Tracing the ghost in the machine usually means following a number that refuses to reconcile. In a report circulating this week, there are two. The first claims SpaceX posted quarterly AI-unit earnings of $7.6 billion, exceeding its connectivity business. SpaceX is private; it has never held a public earnings call, and it has no AI unit — xAI merged with X Corp, not with the rocket company. The second claims CoreWeave carries $50 billion in debt obligations, inside an essay whose entire thesis is that the debt market has slammed shut. Both statements cannot be true at once. When they appear together, the interesting question is no longer whether CoreWeave survives. It is who is holding the paper while the footnote moves.
Neoclouds are a young category: firms that rent GPU clusters rather than build clouds. CoreWeave grew out of a crypto-mining operation into the largest of them, financing thousands of H100s through delayed-draw term loans and special-purpose vehicles, each collateralized by a specific cluster and the customer contract attached to it. By late 2025 the structure had become the market's favorite Rorschach test. Oracle credit-default swaps widened. Meta pushed compute commitments through off-balance-sheet vehicles. A well-known short seller took a position. CoreWeave issued bonds above 9 percent. The narrative was already cracking; the article simply arrived with the wrong numbers and the right direction.
Here is what the report never mentions, and it is the only variable that decides solvency: depreciation. CoreWeave depreciates GPU clusters over roughly six years on a straight-line basis, while much of the buy side believes the economic life of an H100 is two to three years. Shift the assumption from six years to three and current-period depreciation doubles. EBITDA becomes an EBIT loss. No customer has to leave, no contract has to break — the company crosses into technical default purely by changing a footnote. I learned this shape of failure while auditing Uniswap's early contracts in Buenos Aires: the mechanism was sound, but the incentives underneath it decided everything. Depreciation is an incentive written as an accounting convention.
The harder falsification signal is the rental curve. H100 spot pricing fell from roughly $8 per GPU-hour in 2024 to $2–3 across multiple sources in 2025. The debt repayment models assume long-term contracts locked at $2.5–3.5 per hour for five years. Once spot crosses below the contracted rate, backlog begins to carry negative economic value — the customer is paying above market for capacity it could re-procure. That is the moment renegotiation clauses get pulled out.
Which brings us to backlog, the number the CEO cites as proof of solvency. Backlog is not cash. It is a promissory note for future capex. CoreWeave's disclosed backlog sits near $55.6 billion after the Meta contract, with an average term of four to six years — implying roughly $11 billion in annualized revenue. Current annualized revenue is closer to $5.4 billion. Closing that gap requires another $30–40 billion of capex, which requires $20–30 billion of new debt, into a market the article itself says has closed. The larger the backlog, the wider the funding hole. And most long-term contracts carry delivery-milestone clauses: miss a build schedule for lack of capital, and the customer can walk without penalty. Backlog evaporates precisely when you need it most.
Concentration compounds it. Microsoft accounted for about 62 percent of 2024 revenue — simultaneously the largest customer and the party most able to build its own capacity and reclaim the workload. Nvidia, meanwhile, invested in CoreWeave and extended purchase priority, closing a loop: invest, sell, recognize revenue, mark up the equity, refinance, buy again. While credit is cheap the flywheel spins forward. When it stops, it spins backward, and the accounting follows.
There is a crypto chapter the report ignores entirely. The same GPU-backed debt is being packaged on-chain: tokenized compute contracts, lending pools accepting cluster receivables as collateral, real-yield products that are really equipment leases with a hash-rate chart stapled on. Liquidity mining taught us the pattern — subsidize the TVL, and the moment the subsidy stops, the users leave. These pools are that pattern with better branding. The collateral depreciates faster than the yield accrues, and nobody publishing an APY is publishing a depreciation schedule.
The scarcity that actually matters is not silicon. It is interconnection queues, transformers, and gas turbines. A GPU cluster is a depreciating option on electricity; the land and the grid connection are the durable asset. That is why value is migrating toward the physical layer while the reseller tier gets flattened between hyperscalers building their own capacity and utilities that now hold pricing power.
The contrarian reading is not that the article is wrong — it is that being wrong about the numbers has distracted everyone from the right question. The report's figures are almost certainly synthetic or second-hand: entity confusion, timeline drift, a debt number that contradicts its own thesis. Treat them as directional signal only. The real exposure sits with the creditors, not the shareholders. Because the debt is structured through SPVs collateralized by specific clusters, the loss does not land on CoreWeave's equity first — it lands on whoever bought the mezzanine tranches, and on the secondhand residual market for H100s that nobody has priced. This is the Terra playbook with different collateral. I spent three months in Patagonia after algorithmic stablecoins failed, and the lesson was not that the math was broken. The math worked. The incentives were pointed at the wrong people. Here, the collateral's price is set by the same institutions lending against it. That is not a market. It is a mirror.
Watch two lines next quarter: the depreciation footnote in the 10-Q, and the spread on the structured tranches. Those will move before any headline does. The code remembers what the market forgets — and the ledger here is written in depreciation schedules, not press releases. When the herd wakes, the signal will have already faded into a footnote nobody read.
