The AI Bubble Math Finally Broke

Ed Zitron told The Tech Report that something profound has changed in AI finance, and he put the number on debt.

CoreWeave's fixed-term loans touched a 16% yield this week, Zitron said, with every tranche now above 14% and one above 15.3% as of the recording.

That is personal loan pricing, not infrastructure credit, and Zitron argued it reprices all neocloud borrowing. He said most neoclouds and data center special purpose vehicles will now face 10% to 13% if they try to raise anything at all. A brand new neocloud called VA is already trying to raise a leveraged loan above 11%, he said, with JPMorgan and Goldman Sachs trying to sell the debt rather than hold it.

The signal, in his telling, is not a stock wobble. It is the bond market deciding the AI buildout may not pay. Zitron said he has spent months in debt markets precisely because equities are emotional and easier to move with a headline from Nvidia chief Jensen Huang. Debt, he argued, boils slower and is harder to put back in the cup.

The math he laid out was blunt. Goldman Sachs, per Zitron, estimated hyperscalers would need about $360 billion in brand new AI revenue just to break even on 2026 to 2027 capex, and about $425 billion to earn a 10% return on invested capital. Current AI revenue is $140 billion to $170 billion if lucky, he said. Gartner, which Zitron noted had hyped neoclouds as essential in October 2025, now projects 80% of neoclouds will be dead by 2030, and estimates AI data center owners need to generate a cumulative $8.5 trillion through 2030 to cover announced capital spending, financing costs and a 12% return, Zitron said.

CoreWeave is the canary because its numbers are public. The stock is down 18% year over year as of the recording and only about 20% above its IPO debut, Zitron noted, while the Financial Times ran a headline that its bonds are screaming distress. He cited an analyst estimate that CoreWeave needs to raise another $100 billion in debt by 2030 to fulfill its build, at a time when depreciation has not even started on most chips and on data centers being built by Core Scientific. Even on an operating basis cash flow positive, he said, depreciation will eat profits for years.

Anthropic was the other leg to break, according to Zitron. Leaked Q3 figures ahead of its planned IPO showed $16.8 billion in revenue, he said, up from $11.5 billion in Q2 and $4.8 billion in Q1. That is 46% quarter over quarter, down from doubling the prior quarter. The company also cited an 11% adjusted operating margin and an 85% adjusted gross margin, both figures Zitron dismissed as cost exclusions. It did not provide a new annualized run rate after touting $65 billion in July, he said. That omission, he argued, mattered more than the absolute growth.

The commitments behind that revenue are larger than the revenue itself. Anthropic carries $413 billion in non-cancellable commitments, per Reuters reporting cited by Zitron, and that compute is funded by others. Google has taken on more than $800 billion in off-balance sheet agreements to provide compute to Anthropic, he said. Broadcom has a $35 billion private credit facility with Apollo and Blackstone where $30 billion is backstopped by Broadcom if Anthropic does not pay, he said. Hut 8 and Fluidstack are among other counterparties tied to that growth.

Revenue quality looked thin too. Zitron said Anthropic disclosed 5% to 7% of revenue came from distillation, where Chinese model companies use Anthropic outputs to train their own models, and claimed it has cut those attacks by 90% to 99% in the last three days. Even if that headwind is gone, the trajectory does not close the gap. To keep its current growth rate, Anthropic would need about $24.5 billion next quarter and roughly $35 billion to $40 billion per quarter within a year, Zitron calculated. Investors had earlier priced it for a $100 billion annualized run rate by year end and $180 billion to $200 billion by 2028 to justify its IPO valuation, he said, citing Reuters.

What the numbers do not show is costs. Zitron noted Anthropic has not disclosed Q3 costs, and he suspects training was deliberately pulled back to prop up adjusted margins as the company oscillates between inference and training spend. That pullback, he said, is why users sometimes feel models get dumber. Without true cost, the 85% gross margin tells you little about sustainability, and the same opacity has inflated the bubble through run rates that multiply an undisclosed period by an undisclosed factor.

There is a counter-argument embedded in his own metaphor. He described the unwind as a boil or a decay, not an explosion, where contracts do not get signed and debt does not get raised and bodies are found slowly. That framing is less falsifiable than a default date, and it leans on leaks and secondary pricing, not filings, for the most damning figures. CoreWeave and Anthropic have not confirmed insolvency timelines, and hyperscaler capex cuts remain analyst expectations, not announced plans.

One sentence would change the read. If hyperscalers publish credible cuts to 2026 capex or if Anthropic files audited costs showing real operating leverage, the distress pricing could ease and the $8.5 trillion hurdle would look less like a wall.

Until then Zitron argued the usual rescue does not work. He said Japanese banks MUFG and SMBC, once the loosest funders alongside Blue Owl, have become more selective on data centers, per a recent Information story he cited, and that Huang's claim that Nvidia does not need to backstop CoreWeave ignores Business Insider reporting from February that Nvidia already backstops CoreWeave leases. The question his numbers leave is simple. Who generates trillions in new AI revenue by 2030 if the largest customers are venture-backed startups burning $1,200 in tokens for a $200 subscription, and the lenders now charge 14% to find out?