

The EQOS Asia AI study provides the backdrop. Valuations rose ×4 in 18 months, about $200B a year goes into AI, and the enterprise AI market scores 0.337 in the fragile zone (EQOS Asia AI study, April 2026). Rapid price rises raise collateral values, and the same speed can reverse them, so the routes need a stress view and not only a current view.

How does a data centre project loan carry AI bubble risk?

A project loan is repaid from the rent of the finished asset. The key question is who signs that rent. A loan covered by a long lease from a strong tenant has a clear path. A loan for a hall built ahead of tenants relies on a lease that does not yet exist. Lenders then ask for guarantees from the sponsor, which moves part of the risk to the sponsor.

Construction adds timing risk. If tenants delay move-in, interest accrues while rent stays at zero. A lender reads the lease cover ratio, the construction schedule, and the sponsor strength together. Any one of the three can be sound while the other two are weak, which is why the combination matters.

What credit risk sits in loans to technology and AI firms?

Technology firms borrow against revenue growth, which depends on customers who may be funded by the same market sentiment. A firm that sells to enterprises with signed contracts has a clear base. A firm that sells to start-ups funded by investors has a base that follows funding rounds. Credit files should name the customer type, the contract length, and the share of revenue from the top 3 clients.

The study adds a useful reference: 95% of AI projects fail in production (MIT 2025, cited by the study). Customers whose projects stall may delay payments or cancel contracts. A lender who tracks customer project status early sees stress before it reaches an overdue invoice.

How does lending against AI shares link private banking to the bubble?

Private clients often borrow against a portfolio of shares. When AI shares fall, the value of the collateral falls, and the bank asks for more cover. Clients who hold a concentrated set of AI names face the sharpest calls. The bank then faces a chain of calls arriving at the same time, in the same direction, which tests both its operations and its clients.

Diversified collateral changes the picture. A portfolio spread across sectors and regions loses less in one episode. Lenders therefore measure concentration by name and by theme, not only by total value. A single theme that exceeds a set share of collateral deserves a lower advance rate.

Which indirect credit links to AI do banks tend to overlook?

Indirect links hide in ordinary lending. A contractor building halls, an electrical supplier, a property owner letting offices to AI firms, and a logistics company moving servers all depend on AI demand. Each loan looks normal on its own. Together they form a theme, and the theme is visible only when the credit team tags customers by their end market.

Tagging is a simple discipline. Add one field to each customer file: share of revenue linked to AI build-out. Sum the field across the book once a quarter. The total shows the real exposure, which is usually larger than the direct loans and smaller than the loudest fears.

Which indicators show credit exposure to the AI bubble early?

Four indicators work well. Lease cover on data centre loans shows income safety. Share of the loan book tagged to AI shows concentration. Average advance rate on AI share collateral shows cushion. Loan maturities in the next 24 months show refinancing pressure. Each one is a number with a date, and each one changes before losses appear.

Set a threshold for each indicator and write down the action it triggers. For example, a rising AI-tagged share could trigger a review of limits. A thin cushion could trigger a lower advance rate. Pre-set actions turn a nervous discussion into a routine step, which is the main value of an early indicator.

How can a reader use this credit guide as of October 2026?

Bank staff can apply the four routes to their own book. Corporate treasurers can read the same indicators to see how their lenders may behave. Investors can use them to ask better questions at results time. This page is general information, not investment advice, and it gives no view on any bank.

To score your own case, the Λ simulator gives a first reading in 10 minutes. For a decision that involves AI spending or lending in Singapore, a 30-minute pre-diagnostic applies the EQOS Asia scoring and returns a score with ranked options.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile