Hook
Meta sold $30 billion of bonds in October 2024. The largest corporate debt deal of that year. Oracle followed with roughly $18 billion in early 2025. Both framed as data-center and compute financing.
That is the tell. The AI arms race is no longer funded by cash flow. It is funded by credit.
I've watched this pattern before. Not in AI. In 2022, when Terra's oracle feeds went stale and Curve pools drained in minutes. The mechanism differs. The structure rhymes. When an asset's funding shifts from internal cash to external credit, the risk migrates from the income statement to the balance sheet. And the balance sheet does not forgive quietly.
Context
Numbers first. The four largest US hyperscalers — Microsoft, Alphabet, Amazon, Meta — ran combined capital expenditure near $150 billion in 2023. By 2025 guidance, that figure crosses $300 billion. AI-attributable revenue — cloud AI services, Copilot seats, inference APIs — has not scaled on the same slope.
The gap gets filled. Operating cash flow covers part. Debt and lease structures cover the rest.
Leasing, not just bonds, is part of the structure. Operating leases for data centers and GPU capacity function as debt with better optics. Add them to the headline bond figures and the real financing need is larger than any single prospectus shows.
This is not speculation. Meta's bond was one of the year's largest corporate offerings, with tranches stretching to 40 years. Oracle's debt load is now a standing item in its credit profile. And beneath the headline bonds sits a quieter layer: special-purpose vehicles, joint ventures, sale-leaseback arrangements. Meta's $27 billion data-center joint venture with Blue Owl is the template. The compute asset sits off the core balance sheet. The obligation does not fully disappear.
The core contradiction is a timing mismatch. AI capital expenditure now exceeds any verifiable AI revenue, and the difference is being financed with debt rather than profit. That single shift moves the cost of failure from "earnings miss" to "credit impairment." Those are different animals.
Core
Here is where the tape gets interesting.
Bond investors are not equity investors. They do not get upside. They buy downside protection. So when credit buyers start asking harder questions about AI capex, they are pricing something equity holders are still ignoring.
The signal is in the term structure. Long-duration paper — the twenty- to forty-year tranches — is where hesitation shows first. The reason is mechanical. GPU depreciation cycles run three to five years. Data-center debt runs twenty to forty. That mismatch is structural, not cyclical.
A GPU cluster financed in 2024 is economically obsolete long before a bond maturing in 2064 makes its first coupon adjustment. The asset depreciates on a technology curve. The liability amortizes on a financial curve. They do not track.
Yield is never free; it is rented. And the rental term here is badly mismatched.
Now check who holds the paper. If insurers and pension funds dominate the buyer base, the transmission path differs from bank-held credit. These are long-duration, rate-sensitive, slow-repricing buyers. Once a pension fund decides AI infrastructure debt is mispriced, it does not flip back on a single earnings beat. The reprice is permanent.
Then there is the SPV layer. Off-balance-sheet and quasi-off-balance-sheet structures mean the true leverage of the AI buildout is understated. I spent a week in 2022 reverse-engineering Terra's oracle failure with Python scripts. The lesson stuck: the visible balance sheet is never the whole story. The debt that breaks you is the debt you cannot see.
Check the gas, then check the truth.
The distinction that matters is not "how much debt." It is "whose debt." A balance sheet with $100 billion of free cash flow per year can carry this. A balance sheet that depends on rolling credit to fund the next GPU order cannot. The market is starting to draw that line, and the line is where the repricing begins.
Track the issuance calendar, not the press release. When a hyperscaler's next bond deal prices wider than the last one — same issuer, same duration, higher spread — that is the market marking AI credit risk up in real time. Spreads are the price of doubt. Watch them weekly, not quarterly.
Follow the money to the supply chain, because that is where debt converts to cash. The proceeds flow into GPU orders, data-center construction, power purchase agreements, cooling. Nvidia's data-center revenue is the visible beneficiary. Microsoft restarting Three Mile Island, Google and Amazon signing nuclear PPAs — those are debt-financed bets on electricity. Power is the slow variable. You can borrow in weeks. You cannot interconnect a substation in weeks.
So capital gets raised faster than it can be deployed. That is a duration mismatch inside the buildout itself, layered on top of the debt maturity mismatch.
Contrarian
The consensus read is that this is an AI bubble warning. I disagree with the framing.
The implicit message — "financial stability pressure" — is emotionally loud and quantitatively empty. The issuers are among the highest-rated credits on earth. Microsoft and Amazon carry ratings that make a 2008 subprime comparison absurd. Default probability is not the risk. Repricing is the risk.
And the alarmism ignores the asset side entirely. Cloud AI backlog — Microsoft Azure AI, Oracle's remaining performance obligations — is growing fast. Oracle's RPO has ballooned into the hundreds of billions. The liability story only lands if the asset story stalls. Nobody covering the debt is covering the backlog with the same intensity.
Backtest the assumption, not just the data. The assumption here is that AI revenue keeps compounding. Test that, not the headline debt number.
The real contrarian point: this is not a bubble signal. It is a capital-discipline signal. When funding shifts from equity — which tolerates losses — to credit — which does not — the market is imposing a governor on the arms race. Higher funding costs will slow the marginal, reckless builder before they slow the disciplined one. That is the system working.
Takeaway
Watch three numbers, not the headlines.
Credit spreads on AI-linked debt. If they widen, equity valuations follow. The CDS-to-equity transmission is real and fast.
The scissors gap: cloud AI revenue growth versus capex growth. When capex outruns revenue for consecutive quarters, the debt narrative wins.
Data-center utilization and power-on rates. Borrowed money that cannot energize is money that cannot earn.
When the tape freezes, the logic remains. The question is not whether AI gets built. It is who is holding the debt when the buildout meets its first real demand pause.