CRWV just printed the quarter the bulls were waiting for. Adjusted operating income came in at $128M against a $30-90M guide, 42% above the high end, the second straight quarter management beat its own margin promise. Revenue backlog hit $104B, with another $25B of commitments added in the first weeks of Q3. Interest expense even came in below the guided range. The stock jumped 15.7% after hours and has kept going, trading around $106 as I write, up roughly 17% from the pre-print close. All of that is real and I’m not going to argue the quarter was secretly bad, because it wasn’t. The demand is real, the margin inflection is real, and the operating execution is about as clean as this story has ever looked.My problem is with a different line in the same release: total debt grew from $24.9B to $35.1B in one quarter. And the more time I spend with this print, the more it reads like great news for the people who lent CoreWeave that money, and a much more ambiguous update for the people who own the stock.One quarter of operating profit, one-fifth of the interest billStart with the arithmetic the celebration skipped. The $128M of adjusted operating income that drove the pop sits next to $640M of net interest expense in the same quarter, up from $267M a year ago. The best operating quarter in the company’s history covered about a fifth of its own interest bill. For the first half it’s worse: $149M of adjusted operating income against $1,176M of interest. And the Q3 interest guide of $860-940M implies an exit run-rate around $3.5B annualized, versus a full-year adjusted operating income guide of $960M-$1,150M. On the company’s own preferred profitability metric, a full year of operating profit covers roughly 30% of the interest line, and that’s after a genuinely great margin quarter. The year-over-year comparison is worth sitting with too, because the “sequential inflection” framing hides it. In Q2 2025, CoreWeave earned $200M of adjusted operating income on $1.2B of revenue, a 16% margin. This quarter: $128M on $2.6B, a 5% margin. Revenue more than doubled and operating profit went down, because depreciation grew 149% against revenue growth of 112%. Scale is not yet improving the economics; it’s front-running them.Bulls will say the right lens is adjusted EBITDA, which came in at $1,510M, a 59% margin, covering interest 2.4x. But look at what sits between EBITDA and operating profit: $1,393M of D&A in the quarter. Depreciation consumes 92% of CoreWeave’s EBITDA. For a fleet of GPUs, that’s not an accounting fiction, it’s the business model — the assets wear out, or more precisely, get repriced by every new Nvidia generation. Take EBITDA, subtract the depreciation that a GPU fleet genuinely incurs, and you’re left with about $117M a quarter against $640M of interest. A coverage metric that ignores D&A on depreciating silicon is exactly the metric I’d want if I were selling the debt, not buying the equity.The cash flow needed prepayments to stay positiveThe funding math points the same direction, and the cash flow statement deserves a closer read than it got on the call. H1 operating cash flow was $3.7B, which sounds like real self-funding capacity. The composition says otherwise. It starts from a $1.4B net loss, adds back $2.5B of D&A and $0.3B of stock comp, and then the line doing the quiet heavy lifting: deferred revenue up $1.4B. That’s customers paying for compute before they consume it. CoreWeave is now sitting on $9.7B of these prepayments ($2.7B current, $7.0B non-current), a balance that grew $1.5B in six months.The Q2 number in isolation is starker. Operating cash flow was $679M for the quarter, and $790M of it came from new customer prepayments. Strip those out and Q2 operations consumed roughly $110M of cash. In the same quarter, cash capex was $6.4B. For the half: $14.1B of capex plus another $0.7B into joint ventures and strategic investments, against that $3.7B of OCF. That’s an $11B hole in six months, and the capex guide just went up again, to $35-39B for the year against $12.4-13.2B of revenue. The company will spend nearly 3x its revenue this year on the buildout, and it has raised that spending guide four quarters in a row.The hole got filled the only way it can be. Q2 financing inflows: $13.5B of new debt raised, $3.9B repaid, so about $9.6B net, plus a $1.0B private placement to Jane Street, plus $492M spent on capped calls to blunt the dilution from the new converts. Whatever else this is, it’s a machine that has to raise something like $25-30B a year for the next several years. The demand side no longer decides whether that works. The credit market does, and June-July already showed what happens when it wobbles: a 48.6% drawdown in six weeks with no change in CoreWeave’s own execution.The moat was supposed to be non-recourseThe financing story bulls tell, and it’s a good one, is that CoreWeave’s debt is project finance in disguise: asset-level, self-amortizing, non-recourse structures