Six years. That is the number that decides whether a fifty-billion-dollar asset class survives the winter, and it appears exactly once โ a single line in a depreciation schedule, buried three clicks deep in a filing almost nobody reads past the revenue table. Six years is how long CoreWeave assumes a GPU keeps earning. Six years is the assumption on which every delayed-draw term loan, every special-purpose vehicle, every structured note in the neocloud complex quietly rests. And six years is almost certainly wrong.
I want to start there, not with the headline. Because the headline โ the one screaming that two hundred AI cloud firms face a wipeout, that debt markets have slammed shut, that the whole stack is one margin call from vaporizing โ is not new information. It is the ambient weather of the 2026 bear market. What is new, and what almost nobody has priced, is the mechanism. The trigger for the next credit event in AI infrastructure will not be a customer leaving. It will be an accountant changing a number from six to three.
That is a very different kind of risk. And it is the kind you only see if you excavate truth from the code's buried layers โ or, in this case, from the arithmetic that sits underneath the code.
Before the mechanism, the terrain.
A neocloud is a company that does one thing: it borrows money, buys GPUs, racks them, and rents the resulting compute to someone who needs it faster than they can build it themselves. CoreWeave is the archetype. It is not a cloud in the AWS sense โ it does not own the full stack, does not sell a hundred services, does not carry a decade of enterprise trust. It sells raw, fungible, interchangeable GPU-hours. That is the entire product.
The financing follows the product. CoreWeave's principal debt instruments are delayed-draw term loans โ DDTLs โ wrapped inside special-purpose vehicles. Read that structure carefully, because it tells you what the lenders actually believe. A DDTL secured against a specific GPU cluster and the specific customer contracts attached to that cluster is not a bet on the company. It is a bet on an isolated pool of metal and paper. The structure itself is an admission: the creditors do not trust the corporate balance sheet. They trust only what they can wall off. This is equipment finance wearing a software company's valuation.
Then there is the backlog. CoreWeave has publicly cited something in the neighborhood of fifty-five billion dollars of contracted future revenue, inflated by large multi-year deals with hyperscalers and model labs. Leadership has repeatedly framed this backlog as the thing that "more than supports" its debt obligations. On an accounting page, that sentence can be made true. On a cash-flow page, it is close to a lie, and I will show you the arithmetic.
Now, the source material I was handed. It arrived with a set of financial claims that do not reconcile โ a private aerospace company allegedly hosting an earnings call with a single-quarter capex guidance several times larger than its plausible annual spend; an "AI business unit" attributed to an entity that does not contain one; a debt figure roughly three times the publicly disclosed total. I am not going to launder those numbers into my analysis, because a number you cannot reconcile is not a data point, it is a rumor with a decimal. But the direction those claims point โ tightening credit, neocloud consolidation, a debt market that has quietly closed its doors โ is independently verifiable through other channels: credit default swap spreads widening at the large leveraged data-center borrowers, the high coupons CoreWeave has had to pay to place paper, the special-purpose-vehicle architecture hyperscalers adopted to keep this exposure off their own balance sheets, and the fact that prominent contrarian investors have taken public short positions on the depreciation assumption itself. The direction is real. The specifics are noise. I will work with the direction and be explicit wherever the arithmetic is mine.
Here is the machine, disassembled.
First: depreciation is the hidden leverage, and it is the only variable that can trigger default without a single customer leaving.
CoreWeave depreciates its GPU fleet on roughly a six-year straight-line schedule. The industry argument around that number is not academic. Some buyers model H100 economic life at two to three years. At least one major customer has internally used five. The gap between six and three is not a rounding error โ it is the difference between a company that prints EBITDA and a company that bleeds EBIT.
Run the sensitivity yourself. Take a fleet, any fleet. Halve the assumed life, and you double the annual depreciation charge. That charge flows straight through operating income. A business that looks comfortably cash-generative at six years can flip to a headline operating loss at three, with identical revenue, identical customers, identical contracts. No churn required. No price war required. Just a change in an accounting estimate that a new auditor, a new CFO, or a nervous board can force at any moment. Every bug is a story waiting to be decoded, and this one is written in the least glamorous line of the financial statements.
Why would anyone force it? Because the physical evidence is pointing that way. The transition from Hopper to Blackwell to the next generation compresses the useful economic window of the prior fleet. A GPU cluster is not a building with a twenty-year life. It is a depreciating option that burns value along a curve set by somebody else's roadmap. When the next generation delivers a step-change in performance-per-watt, the residual value of the old cluster does not gently decline โ it cliffs. And the financing amortization schedule does not cliff with it. It grinds down linearly, on that same six-year line, oblivious.
