Over the past seven sessions, three tickers the market insists on treating as a single idea โ Nvidia, Broadcom, CoreWeave โ have drifted lower in near-lockstep, while the ten-year Treasury yield pushed higher. The financial press has settled on a tidy explanation: AI debt concerns. I want to resist the tidiness. What actually happened is that a credit market cleared its throat, and an equity market heard it. The mechanism is not demand. The mechanism is the price of patience.
My eye is on the horizon, not the hourly candle, so let me say plainly what the headline buried. The story arrived with four facts and zero numbers: no debt-to-EBITDA, no interest coverage ratio, no maturity ladder, no named creditor. In my line of work, an absence of numbers is not an absence of meaning. It is a signal about who is doing the talking. When a report names three companies with radically different balance sheets and calls them one trade, it is not describing a market. It is describing a mood.
For three years the buildout of AI infrastructure was financed by the least fragile instrument available to a corporation: its own cash flow. Hyperscaler capital expenditure, on the order of two hundred billion dollars annualized, was funded largely from operating cash and topped up by equity markets willing to pay a generous multiple for the privilege. That regime had one great virtue โ nobody had to be repaid on a schedule. Growth could be wrong for a while and still survive.
That regime is ending at the margin, and margins are where regimes turn. The new instruments are GPU-backed loans, private credit facilities, and asset-backed structures that securitize the rental stream of a depreciating machine. CoreWeave is the archetype: borrow against accelerators, rent them out, service the debt from the rent. Strip the branding away and it is a levered equipment finance business wearing a software multiple.
The scale is the part that unsettles me. AI-adjacent infrastructure debt has migrated into private credit funds, insurance balance sheets, and bespoke securitizations โ vehicles valued quarterly, by the same people who originated them. Public bond markets price doubt continuously. Private marks price it annually, if at all. A risk that cannot be observed cannot be transmitted โ until the moment it is.
I have seen this film and I remember how it ended. Between 1997 and 2001, competitive carriers borrowed against fiber-optic capacity assumed to be perpetually scarce. Equipment vendors โ Lucent and Nortel chief among them โ financed their own customers, booking revenue on gear paid for with money the vendor had effectively lent. When capacity arrived faster than demand, lease rates collapsed, residual values evaporated, and the loans went bad. The technology was never the problem. The fiber was real, the internet was real, and the companies that built it went bankrupt anyway. Vendor financing is not a new sin. It is an old one with better branding.
The three companies being sold as one thing are not one thing. Nvidia designs chips, holds something like eighty to ninety percent of the accelerator market, and defends it with a software ecosystem that competitors have spent a decade failing to replicate. Broadcom sells custom silicon to the hyperscalers plus the networking fabric that stitches clusters together โ it is, structurally, a beneficiary of every customer who decides to reduce dependence on Nvidia. CoreWeave owns depreciating metal, rents it to a small number of very large tenants, and differentiates itself mainly by having borrowed faster than its peers.
That difference is the category error at the heart of the selloff. Nvidia and Broadcom are being repriced for CoreWeave's capital structure. Correlation is not causation; often it is just a shared ticker tape.
The most consequential mechanism is a clock. A traditional data center asset is depreciated over ten years or more. The economic life of an AI accelerator is closer to three to five, because the next generation does not merely improve on the last โ it redefines what a unit of compute is worth. If a facility is financed with five- to seven-year paper against an asset that loses competitive value in four, the loan outlives the collateral's usefulness. Collateral coverage then becomes a function of residual value, and residual value becomes a function of a benchmark nobody has published yet.
There is also the matter of the denominator. High-multiple equities are long-duration instruments, and the AI complex carries the longest duration in the market because so much of its value sits in terminal assumptions. A move in the risk-free rate repriced the whole complex before a single lender tightened a single covenant. That is why Nvidia fell alongside a company with a fraction of its balance sheet strength. Same denominator. Very different numerator.
The rate backdrop deserves its own breath, because it is not merely a Fed story. Long-end yields have been pushed up by term premium โ heavy sovereign issuance meeting indifferent foreign demand โ which is a slower, stickier force than a policy rate. Sticky rates are precisely the wrong environment for assets whose value is concentrated in distant cash flows. A central bank can cut. A term premium takes years to unwind.
