There is a particular silence that follows a number that does not add up. I first felt it in 2017, alone with MakerDAO's early governance contracts, watching a stability fee calculation drift a few decimal places away from solvency. Nobody had asked me to look. I looked anyway. The silence arrived later, when the fix shipped and no one wanted to discuss how close the edge had been. I have been chasing that silence ever since, because it is usually where the real story lives.
This week the silence came from a single sentence in a news story. The ten-year Treasury yield had risen 'about 30 percent' from its March low, the story said. It had not. A move from roughly 4.1 percent to 5.4 percent inside eight months would be a generational repricing of the American state's borrowing cost. What actually happened was thirty basis points, a rounding error in monetary history and a catastrophe for exactly one kind of balance sheet.
The unit error is not the story. The unit error is the fingerprint. Somewhere between the analyst's model and the reader's screen, a variable lost its magnitude, and nobody in the chain felt the need to catch it. That is the same failure mode I spent six months documenting in 2017: not malice, not fraud, just a system so confident in its own narrative that it stops checking whether the numbers still describe the world.
Here is what the numbers actually describe. CoreWeave's existing high-yield debt carries a coupon in the nine percent range. Its bonds now trade at a secondary yield close to fifteen percent. That gap, roughly six hundred basis points, is the entire article. Six hundred basis points is not what a thirty-basis-point move in the risk-free rate does to a credit. It is what a market does when it stops pricing a rate cycle and starts pricing a probability of default.
On the equity screen the same week, Oracle fell 8.8 percent in a single session and CoreWeave fell 15 percent. If this were purely a rate story, the investment-grade issuers, Oracle and Meta and Alphabet and Microsoft among them, would have been hit hardest in duration terms, because their bonds are the ones most sensitive to the benchmark. They were not. The dispersion in the equity tape maps almost perfectly onto leverage and customer concentration, not onto rate sensitivity. Chop is for positioning, and the credit market had already positioned while the equity crowd was still arguing about the tape.
Let me be precise about what a neocloud is, because the word hides the structure. These are companies that buy GPUs at scale, finance them with debt, and rent them out under multi-year take-or-pay contracts. The revenue side is fixed at signing. The cost side floats with credit markets. The asset side depreciates on an accounting schedule that may or may not resemble reality. Written out that way, the business is not a cloud business at all. It is a leveraged lease on a depreciating industrial asset, wrapped in the language of software margins.
The timeline matters. CoreWeave listed in March 2025. Oracle issued a large slug of debt in September 2025 to fund data center buildout and watched its equity slide for weeks afterward. In November 2025, Gartner published a forecast that roughly 80 percent of neoclouds would lose their independence by 2030. Three data points, one direction. The equity story and the credit story were telling different things about the same assets, and only one of them was being marked every day.
The sell-side framing that dominated coverage came from Citi's co-head of US software research, who argued the neocloud drawdown was a rates problem rather than a demand problem. The recommendation that followed was Microsoft, on the grounds that it generates free cash flow, does not depend on the bond market, and therefore spares the analyst from having to judge how fast chips depreciate. I want to be fair to this argument, because parts of it are correct. Rates do hit unrated borrowers hardest. But the argument quietly swaps one risk for another and then calls the swap safety.
Why should a person who spends her days auditing governance contracts and DeFi composability care about the cost of a GPU lease? Because compute just became a financial asset, and financial assets behave according to the people who finance them rather than the people who use them. That is the same transition crypto lived through in 2020, when yield farming turned protocol tokens into collateral and the collateral became the product. The moment an asset becomes borrowable, its price stops being a technical question and becomes a governance question. The neocloud sector is living through that transition in public, and most of its participants have not noticed.
Decompose the six hundred basis points and the story writes itself. The risk-free component moved thirty basis points. The remainder is issuer-specific: customer concentration in a handful of hyperscalers, a depreciation schedule that assumes five to six years of economic life for silicon that may be obsolete in three, and a duration mismatch between short, expensive debt and long, locked contracts. None of that is rate risk. All of it is credit risk, and credit risk is a statement about the borrower's ability to survive a bad quarter, not about the Federal Reserve.
Test the rate thesis against the investment-grade cohort and it fails cleanly. Oracle's debt is investment grade and its equity fell 8.8 percent. CoreWeave's debt is high yield and its equity fell 15 percent. If the benchmark were the driver, the ordering should be closer to the opposite. What the market actually priced was the difference between a company that can refinance from operating cash flow and a company that cannot. The spread is the message: this is a solvency repricing wearing a rate story as a coat.