Second: the backlog is not cash. It is a promissory note for future capital expenditure, and that makes it a liability in disguise.
Do the division. Fifty-five billion of backlog, spread across contracts averaging four to six years, implies something like a hundred to a hundred and ten billion in annualized revenue at full delivery. Current annualized revenue is roughly in the mid-five-billion range. To climb from five to a hundred and ten, you must build. Building means three to four hundred billion of additional capital expenditure. Capital expenditure means two to three hundred billion of additional debt.
Read that again, slowly. The larger the backlog, the larger the financing gap. Backlog is being marketed as the proof of solvency, when in fact it is the quantification of the hole. A company with a small backlog needs little new money. A company with a fifty-five-billion-dollar backlog needs a fortune, and it needs it now, on a schedule dictated by delivery milestones it has already promised. This is the structural paradox at the heart of the neocloud model, and it is the reason the credit market's closure is not a temporary inconvenience but an existential event. The backlog is a treadmill that speeds up the longer you stay on it.
Third: backlog evaporates precisely when you need it most.
Most of these multi-year contracts contain capacity-delivery milestones. Miss the milestone โ because you cannot finance the build, because the transformer did not arrive, because the grid interconnection slipped eighteen months โ and the customer can walk, typically without penalty. This is not hypothetical. It is the standard contract architecture, and it means the backlog is contingent on the very financing that the backlog's existence is supposed to justify. A closed credit market does not just slow the build. It vaporizes the order book that was collateralizing the debt. The asset and the liability are coupled, and they fail together. Where are the take-or-pay clauses? What percentage of the book carries a binding minimum purchase commitment rather than a best-efforts forecast? That disclosure is the one number that would tell a lender whether the backlog is a contract or a mood, and it is the number the industry is least eager to publish.
Fourth: customer concentration turns a business risk into a counterparty risk.
In 2024, Microsoft accounted for the substantial majority โ publicly around sixty percent โ of CoreWeave's revenue. The largest customer is also the most capable competitor, a company that can build its own capacity and reclaim that demand the moment the economics favor insourcing. That is not a partnership. That is a stay of execution, renewable at the counterparty's discretion. And the offsetting contracts with model labs carry their own counterparty risk, because those labs are themselves loss-making entities dependent on external financing. You are lending against the cash flow of companies that are also borrowing to survive. Navigating the labyrinth where value flows unseen means noticing that the same dollar is being counted as revenue by three different companies in the same loop.
Fifth: the Nvidia flywheel, and what happens when it reverses.
Nvidia has historically invested in neoclouds, extended them supply priority, and benefited from their insatiable purchasing. Trace the loop: invest, enable purchase, recognize revenue, mark up the investee's valuation, support its refinancing, enable the next purchase. It is a beautiful flywheel while credit is cheap. It is a beautiful flywheel right up until the moment the refinancing step fails โ and then every subsequent rotation runs backward. The investment marks down. The purchase orders soften. The revenue guidance trims. The valuation that supported the debt collapses, which tightens the debt, which kills the purchase orders. The same architecture that amplified the up-cycle amplifies the down-cycle, and the amplification is not symmetric โ it is faster on the way down, because margin calls are instant and construction is slow.
Sixth: the value is not being destroyed, it is being redistributed โ and the middle layer is what gets flattened.
This is not a simple industry shakeout. It is a migration of value along the chain. Upstream โ the chip designers, the fabs, the memory suppliers โ benefits in the short term but loses its fastest-growing marginal buyer, because the neoclouds were among the most aggressive purchasers in the market. The physical layer โ power, land, interconnection rights, transformers, gas turbines โ becomes the new bottleneck and gains pricing power it has not held in a generation. The downstream hyperscalers and model labs strengthen their insourcing trend and simply skip the intermediary. And the middle layer, the pure GPU reseller with no power, no land, and no proprietary model, gets compressed from both ends at once. The employment story follows the same geometry: this is structural mismatch, not aggregate loss. The jobs that disappear are finance, sales, and generic operations roles. The jobs that appear are electricians, HVAC engineers, power systems specialists, and data-center construction crews. The skills do not transfer. The geography does not transfer. A laid-off capacity planner in the Bay Area is not going to relocate to a substation in Virginia to pull cable.
Seventh โ and this is the part that belongs to my own discipline โ none of this compute is verifiable.
