Over the past seven quarters, four companies rewired how the AI buildout gets paid for. Microsoft, Alphabet, Amazon and Meta lifted combined annual capital expenditure from roughly $150 billion in 2023 toward a $300 billion-plus run-rate by 2025. Meta priced a $30 billion bond in October 2024 — the largest corporate debt sale of that year, and it carried a 40-year tranche. Oracle followed with roughly $18 billion in early 2025, tied explicitly to data center and cloud capacity.
The most important number in AI is no longer a benchmark score. It is a credit spread.
Bond investors are not equity tourists. They buy downside protection first and upside second. When they start asking whether AI capex can service its own interest, the question stops being about model capability. It becomes about cash conversion. The equity market stopped asking that question somewhere in 2023.
The shift is structural, not cosmetic: the AI arms race has moved from the income statement to the balance sheet.
For three years the story was simple. Hyperscalers generated enormous free cash flow, spent a slice of it on GPUs, and wrote the rest off as a growth investment. No leverage. No maturity walls. No creditors. The buildout itself was real — Microsoft restarted Three Mile Island, Google and Amazon signed nuclear power purchase agreements, and Nvidia's data center revenue became the cleanest single expression of the capex wave.
That era is over. The capex curve outran the cash flow curve. Microsoft, Alphabet, Amazon and Meta still print money, but the gap between what they spend on compute and what they earn directly from AI services has widened. Filling that gap now requires three things beyond operating cash flow: direct bond issuance, structured vehicles, and private credit.
Direct issuance is the visible layer. Meta's $30 billion deal and Oracle's $18 billion deal are the two cleanest public data points. They confirm that "borrowing hundreds of billions" is not hyperbole. It is a description of a financing channel that did not exist for AI at this scale in 2021.
Structured vehicles are the opaque layer. Meta's $27 billion data center joint venture with Blue Owl is the template: an SPV holds the physical asset, a private capital partner holds equity, and the hyperscaler signs a long-term lease. Sale-leaseback, joint ventures and asset-backed structures move debt off the headline balance sheet while leaving the obligation fully intact. This is where the real leverage lives, and it is where disclosure gets thin.
The buyers matter as much as the borrowers. If the marginal purchaser of AI-linked debt is a pension fund or insurer chasing duration, the risk transmits through mark-to-market losses rather than a banking channel. If it is a leveraged private credit fund, the transmission is faster and uglier. "Pressure on financial stability" is a meaningful phrase only after you know who is holding the paper.
Here is the uncomfortable part. The 2008 comparison is lazy — but the 2008 mechanism is not. The mechanism was never bad borrowers. It was leverage hidden inside structures that looked safe until the underlying cash flow stopped.
My Celsius framework applies almost directly. In mid-2022 I built a standardized audit that compared on-chain reserves against reported liabilities, line by line. The headline numbers looked fine. The reconciliation did not. A 15% discrepancy in Bitcoin reserves — visible only when you matched the two sides against each other — was the tell. I published a bullet-pointed note titled "Celsius is Insolvent" and predicted bankruptcy within 72 hours. The prediction held. The lesson did not expire: the risk is never in the number you are shown. It is in the number you have to derive.
Apply that lens to AI infrastructure today. The debt is not secured against AI revenue. It is secured against physical assets — GPUs, buildings, power contracts — and against the credit of a tenant. Two mismatches jump out immediately.
Mismatch one: asset life versus debt life. A GPU cluster depreciates on a three-to-five-year curve. Data center bonds and leases run twenty to forty years. Meta's 40-year tranche is being serviced by an asset class whose useful economic life is a fraction of that. If model architectures shift and today's accelerator generation becomes yesterday's silicon, the collateral marks down long before the debt matures. That is not speculation. It is arithmetic.
Mismatch two: capex velocity versus power velocity. Capital can be raised in a week. A grid interconnection queue takes years. Money is no longer the bottleneck. Electricity is. Debt raised faster than it can be electrified sits idle, accruing interest, converting nothing.
Then there is the circular financing layer. Chip vendors invest in model labs. Model labs commit to cloud capacity. Cloud providers buy chips. Each leg looks like demand. Aggregated, some of it is a mirror. When capital flows in a loop, a single node failing does not subtract from the system — it divides it.
