Over the past seven days the loudest number in my feeds was not a price. It was a sentence โ a single line from an executive at a firm managing more than $1 trillion in alternative assets, telling a reporter that the market's anxiety about AI debt was really a question of distinguishing risks rather than avoiding them. The line travelled fast because it was comforting. I read it three times and felt what I felt in 2022, reading a now-defunct lender's blog post about "risk-adjusted collateral." Not because the sentence was wrong. Because it was smooth. Smooth language is what capital sounds like when it is managing a price it does not want to move.
What stayed with me was the shape of the report around it. No exposure figures. No instrument types. No maturities. No denominator. A reassurance with no denominator is not analysis; it is weather. And in my experience, the absence of numbers is never neutral โ it is where the leverage lives.
Set the facts plainly, because the story cannot be read without them. AI infrastructure has become a debt-financed industry. The four largest hyperscalers committed well over $200 billion of combined capital expenditure in 2024, and a meaningful share of that ran through debt, equipment leases, and long-dated power purchase agreements rather than retained cash. A single modern AI data center clears $1 billion before the first rack is energized. A serious GPU cluster runs into the hundreds of millions. These are not venture cheques. They are balance-sheet obligations with covenants attached.
The lender of record is no longer only the bank. Private credit โ a market that has grown past $1.5 trillion โ now sits at the centre of this financing. Blackstone, Apollo, Ares and their peers arrange and hold the paper. These are assets without public prints, valued by model rather than by trade, and that single structural fact does more to explain the current mood than any interest-rate chart.
This is why one executive sentence carries weight, and why it deserves a second reading. The person describing the risk is also the person holding it, and often the person who arranged it. That does not make the statement false. It makes it situated. I learned this in 2017, when I spent a summer reading fifteen ICO whitepapers for their philosophy and wrote twenty pages arguing that most of them had no community to speak of. The founders were not lying. They were describing the asset they owned.
And it is why this reaches my desk at all. Crypto has spent four years building an on-chain credit market โ tokenized treasuries, private-credit wrappers, GPU financing, compute-indexed yield. The AI debt question is no longer a TradFi question with a crypto footnote appended for SEO. It is the same question asked twice, in two languages, by two markets that do not share a marking convention. That gap is the most interesting thing in the industry right now, and almost nobody is trading it.
Here is the technical core, and it is not about totals. The risk in AI infrastructure debt is not the size of the stack; it is the shape of it. Follow the durations. Data center leases and power contracts run ten to twenty years, often as take-or-pay. The GPU inside the building has an economic life of roughly three to five years before an architecture announcement resets its residual value. The revenue that services both is underwritten on application-layer growth measured in quarters. Three clocks, three tempos, one amortization schedule. Any honest credit model has to reconcile all three, and almost none of them do, because the model is built by the party who benefits from the reconciliation looking tidy.
Then there is the collateral itself, which is where my attention stays. GPU-backed lending treats a chip as an asset in the ordinary sense โ something with a market, a residual, a liquidation path. But a GPU is not a building. It does not decay gracefully. It depreciates by announcement: a new generation lands and the prior one's secondary price steps down in weeks, not years. The mark on the loan does not move, because the mark is modelled, and the model was written by someone who needed the mark not to move. Every broken token taught me how to hold value; every broken chip teaches the same lesson more slowly and far more expensively.
I spent 300 hours auditing Uniswap V2's contracts during DeFi Summer โ not hunting exploits, but reading the fair-launch philosophy inside the math. What I carried out of that audit was a single conviction: lending markets are not honest because they are decentralized. They are honest because they are observed. Price feeds, loan-to-value ratios, liquidation thresholds, public bad-debt balances. Their honesty is a byproduct of visibility, not ideology. Strip the visibility and the same design becomes a very sophisticated way to be wrong quietly.