Here is the gap that has been bothering me since I started prototyping proof layers for model inference. Every dollar of neocloud revenue rests on an assertion that cannot be independently checked: that a specific number of GPUs ran a specific workload for a specific duration at a specific utilization. The lender financing a GPU cluster has, in practice, no cryptographic or even robust statistical proof of that. They have a meter reading, an invoice, and a promise. Utilization is an unobservable variable. Uptime is self-reported. The physical asset โ the thing the whole DDTL structure is secured against โ is a machine whose economic output is invisible to the party taking the risk.
Sit with that. We built an entire global industry of collateralized lending on top of an asset class whose productivity is, from the creditor's vantage, a black box. In my own work this stopped being abstract. When I forked the Circom compiler in 2021 and built a simplified tutorial that let a few thousand developers deploy their first circuit, the lesson was not about privacy. It was about what a proof actually buys you: a succinct, cheap, independently checkable statement that a computation happened as claimed, without re-executing it. Later, working with a handful of AI startups to prototype a zero-knowledge proof layer for large-model inference, the same lesson reappeared in a commercial register. If you can prove that a declared inference workload executed on declared hardware for a declared interval, you have converted an unobservable asset into a financeable one. The missing primitive is not capital. It is proof of compute.
This is not a philosophical aside. It is the difference between a collateral pool a lender can mark and a collateral pool a lender must trust. Trust is what gets repriced first in a bear market. And the neocloud complex has an enormous amount of it outstanding.
Now the part most analysts will get wrong, and the part I think is the actual story.
The consensus framing is that this is a demand problem. AI capex is enormous, the returns are unproven, therefore the neoclouds are overbuilt. That framing is comfortable because it lets everyone debate demand โ a number nobody can measure โ while ignoring the two numbers that actually determine the outcome.
The first is the depreciation schedule, and I have already shown you why. The second is the physical layer, and almost nobody is talking about it correctly.
The real bottleneck is not GPUs. It is grid interconnection, transformer lead times, and gas turbines. You cannot deploy capital faster than you can get power. That constraint does not disappear when credit tightens โ it becomes the only thing that matters. Which means the value in this chain is not going to the company that rents out the metal. It is going to whoever holds the land, the power purchase agreement, and the queue position. The intermediary is left holding depreciating silicon financed with non-depreciating debt.
The second thing the consensus gets wrong: this is being valued as a software business and financed as an airline. Fungible output, near-zero switching costs, brutal capital intensity, no pricing power, and a product whose unit price is in structural decline. Spot rental prices for the previous generation have already compressed dramatically from their peak โ a decline of the kind you see in shipping rates, not in SaaS. When spot pricing falls below the locked-in long-term contract price, the economic value of the backlog goes negative relative to its stated value. The contracts remain legally binding. The economics underneath them do not. A business with that profile deserves an airline multiple, and the market has been paying a software multiple. That gap is the trade, and it has nothing to do with whether AI demand is real.
And the deepest blind spot: the lenders think they are secured. The DDTL structure isolates a GPU pool and its contracts, which sounds like protection until you ask what that pool is worth in a forced liquidation. What is the secondary market for a two-year-old accelerator cluster with eighteen months of remaining economic relevance, sold by a distressed seller into a market where the marginal buyer just watched the spot price halve? The recovery assumption embedded in those structures is almost certainly optimistic, and recovery assumptions are the one input in credit modeling that never gets marked until it is too late. The collateral is real. The liquidity of the collateral is a fiction nobody has tested.
Watch three things in 2026, and none of them are revenue.
Watch the depreciation note in the next set of filings. If any major neocloud shortens its assumed GPU life, treat it as a credit event, not an accounting adjustment โ because that is what it is, and it will propagate through every DDTL and every SPV that used the old number.
Watch the ratio of delivered to undelivered backlog. It is the only disclosure that separates an order book from a wish list, and it is the number the industry is least eager to publish.
Watch whether anyone ships a credible proof-of-compute primitive and gets it accepted by a lender. If that happens, the entire risk architecture of this industry changes โ collateral becomes verifiable, and verifiable collateral is cheaper than trusted collateral by an order of magnitude. Composability is not just function; it is poetry, and the poetry here is that the one component missing from the stack is the one that would have made the whole thing honest.
So here is the question I keep returning to, the one the wipeout headlines never ask. When the first auditor shortens the life of a GPU from six years to three, and a hundred billion dollars of "secure" equipment debt reprices overnight, will we finally admit that the thing we were lending against was never the hardware?
Or will we do what we always do โ build the next layer on top of the same unverified foundation, and call it innovation?