matched against contracted revenue from quality counterparties. If a facility’s contract goes bad, the lenders get the GPUs and the parent walks away. That’s not what printed this quarter. Of the $35.1B of total debt, $31.4B is recourse to the parent; the non-recourse piece is $3.7B, about 10% of the stack. The roughly $10B of new debt added in Q2 was mostly parent-level unsecured paper: the release itself touts “more than $10 billion of unsecured debt and convertible bonds,” including an inaugural Eurobond. The moat narrative rests on the non-recourse share scaling up over time, and this quarter it shrank as a share of the stack. Management did disclose one genuinely impressive data point, the $3.1B DDTL 5.5 term loan priced during one of the most dislocated credit weeks of the year, closed at full size, and spreads have retraced since. Credit where due, that’s an in-hand stress test passed. But it tells you the debt is available but says very little about what’s left over for the equity after servicing it. And debt isn’t the whole stack of claims ahead of the common. Operating-lease liabilities, the data-center leases behind the buildout, roughly doubled in six months, from $8.2B at year-end to $16.3B. Add it up and there’s about $51B of debt and lease obligations standing in front of $5.0B of book equity. The common stock’s claim on this balance sheet is 6.5% of $77B in assets, levered against everything above it.Meanwhile the share count is doing its own quiet work. Diluted shares went from 529M to 551M in a single quarter, up 4%, and from 487M a year ago, up 13%. Paid-in capital grew $3.1B in six months. Any honest multi-year model has to assume more of that, because equity is the other half of the funding stack. Even in the bullish scenarios, you get paid in a currency that’s being printed.Nobody knows what a five-year-old GPU earnsThe deepest problem is terminal value, and here I think the market let management’s best talking point do too much work. The quote that made the rounds: CoreWeave recontracted a 2020-vintage A100 SKU into 2029 at an attractive price, with renewals trading at or above year-ago ASPs. That sounds like proof the fleet has annuity value beyond the initial contract term. Read the fine print, though. Management itself described the capacity coming up for renewal as “a very limited part of our fleet.” One renewal cohort, on the oldest and smallest slice of assets, repriced during the tightest compute market anyone has seen. Of course it renewed well, everything rents well in a shortage.The equity case requires something much stronger: that GPUs bought today, inside a $35-39B annual capex program, still earn attractive returns in years four through eight of their lives, after Nvidia has shipped another generation or three and after the industry has added the capacity everyone is announcing. Nobody has that data. It doesn’t exist yet, not for CoreWeave, not for anyone. What we do know is the depreciation clock is already loud: quarterly D&A has more than doubled in a year, to $1.4B, chasing a net PP&E base that grew from $30.6B to $46.7B in just six months. Every terminal-value model on this name, bullish or bearish, is guessing at the single number that matters most, what a depreciated GPU earns through a full supply cycle, and one shortage-era renewal on 2020 silicon doesn’t settle it. The honest range on that number is wide enough to swing the equity from a multi-bagger to roughly zero, which is what the stock’s own 17-month trading history keeps telling us: two drawdowns past 45%, now two violent rallies.What the price now asks you to believeAt the pre-print close of $90, you could argue the market was paying for something close to the base case; the exit multiple required to earn 10% a year worked out to under 4x Year-3 revenue on a reasonable operating path. That argument was fair. At $106, with EV around $88B, you’re paying about 6.9x this year’s revenue guide, and the same hurdle math needs a low-4s exit multiple on roughly $34B of 2028 revenue. Still not heroic on paper. But every step of that path runs through something like $60-80B of additional external financing, a share count drifting toward the mid-600s, and a terminal multiple on assets whose through-cycle economics are unknown. The “cheap on the required multiple” frame quietly assumes the financing and the terminal value, which are the two things actually in question.My read is that last night validated CoreWeave the business but not the equity investment. The company proved it can rent GPUs profitably and borrow through a storm, it has not proved that the common stock, sitting behind $51B of debt and lease obligations, dilution that ran 4% in the latest quarter alone, and a terminal asset value no one can yet underwrite with any confidence, is the instrument that captures the value being created. After a 17% two-day rerate, I’d rather own the demand some other way and let the capital structure prove itself first.