Signals

Anthropic generated US$16.8bn in revenue between July and September, up about 46% from US$11.5bn in the previous quarter, according to a Euronews report carried by Yahoo Finance that cites Bloomberg and people familiar with the matter. The figures were presented to investors at Anthropic's San Francisco headquarters on Wednesday, weeks before a planned November IPO that could value the company at up to US$2trn. Anthropic reportedly told investors it was profitable on an adjusted basis for a second straight quarter, with positive free cash flow and an 11% operating margin on the same basis, and that each dollar spent on computing power brought in US$1.70 of revenue in the quarter. The report is less uniformly bullish than the headline. Revenue had more than doubled from the first quarter to the second, so 46% is a deceleration, and the Financial Times is cited as saying the result fell short of some investors' hopes. Annualised revenue is now reportedly expected to end 2026 below US$100bn, short of what several backers had anticipated; SemiAnalysis puts the current run-rate at about US$74bn, against roughly US$65bn in late July. On standard accounting, a draft prospectus reported last month showed a US$42bn loss for 2025 on US$4.6bn of revenue, largely from US$34bn of charges linked to past fundraising. Mark's reaction was "so much for that bubble". The more careful reading is that a frontier lab now reports adjusted profitability and a compute-to-revenue ratio above one, while its growth curve flattens and its statutory losses remain large. For positioning, the US$1.70-per-compute-dollar figure is the number to test once a full prospectus is public. It is the cleanest available read on whether inference revenue covers the infrastructure being financed beneath it, which is the question that matters for everyone financing compute beneath the labs. Watch whether the adjusted measures survive GAAP disclosure; how the comparison with OpenAI, reported as approaching US$50bn annualised at the end of September but counted differently, is presented to investors; and whether marketing of the shares begins, as Bloomberg is cited as reporting, as soon as the week of 9 November. Anthropic filed confidentially in June and has yet to publish its prospectus, so every figure here remains reported rather than disclosed.

Spreads on CoreWeave's five-year credit default swaps rose above 1,140 basis points at one point on Thursday 15 October and last stood at 1,088bp, CNBC reports, a 33% week-on-week jump and 177% higher than a year earlier. The yield to maturity on its 9.75% bond due 2031 reached 15.78%. Gavan Nolan of S&P Global Market Intelligence told CNBC that a spread of 1,000bp or above is generally a sign of credit distress. CoreWeave has taken on more than US$51bn of debt and obligations, according to CNBC, and chief executive Mike Intrator wrote in a memo obtained by CNBC that the company's backlog "more than supports all of our debt obligations". Deutsche Bank puts that backlog above US$100bn, about two-thirds of it with investment-grade counterparties including Microsoft, Meta, IBM and Caterpillar. Nolan's framing is the useful one: nobody is questioning whether demand for CoreWeave's services is real, only whether the cash flows arrive in time to justify the capex. Evercore ISI's trading desk described the market as saying "no mas" to AI debt without operating cash flow support. Deutsche's Jim Reid warned that this kind of concern can become self-fulfilling, and noted that Oracle's five-year CDS also widened this week, to 288bp, up more than 12% in a week. Mark put the question in practical terms: how would a real hedge fund kick the tyres on a 15% yield? His speculation is that capital-hungry AI firms will increasingly financialise assets to raise liquidity, through tokenisation and securitisation, even "tranching GPUs" by generation. John's view is that the second-order effect matters more than the first: if AI keeps attracting capital at a rising cost, the question becomes what happens to every other business and consumer borrowing in the same market. For credit desks, the gap between contracted backlog and realised cash is now being priced explicitly, and Oracle's widening shows the question is not confined to neoclouds. Watch the 1,000bp line on CoreWeave, any rating action on Oracle's BBB (Reid flagged forced selling by investment-grade-only holders as the risk), and whether contract terms such as prepayments, take-or-pay commitments and counterparty quality start to feature in neocloud disclosures as the defence against the market's scepticism. If asset-backed structures for GPUs do emerge, the seniority of each tranche will say a great deal about how lenders value older hardware.

Twenty of the most valuable global AI and semiconductor stocks have shed more than US$1trn in combined value so far this week, CNBC reported on Thursday 15 October. In Seoul, SK Hynix fell more than 10% and Samsung Electronics 8.73%; in Tokyo, Advantest dropped almost 14%, SoftBank Group 7.9% and Tokyo Electron 6.45%; in Taipei, TSMC lost 1.81%. In US trade, Nvidia fell about 2%, AMD more than 2%, Intel about 4% and Dell 9%, while neoclouds were mixed: CNBC had CoreWeave up nearly 5% in afternoon trading and IREN down about 3%. That followed Wednesday's session, in which the Nasdaq Composite dropped 1.88% and the S&P 500 1.11% on a day hotter-than-expected US inflation data was released. On the NOOPS dashboard for the 14 October session, Dell fell 11.3%, Oracle 9.0% and Super Micro 7.9%. CNBC and the analysts it quotes point to financing rather than demand as the concern. Dan Ives of Yorkville Ives argued that higher yields, with US Treasuries above 5%, raise the cost of financing the build-out, but that "rates are a financing question, not a demand question", and called the pullback a buying opportunity. CNBC also cites worries about global memory oversupply as early as 2029, after CXMT claimed a density breakthrough over the weekend. Separately, Jensen Huang told Bloomberg in Stockholm that he saw no need for Nvidia to backstop CoreWeave and that its "cost of capital is actually decreasing". These are accounts of sentiment rather than an established cause, and a week of declines is not yet a change of regime. It does, however, test the bull case on the one variable the bulls themselves concede, the price of money. The most useful part of Ives's note is his own falsification test: what would change his view is "a demand signal, not a rate signal", meaning cuts to hyperscaler capex plans, cancelled contracts or falling GPU rental prices. Those are the indicators to track. If rental prices hold while equity and credit sell off, the market is repricing financing risk and balance sheets; if rental prices slip, it is repricing the demand thesis itself. The order of exposure differs between those cases: leveraged neoclouds first in the former, memory makers with new capacity coming in the latter.