The duration mismatch deserves its own paragraph, because it is the least discussed and most mechanical failure mode. Take-or-pay contracts lock revenue for years, which looks like stability until you notice that the debt funding those contracts is shorter and reprices faster than the contracts do. In a rising-credit-cost environment, the asset side is fixed and the liability side floats. That is not a hedge. That is a short volatility position on your own funding costs, held by a company whose only collateral is depreciating hardware.
Which brings me to the claim that costs can be passed through to future contracts. On the margin, yes. In aggregate, this is the most overrated buffer in the sector. Existing contracts were priced before the credit market moved, so the first thing a rate increase compresses is the margin on capacity already sold. Raising prices on new contracts pushes customers back toward building their own capacity or toward a cheaper competitor, which caps how much of the increase can actually be passed. Transferability is not a cushion; it is a negotiation, and the counterparty holds most of the cards.
I also want to flag a factual problem in the demand argument, because it sits at the center of the whole thesis. The claim that H100 rental rates are 'near historic highs' does not survive contact with public GPU rental indices, which show spot rents down by more than half from their 2024 peak and sitting in a low band through 2025. The claim can only be true if it refers to long-term contract pricing, to Blackwell-class capacity, or to a recent bounce off the floor. None of those distinctions were made. When the single strongest piece of evidence for 'demand is healthy' cannot be verified as stated, the conclusion built on it needs to be rebuilt rather than repeated.
Then there is the backlog. The company's public response to questions about its debt was that contracted backlog covers it. Backlog is an undiscounted nominal figure that requires continued capital expenditure to convert into cash, assumes customers neither default nor reduce volume, and ignores the time value of everything in between. Covering nominal debt with nominal revenue is the kind of arithmetic that works on a slide and fails in a workout. I have read fifty post-mortems from the 2022 cycle, and the phrase 'our backlog covers it' appears in a startling number of them.
Now the assumption nobody wants to mark: depreciation. The industry broadly assumes five to six years of useful life for GPUs. If model architectures keep iterating at the current pace, the economic life may be materially shorter. A downward revision would hit two places at once, the solvency of every leveraged neocloud and the reported profit of every hyperscaler carrying the same assets. That is a cross-sector accounting risk disguised as an industry-standard convention, and buying Microsoft specifically because it avoids the depreciation question is a transfer of that risk into the footnote, not its removal.
And then there is the channel the article never mentions, which is where I think the real exposure sits. Alongside public bonds, 2025 saw rapid growth in private credit, GPU-backed lending, and data center asset-backed securities. These instruments do not print a price every day. Their valuations lag, their redemption terms are gated, and their covenants often trigger on collateral revaluation. If GPU rental rates fall while loans are marked against historical purchase prices, collateral and loan value diverge, and the trigger fires quietly. The risk that is hardest to see is not the risk that is smallest; it is the risk that is repriced last.
I recognize this shape. In 2020 I spent four months in a cabin outside Seattle mapping composability risk across a single yield aggregator, and what I found was not a flaw in any one vault but a topology: every vault assumed the others were solvent, and each assumption was individually reasonable. The neocloud sector has the same topology with different nouns. Every lender assumes the GPU collateral holds, every customer assumes delivery arrives, every analyst assumes the others have done the depreciation work. In the chaos of DeFi, I found my silence, and the silence was telling me that contagion is a property of the map, not of any single node on it.
There is also a harder constraint than silicon, and it is electricity. Data center delivery slips are more often caused by interconnection queues and local permitting than by GPU supply. That matters here because take-or-pay contracts carry delivery obligations, and a missed delivery date is a non-technical default. The backlog that supposedly covers the debt assumes capacity arrives on schedule. Some of that schedule is not in the company's hands, which means part of the collateral story is a regulatory timeline dressed as an engineering one.
Now the part of the story that gets flattened into a headline. Microsoft is CoreWeave's largest customer, historically above sixty percent of revenue, with contracts that have scaled into the ten-billion-dollar range. So when neocloud financing tightens, Microsoft gains twice: it gains bargaining power over a supplier that needs capital, and it gains an option to acquire capacity and contracts at a discount if valuations break. It also loses, quietly, because it is the counterparty on those purchase commitments. The recommendation to buy Microsoft is not wrong. It is incomplete in a way that flatters the recommendation.