None of this means the hyperscalers are insolvent. They are not. Microsoft, Alphabet and Amazon carry the highest credit ratings in the corporate world. Their default probability is negligible, and the 2008 subprime analogy collapses on that fact alone. The issuers are not the risk. The layer beneath them is.
That layer is private credit and the SPV ecosystem — less transparent, less rated, more sensitive to demand. If AI utilization disappoints, the sequence is predictable: data center vacancy rises, rents soften, SPV cash flows compress, and the private credit funds that financed them feel it first. The hyperscaler parent absorbs a write-down. The SPV lender absorbs a default. Those are very different outcomes, and the market tends to price them as one.
The on-chain version of this structure already exists, and it failed once. Celsius, BlockFi and Voyager were private credit funds wearing a consumer app. They borrowed short, lent long, and marked illiquid positions as safe. The AI SPV layer is the same architecture at larger scale, with better collateral and worse transparency. DeFi lending markets are starting to price the echo: tokenized treasury and private credit vaults now compete for the same duration-hungry capital that AI infrastructure wants. When two markets chase the same dollar, the spread is the message.
I have seen this pattern before, in a smaller venue. In early 2021 I built a scraper tracking BAYC sales across OpenSea and Blur, watching volume and floor together. A single whale wallet was manufacturing volume that looked like organic demand. The chart separated the two cleanly — wash volume has a signature, and it does not behave like conviction. The algorithm priced the ape before the crowd did. Twelve hours later the floor dropped 30%. The subscribers who had the chart got out. The ones who had the narrative did not.
The AI trade has the same tell. Volume — capex, bond issuance, GPU orders — is loud. Demand — realized AI revenue, RPO conversion, utilization — is quiet. When the two diverge, the loud one is not the signal. It is the noise.
Which brings us to the point the bearish framing misses entirely. Every "AI debt is a bubble" take I have read shares one flaw: it reports the liability side and ignores the asset side. That is not analysis. That is half a balance sheet.
The asset side exists. Cloud providers are carrying record backlogs. Microsoft's Azure AI and Oracle's remaining performance obligations — contracted future revenue — have grown at rates that would have been unthinkable three years ago. Oracle's RPO has become a genuine justification for its debt load rather than a rationalization. Contracted backlog is not revenue. But it is a claim on future revenue.
Value is a consensus, not a contract. A backlog is a contract. Its conversion into cash is a consensus — between the buyer's willingness to deploy and the operator's ability to deliver power. That consensus is currently intact. It is not guaranteed.
So read the bond market correctly. Cautious bond investors are not announcing that AI is over. They are announcing that capital discipline is returning. When financing gets more expensive, marginal projects die first. That is not a crash. That is a filter.
The real story is that the AI race has changed its scoring system. For three years it rewarded imagination. Now it rewards cash flow, credit rating, and the ability to fund a twenty-year asset without a twenty-year liability. That favors Microsoft, Alphabet and Amazon — cash-rich, low-leverage, top-rated. It penalizes the debt-driven challengers. Oracle is the clearest example: its expansion is the most credit-financed, and therefore the most exposed if spreads widen. Structure is not a cage; it is a launchpad — but only for the players who can still borrow cheaply.
One more mechanism deserves a name, because it is self-reinforcing. Media warns about AI debt. Bond investors read the warning and demand a wider spread. The wider spread raises the cost of capex. The higher cost slows expansion. The slowdown appears to confirm the warning. The narrative becomes the risk. I watched that exact loop run on Celsius in 2022 — the coverage did not so much predict the collapse as accelerate it.
Watch the spread, not the headline. Three signals matter over the next two quarters. First, the direction of AI-linked corporate bond spreads — mild widening is repricing, violent widening is a regime change. Second, the scissors gap between capex growth and RPO growth at the hyperscalers; if capex keeps climbing while backlog flattens, financing is outrunning demand. Third, data center vacancy and power interconnection timelines — the physical constraints debt cannot buy its way past.
Liquidity didn't vanish. It migrated — from the equity market, which prices hope, to the credit market, which prices survival. In a bear market that migration is not a warning sign. It is the only thing standing between the AI buildout and a much harder landing.
The question is not whether AI capex gets funded. It is who is still holding the paper when the music stops — and whether they checked the reconciliation before they bought.