CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer

CoreWeave, IREN, and Nebius fell about 20%, including 10%–15% drops on Wednesday, as investors reassessed the financing risk in AI infrastructure. CoreWeave bonds traded around 80 cents on the dollar, implying yields near 15%; that does not change the 9.6% coupon on existing debt, but could raise the cost of future borrowing. Chamath cited a 953-basis-point CDS spread, corresponding to market-implied default probabilities of roughly 16% over one year and 55% over five—not a forecast that CoreWeave will default. The company has about $35 billion in debt and needs another $35 billion to fund contracted capacity, with roughly half still unfinanced. Its average borrowing rate was described as about 8.5%, while one GPU-loan tranche recently reached 15%. Speakers disagreed about immediate distress: one estimated CoreWeave was about $5 billion from break-even and warned that higher financing costs could squeeze margins; another noted roughly $78 billion in assets against $35 billion of debt. Three customers contribute 75% of revenue, and some are lowering future revenue expectations.

CoreWeave says each “box”—a customer contract, GPUs, and a data-center contract—covers data-center costs, power, interest, and principal before generating a remainder for the company, and that each box pays for itself within two and a half years of a five-year deal. Chamath countered that fixed-price commitments limit upside if spot prices rise, while borrowing costs remain exposed to market rates. CoreWeave has built two of four gigawatts of contracted capacity, and one speaker estimated the gap between contracted and spot pricing could leave $20–30 billion on the table; exiting some contracts might require paying penalties. Yet five-year commitments help secure financing, and large offtakers can impose substantial penalties for failing to deliver. A comparison put CoreWeave’s data-center payback at 2.5 years versus roughly one year for SpaceX, which was estimated to monetize about $50 billion per gigawatt against CoreWeave’s $20 billion. SpaceX reportedly can capture spot pricing on deals cancellable with 90 days’ notice, unlike CoreWeave’s longer fixed commitments. Hotter chips and operating complexity add execution risk.

The financing debate extends beyond CoreWeave: one speaker warned that a 10-year Treasury yield of 5.5%–6% could damage businesses dependent on financing spreads, and said Microsoft, Google, Meta, and Nvidia might need to support infrastructure companies through a one- or two-year correction. Other speakers argued that AI demand remains strong and that compute, power, and suitable data-center capacity are scarce. Older H100 and A100 GPUs still rent at premium prices, while construction and regulatory obstacles slow new supply; one speaker cited a forecast that supply may not return until after the end of 2028. Power was described as the ultimate AI moat, with a projected US electricity shortfall of about 1.4 terawatt-hours by 2050 even if planned projects arrive on time and on budget, and without additional AI demand. Speakers also warned that cheap AI tools and open-source models could erode software-product value. Sacks said coding assistants enable bespoke software for previously underserved uses; Ben argued that more software to manage and maintain could increase demand for coders, while durable value may shift toward products, licensing, specialized models, or data.

Anthropic’s Investor Day produced both strong numbers and disclosure concerns. The company reported $16.8 billion in Q3 revenue, 900% net revenue retention, positive free cash flow, and adjusted gross and operating margins of 85% and 11%, but did not disclose ARR or explain its adjusted-margin calculations. Investor estimates of ARR ranged from $70 billion to over $100 billion, were second-hand and questioned, and revenue growth reportedly slowed substantially from Q2 to Q3. Anthropic was reportedly planning an IPO around Thanksgiving, though prediction-market odds of an offering this year fell to 74% from 95% in early September. Sacks called it a strong business but put its debated valuation around $1.9 trillion to $2.3 trillion; Chamath argued it should raise at least $200 billion. Proposals ranged from $200 billion to $250 billion, while Ben said the amount should be derived from three to five years of compute needs. Compute remains constrained across chips, packaging, and high-bandwidth memory; one cited projection has Anthropic and OpenAI taking 60% of new compute coming online over the next year. Speakers disagreed over whether a $50–70 billion raise would leave Anthropic unable to afford spot-market compute in two or three years, or whether a much larger raise is justified by securing supply.

The discussion also challenged dramatic AI-risk claims. Anthropic warned that small models might self-propagate across IoT devices within months, but said its tests showed only “signs of life,” not a rogue replicator; a critic called the framing fearmongering and noted that existing IoT worms do not establish that slow, purpose-built models can spread broadly. Speakers said Bill Gates described AI as capable of causing “a billion deaths” and as “the most dangerous thing that humans have ever gone near.” Some interpreted his warnings as an effort to regain relevance or centralized influence, while others said his Microsoft record and political reach make his views consequential. Speakers compared his pandemic-preparedness advocacy with his AI warnings, including the argument that gain-of-function research on bat viruses helped create the COVID threat it aimed to anticipate. They also said runaway AI self-improvement remains theoretical: the proposed cycle would require successive models to use far fewer GPUs, train faster, become more accurate, and repeat, conditions they said have not been demonstrated, while allowing that a major lab could be making private progress.