Some of that incompleteness is structural rather than dishonest. Microsoft sits inside the analyst's coverage universe; neoclouds do not. A framework built to rank software companies will naturally find the software company safest. I do not read this as cynicism. I read it as the ordinary gravity of research desks, and it is exactly the kind of gravity readers should price in when they encounter a cross-asset conclusion drawn from a single-sector model.
Here is my contrarian read. The 'Microsoft wins' framing is not false, but it is a snapshot of a transition rather than its destination. If neoclouds slow their expansion, Microsoft's third-party capacity gets more expensive and less certain at precisely the moment its own buildout is straining. The safe harbor is safe in part because its suppliers are fragile, and fragility is not a stable foundation for a supply chain. A moat made of your suppliers' distress is a moat that eventually surrounds you.
The supporting examples deserve scrutiny too. Palantir was cited as evidence that AI software has good cash flow, which may be true, but a company trading at an extreme multiple is not a defensive asset in any conventional sense. And on the other side of the trade, Tom Lee publicly recommended buying neoclouds and chip stocks after the selloff. Two well-known strategists, same week, opposite conclusions. That disagreement is itself the most honest data point in the whole episode: the market has not decided whether this is a liquidity shock or a fundamental turn, and anyone who tells you otherwise is selling certainty.
The downstream consequence is the one I keep waiting for someone to model. If compute suppliers cannot finance expansion, the near-term effect is higher inference costs and a slower decline in the price of intelligence, which compresses margins at the application layer. Over a longer horizon, if consolidation clears excess capacity and the hyperscalers capture the scale, inference costs could fall faster than anyone currently projects. Both are plausible, and they point in opposite directions for every builder downstream. The financing structure, not the model architecture, decides which one arrives first.
There is a governance lesson buried here that my readers will recognize. When a neocloud needs to refinance, the decision is not made by its users, its engineers, or its community. It is made at a table with perhaps five seats, occupied by creditors and one or two customers. I have spent years documenting on-chain governance where turnout sits below five percent and the real decisions happen in a group chat. The neocloud refinancing table is the same structure with better stationery. Decentralization without accountability is anarchy; centralization without disclosure is something worse, because it looks like order.
And the loop is self-reinforcing in a way that no single actor controls. Debt issuance rises, spreads widen, financing costs climb, hurdle rates for projects rise, some projects slip, revenue expectations get cut, spreads widen again. Each participant behaves rationally and the aggregate is a slow tightening. That is what a debt glut looks like from the inside: not a crash, a ratchet, and a ratchet that no committee ever voted to turn.
Which brings me back to the thing I actually believe, and the reason I keep auditing systems nobody asked me to audit. Truth emerges when the ledger is transparent. Had these obligations been recorded on a public ledger with real-time marks and verifiable reserve attestations, the six-hundred-basis-point gap would have been visible to every counterparty simultaneously, and the repricing would have been faster, shallower, and less concentrated among the people who found out last. Openness is not a feature; it is a philosophy, and its absence is a cost paid by whoever holds the bag when the marks finally land.
This is why I still find the DePIN experiments interesting despite their obvious immaturity. A tokenized compute market that prices capacity continuously, on-chain, with collateral rules anyone can read, would have surfaced this exact stress months earlier than a private credit book marked at cost. It would also fail in public, loudly and quickly, which is a feature. We minted souls, not just tokens, and a soul that cannot be audited is only a brand.
Let me state what would change my mind, because an argument that cannot be falsified is a mood rather than a thesis. If CoreWeave's new contract pricing rises faster than its cost of debt through the next two quarters, the transferability thesis holds and I am wrong. If GPU collateral gets marked down in a disclosed filing, the private credit channel is real and already firing. If the hyperscaler capacity release schedule slips past 2027, the arbitrage window stays open longer than I expect. Three observable events, none of them requiring me to trust anyone's narrative.
So watch the credit tape, not the equity tape, over the next four to eight quarters. Watch the refinancing calendar against the capacity release schedule, because the overlap of those two dates determines whether this ends in bankruptcy or acquisition. Watch whether GPU collateral gets revalued, and whether anyone discloses it. Humanity remains the only non-fungible asset, and the systems we build around scarce compute will be judged by whether they protect the people downstream of the leverage or only the people who arranged it.
The fork is coming. Join it if you must, but keep the lineage, and keep the receipts.