Revenue Doubles, Interest Eats a Fifth of It: The First Impairment Bill for AI Infrastructure Has Arrived

CoreWeave's Q3 report looks at first glance like a growth story: revenue of $3.66 billion, more than double year-on-year, with guidance raised. But the bottom half of the same income statement tells another story—net loss widened to $709 million, mainly due to debt interest costs climbing to $846 million. Put the two numbers side by side: interest has already eaten 23% of quarterly revenue. A company whose revenue doubles while its financing costs double even faster isn't growing—leverage is paying for growth.

The compute leasing business model dictates a naturally heavy capital structure: borrow to buy chips and build data centers, then recover costs through long-term rental income. Revenue is recognized upfront; depreciation and interest pile up later. The prettier the headline revenue, the heavier the liabilities and fixed assets on the balance sheet. CoreWeave's problem isn't that it isn't growing—it's that its growth must be fueled by debt. Once rental growth slows, the interest line won't slow with it. What the company didn't disclose: the average financing cost behind that $846 million interest bill, and the debt maturity structure for the next four quarters. Without those two, any 'profit inflection is near' call lacks a denominator.

Zoom out. On the same day, SoftBank presented another bill: a $3.7 billion impairment on non-refundable prepayments for data center equipment, $900 million of which relates to contract modifications. 'Non-refundable' was supposed to be the seller's moat for locking in demand—instead it became a write-down line on the buyer's books. Put the two together: across the AI infrastructure funding chain, upstream is writing down prepayments, midstream is paying high interest, downstream is compressing valuations—all three tightening at once, something that hasn't happened in the AI narrative of the past two years.

Revenue doubling is the language of the income side; interest and impairments are the language of the capital side—they're not talking about the same thing.

On November 10, Anthropic did two contradictory things. On one hand, it launched office software, embedding Claude into spreadsheets, presentations, and meeting notes, directly targeting Microsoft Office and Google Workspace—Microsoft shares fell sharply on the news. On the other, its prospectus disclosed an additional $10 billion in cloud server leasing over the next decade, and a $34.9 billion Lambda lease agreement that may be canceled. The former is a revenue story, the latter a cost story—and the latter is written in a legally binding document.

The phrase 'may be canceled' on a $34.9 billion lease carries more information than any product launch. It means long-term commitments previously counted as 'future compute security' could be repriced or even voided. For Anthropic, that's good news on cost flexibility; for compute suppliers treating it as certain revenue, it's bad news. The same document: the buyer reads flexibility, the seller reads uncertainty. What wasn't disclosed: how prepayments already made or committed would be handled if the agreement is canceled—exactly the same issue as the $900 million 'related to contract modifications' in SoftBank's $3.7 billion impairment.

AI critic Gary Marcus published a piece on IPO launch day titled 'Just Say No to Anthropic's IPO,' citing lack of a moat and massive capex; he said he wouldn't participate. That's one person's view—Marcus has long been skeptical of generative AI commercialization, and his judgment shouldn't be used as a pricing basis. But set his critique against the numbers in the prospectus, and at least one point is irrefutable: when a company commits to tens of billions in cloud leasing over a decade while acknowledging a multi-tens-of-billions agreement may be canceled, its cost structure isn't 'predictable'—and what valuation models fear most is unpredictable capex.

'May be canceled' in a prospectus does more damage than any bearish article—it puts a question mark on certainty itself.

Anthropic's IPO price range is $70 to $78, aiming to raise up to $64.4 billion, plus $26 billion in private placement, for a total of up to $90.4 billion. That corresponds to a valuation of $1.26 trillion to $1.4 trillion, below the market's earlier $2 trillion expectation. Earlier versions were messier: a $69.3 billion prospectus was withdrawn due to inaccurate pricing details; the initial version sought up to $70 billion. A prospectus withdrawn and reissued is itself a signal that pricing discipline is returning.

Lay out the timeline: the valuation expectation was cut repeatedly—from $1.5 trillion to $1.1 trillion–$1.3 trillion, with some views suggesting it could approach $1 trillion; then the roadshow range was set at $1.26 trillion–$1.4 trillion. This isn't a one-day mood swing; it's institutional investors repeatedly pressing prices before the roadshow. CNBC's Jim Cramer called the IPO a bellwether for the 'AI trade'—the core discussion isn't whether the company is good, but valuation pullback and institutional pressure. In other words, pricing power is shifting from the storytellers to those who sign the checks.

A cross-check makes it clearer: OpenAI's Q3 revenue jumped to $10 billion, up nearly 50% quarter-on-quarter, a marked recovery from 17% growth in Q2, and it's raising at a $1.25 trillion valuation. The two valuations are in the same order of magnitude, but one has proven quarterly revenue scale while the other is still using a decade of lease commitments to support its cost narrative. The market pricing them similarly shows that what's being anchored isn't current cash flow but 'who can survive until compute costs fall.' It's a bet on endurance, not on profitability.

Our calculation: using the $64.4 billion fundraising cap and the $1.26 trillion valuation floor, the price-to-sales ratio depends on Anthropic's undisclosed annualized revenue. The company hasn't disclosed that number, so any P/S multiple remains hypothetical. This must be stated clearly—most multiples circulating in the market use estimates, not disclosures, as denominators.

Pricing power has shifted from the storytellers to those who sign the checks—that's what this IPO has truly changed.

The first number is CoreWeave's interest coverage. The 23% ratio of $846 million interest to $3.66 billion revenue: if it rises next quarter, debt is still expanding faster than revenue and the asset-heavy model isn't yet self-funding; if it falls, then we can talk about scale effects. One quarter isn't a trend—two consecutive quarters are needed to confirm.

The second number is the final status of the $34.9 billion Lambda agreement and the $10 billion cloud leasing commitment in Anthropic's updated prospectus. If the agreement holds and prepayments amortize normally, that's evidence costs are controllable; if it's canceled or substantially modified, the problem shifts from Anthropic to its suppliers' books—see SoftBank's $900 million 'related to contract modifications' impairment, a transmission chain already previewed once.

The third number is how many other 'non-refundable prepayments' sit on AI infrastructure players' balance sheets besides SoftBank's. SoftBank's $3.7 billion impairment is the first public sample, but it's probably not the only one. What wasn't disclosed: the total scale of such prepayments across the industry and each company's impairment policy. Until definitions are unified, any claim that 'industry-wide risk is controllable' lacks a comparable basis.

The boundary to state clearly: impairment isn't cash loss—it's a reassessment of future economic benefits; high interest costs don't mean imminent default—as long as refinancing channels stay open, leverage can keep rolling. The three numbers in this chapter are signposts, not verdicts. They can be overturned by subsequent earnings—which is exactly why they're worth watching.

The first impairment sample has appeared. The question is whether it's an isolated case or the first in a series.