Which is why 2022 keeps returning to me. Celsius collapsed holding "uncorrelated collateral" that turned out to be the same asset wearing different names. Three Arrows borrowed against positions that were, in the end, one position. Rehypothecation turned five balance sheets into one and everyone called it diversification. The 2025 version is subtler and better lawyered: a data center leased to a neocloud, financed by private credit, whose anchor tenant is a startup funded by the same cycle that financed the building. Correlation wearing a lease. Legally distinct. Economically singular.
The differentiation narrative, then, is not wrong โ it is coarse. Investment-grade hyperscaler paper and a neocloud facility are genuinely different instruments, and pricing them alike would be an error. But the sentence as delivered to the public collapses them into one phrase, "AI debt," and that is precisely how a differentiated risk becomes an undifferentiated narrative at exactly the wrong moment.
I keep returning to an older analogy from my own industry. Liquidity mining APY was never yield; it was a project subsidizing its own TVL number, and when the emissions stopped, the users left with them. The same geometry appears here under different vocabulary: anchor tenants signed at subsidized rents, capex underwritten on occupancy that exists only for as long as the incentive does. The lease is the emission schedule. Nobody calls it farming because the tenants have logos.
And there is a second parallel I have been chewing on since the Layer2 buildout. We spent years watching rollups pay for dedicated data availability that their blocks would never fill โ capacity provisioned for demand that had not been negotiated, priced as if it had. Dedicated data centers for tenants who will never fill racks are the same essay with different footnotes. Overbuilt capacity is not fraud. It is a fixed cost waiting for a demand curve, and fixed costs are patient right up until the refinancing date.
Here is the piece I think most readers are missing, and it is the information gain I would stake this essay on. The honest price of AI credit is now discoverable, but not where the headlines look. Where AI compute exposure enters a public or semi-public wrapper โ tokenized GPU financing, compute-indexed yield products, on-chain private-credit tranches โ the marking is continuous, the collateral is disclosed (unevenly), and the liquidation mechanics are visible. The spread between the private mark and the wrapper's implied mark is the single most informative number in this entire story. It is where a modelled valuation meets a traded one. My code was the covenant, not just the contract, and covenants that fail on-chain fail in public โ which is the only reason anyone ever learns from them.
Practically, that means watching four things rather than one: spread widening on AI-linked private credit, secondary GPU pricing and rental rates as the collateral's true print, the ratio of PPA duration to offtake contract term, and whether managers' public language matches their quarterly disclosures. Words are cheap. Filings are not.
The consensus reading is that AI debt is a bubble waiting for a trigger. Mine is quieter. The deflation has already begun โ at the collateral layer, where nobody is watching โ and what we heard was not a forecast but a sequencing. Risk differentiation, done properly, protects the head and clears the tail, and the people saying it usually hold the head. That is not cynicism. That is a portfolio, described out loud.
So apply the pragmatism test. Take that same sentence and replay it with AI-linked credit spreads 150 basis points wider and one visible neocloud default. Does "we are comfortable with the risk, provided it is properly differentiated" still read as analysis โ or does it read as please do not reprice my book? A framework that only survives calm weather is not a framework. It is a mood with a compliance department.
And the blind spot in my own argument deserves the same blade. If AI capacity is genuinely overbuilt, the correction may not arrive as a credit event at all. It may arrive as a price event โ cheap compute, cheap leases, distressed capacity sold to whoever can run it at a lower cost of capital. That outcome is not bad for everyone downstream; it is arguably excellent for the decentralized compute market that spent two years being told it was too small to matter. In the silence of the bear, we heard the truth: the bear was never the enemy of the builders. It was the only thing that made their costs legible.
So I am not waiting for the bubble to pop. I am waiting for the marks to move โ for the modelled number and the traded number to be forced into the same room. The question worth holding through this sideways chop is not whether AI debt is sustainable. It is who gets to decide when the number changes: the party who owns the asset, or the market that finally gets to see it. Ask that of every AI-linked yield product you are offered this quarter, and see which side of the sentence you are standing on.