The mechanism I distrust most is marks. Private credit does not mark to market; it marks to appraisal. During 2021 I spent eight months modeling the sustainability of yield-farming protocols and found that most headline APYs were a function of inflow velocity rather than value creation โ the yield existed because new money kept arriving. I wrote an internal memo about it, was politely ignored, and watched the shape repeat itself across three subsequent cycles. The same shape is visible here. Debt service coverage in a leveraged compute lessor is a function of refinancing velocity, not of the utility of the computation being sold. When the window is open, everything looks solvent.
And there is revenue quality. When a chip vendor invests in a compute lessor, and the lessor uses the proceeds to buy that vendor's chips, a portion of reported revenue is a closed loop rather than a sale to an end user. No one publishes the percentage. That percentage is the single most important undisclosed number in this sector.
Here is the insight I would ask you to keep after the rest of this essay fades: utilization is not solvency. A cluster can run at ninety-five percent occupancy and still default, because what matters is whether the rent it earns clears the debt service on the price that was paid for it. Full utilization tells you the machine is busy. It tells you nothing about whether the entity that owns the machine will exist in three years. The industry's favorite metric is a measure of activity, and activity has never once paid a coupon.
The consensus reading of this week is that the AI bubble is deflating. I think that frame is lazy. Bubbles are arguments about valuation; this is an argument about term structure. Valuations can stay irrational for years if the refinancing window stays open. Term structures resolve on a calendar, quietly, in a room with a lender and a lawyer, and they resolve whether or not the technology is real. The bust, when it reaches the leveraged layer, was not an end, but a necessary pruning โ and the pruning will not touch the demand curve. It will touch the balance sheets that speculated on it.
Which brings me to the part of this story almost nobody is writing. The neocloud sector now contains dozens of operators chasing the same handful of anchor tenants, and the number of operators is growing faster than the number of tenants. That is not capacity creation. It is leverage arbitrage on a shared demand base โ a bet that someone else's future contract will service your present debt. The instinct that made me skeptical of a dozen Layer 2s carving up the same scarce liquidity applies here with considerably more force, because in this case the slicing is done with borrowed money, and the collateral depreciates while you wait for the demand to arrive.
There is a second-order effect the bearish framing misses entirely. If debt-funded third-party compute becomes expensive, the hyperscalers do not stop building. They build themselves, on their own balance sheets, with their own silicon. Risk in this industry does not vanish when credit tightens; it redistributes โ away from the leveraged intermediary and toward the vertically integrated incumbent. The companies that emerge stronger from a compute credit cycle are the ones that never needed the loan.
I should name the framing problem directly, because it shapes how readers absorb all of this. The report that started the discussion came from a crypto-native outlet, and crypto-native outlets carry a long-standing rhetorical investment in the comparison between AI exuberance and their own asset class. That does not make the debt concern false. It makes it selectively told โ the bear case, without the countervailing fact that the largest buyers of compute fund their capex from operating cash flow, not from the bond market. Read the mechanism. Discount the mood.
This is a sideways market, and sideways markets are where positioning is done, not where convictions are declared. In the fund I run, consolidation phases are for building the watchlist and the covenant checklist, not for expressing a directional view on a narrative. The instruments I want ready are unglamorous: the maturity schedules of every listed compute lessor, the customer concentration disclosures buried in the risk factors, and the rental rate series that nobody has bothered to index. Chop is for positioning. The signal, when it comes, will arrive as a refinancing headline at seven in the morning, and I would rather have done the work already.
So what do I actually watch from here? None of it is a stock price. Watch published GPU rental rates โ the spot market for compute is the closest thing this sector has to a real-time credit spread. Watch secondary residual values on previous-generation cards, because that is where an accounting assumption meets an auction. And watch for the first covenant breach or failed refinancing at a leveraged lessor, which will tell us whether this is a repricing or a regime.
The question is not whether artificial intelligence is real. It plainly is. The question is narrower and harder: when the collateral ages faster than the loan, who is holding the clock?