CICC: AI Capex to Generate $3.5 Trillion Financing Demand Over Five Years
N.R. Finch
CICC estimates AI investment will generate roughly $3.5 trillion in external financing needs over the next five years, as cloud giants' capex-to-revenue ratios could surge from 12% to above 40% — this means AI is shifting from self-funded growth to large-scale borrowing, and debt serviceability becomes the central test for the entire AI value chain.
Where does $3.5 trillion come from?
CICC projects five capital pools will absorb the $3.5 trillion: investment-grade bonds ~$1.5T, private capital ~$1.1T, public equity ~$0.4T, leveraged finance ~$0.3T, and securitized products ~$0.3T.
This means → bonds and private capital are the two main pillars, together accounting for over 70% of the total.
In plain terms = cloud companies used to fund data centers from their own earnings. Now earnings can't keep up — they need to borrow at scale, mainly through bonds and private capital.
Capex ratios heading to 40% — how extreme is that?
Major cloud companies' capex-to-revenue ratio rose from ~12% in 2023 to ~23% in 2025. Consensus estimates put the 2027 figure above 40%.
This reflects a "cash burn" pace approaching or exceeding historical capex-cycle peaks in the internet and energy sectors.
In plain terms = for every $100 earned, $40 goes to building infrastructure — a ratio only a handful of industries have reached in history.
How much must AI applications earn to service this debt?
CICC estimates that at a 10% return on invested capital (ROIC — how much profit each dollar of investment generates), AI applications ultimately need to produce roughly $1 trillion in sustainable annual revenue.
Reaching that scale by 2030 requires AI application revenue to nearly double every year.
This means → debt serviceability is not just a finance problem; it is a hard test of AI commercialization speed — any slowdown in revenue growth immediately amplifies debt pressure.
Margins and asset life — the two variables that matter more than revenue?
CICC calculates that at a mature-phase EBITDA margin (earnings before interest, taxes, depreciation, and amortization as a share of revenue) of roughly 50%, infrastructure depreciation life must reach ~5 years for investments to earn a positive return.
If margins fall below 20%, even long asset lives cannot cover the cost of capital.
In plain terms = revenue alone is not enough. What matters is how much profit each dollar of revenue leaves behind, and how long the equipment lasts — thin margins or rapid obsolescence turn even large revenues into losses.
A refinancing wave is coming — how does risk transmit?
CICC projects that 2027–2032 will be a peak period for cloud-company bond maturities, with roughly $28 billion coming due each year — up about 60% from the 2024–2026 pace.
Refinancing buys time but cannot substitute for cash flow. If utilization, margins, and asset life persistently disappoint, rolling debt over only pushes leverage and funding costs higher.
This means → the real test is not whether companies can borrow, but whether they can build a positive cash-flow cycle before funding costs rise and assets depreciate.
Banks profit early, risk arrives late — will a bubble form?
In Q2 2026, the six largest U.S. banks saw non-interest income rise 32% year-on-year; bank lending to non-bank financial institutions grew roughly 25%, with AI financing a significant driver.
But bank revenue is booked during the financing and construction phase, while credit risk only surfaces at the operational and refinancing stage — "revenue front-loaded, risk back-loaded."
CICC's view: AI investment is still led by large cloud companies with strong cash flows, so risk is more likely to appear first as valuation adjustments and capex slowdowns, not an immediate financial-system crisis. However, if commercialization consistently lags capex, risk shifts from valuation correction toward credit-quality deterioration, transmitting through revolving finance, private credit, and securitization channels.
Content is for reference only, not financial advice.