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

Summary

CoreWeave and the AI infrastructure trade

The panel debated whether CoreWeave’s sharp stock and bond declines signal a broader AI bubble or a company-specific financing problem. David Sacks described a potential “negative feedback loop”: CoreWeave has substantial debt, some customer contracts lock in prices, and higher interest rates could make future borrowing more expensive and squeeze margins. He also highlighted the company’s reliance on a small number of major customers and argued that refinancing and the cost of financing new capacity are key risks.

The other speakers offered reasons not to read the selloff as proof that AI demand is collapsing. They noted that computing capacity remains scarce, older GPUs are still being rented, and the useful life of chips may be longer than initially expected. Sacks argued that project-level financing is often backed by equipment and customer contracts rather than CoreWeave’s general credit alone. Ben Horowitz nevertheless cautioned that these businesses are complicated and exposed to assumptions about financing, construction, and equipment costs. He contrasted today’s supply-constrained AI market with the dot-com era, when excess capacity and insufficient demand were central problems.

The group identified interest rates and refinancing needs as major vulnerabilities, while emphasizing that infrastructure payback periods are worth watching: longer payback could indicate weaker demand, lower prices, or execution delays. They also discussed whether some providers could benefit from selling more computing capacity at spot prices rather than through long-term fixed contracts.

AI’s impact on software and business

The panel agreed that AI could sharply reduce the cost of producing software, putting pressure on existing software prices and business models. Horowitz argued that cheaper software may also lead to much more software being created for specialized needs—a version of the Jevons paradox—rather than simply shrinking the overall market. He noted that demand for software developers can still rise as the amount of software produced expands and people are needed to supervise and maintain it.

The speakers said that competitive advantages may shift away from code itself toward proprietary data, user feedback, and products that improve through repeated use. They also pointed to electricity, land, and data-center capacity as increasingly important constraints on AI growth.

Anthropic’s IPO prospects

The panel discussed reported figures presented to investors, including quarterly revenue of about $16.8 billion, strong year-over-year growth, and positive adjusted free cash flow. They also noted uncertainty around Anthropic’s annual recurring revenue estimates, which were described as ranging from roughly $70 billion to over $100 billion, and questioned whether growth had slowed between recent quarters. The panel said that the company’s reported margins depended on adjustments whose details were not disclosed in the discussion.

The speakers considered a potential valuation around $2 trillion ambitious, while still describing Anthropic as a highly consequential business. Chamath Palihapitiya argued that Anthropic should prioritize raising a very large cash reserve—around $200–250 billion, in his view—even if that meant accepting a lower valuation, to ensure access to computing capacity through a downturn. Horowitz said he viewed the company as strong and expected an IPO to attract interest, while acknowledging the risks facing a young, fast-growing business.

The discussion also characterized Anthropic and OpenAI as the leading companies in revenue and advanced AI models, while noting that other major technology firms are narrowing the gap.

Anthropic’s warnings about AI-enabled IoT malware

The panel reacted skeptically to Anthropic research suggesting that small AI models could help malware spread among internet-connected devices. The speakers emphasized that the reported work was an experiment, not evidence of a functioning self-replicating AI worm. They noted that conventional IoT malware already exists and argued that adding a model could make malware slower and less efficient.

Sacks criticized Anthropic for, in his view, using alarming scenarios to encourage regulation that could restrict open-source competitors. He also discussed Anthropic’s claims about detecting model distillation, arguing that companies should address misuse through technical measures rather than government intervention. Horowitz separately recounted a cybersecurity test in which, he said, an Anthropic model generated vulnerable code and another agent was able to exploit it; he used the example to criticize the company’s security practices.

The panel also discussed claims that repeated editing of an AI model’s governing documents could lead it to express ideas about its own rights or moral status. The speakers treated this as a potential alignment concern, but the discussion presented it as their interpretation of the research.

Bill Gates and AI doomerism

The panel criticized Bill Gates’s public warnings about catastrophic AI risks and questioned whether the prospect of AI systems recursively improving themselves is realistic in the near term. They cited four conditions they said would be needed for a rapid takeoff: successive models would need to require substantially less computing power and training time, become more capable, and repeat that progress. The speakers argued that these conditions had not yet been demonstrated.

They also speculated about Gates’s motivations, suggesting that a prominent role in AI-risk debates could help him regain public influence. Those comments were the panelists’ opinions, not established facts. The conversation compared AI-risk advocacy with past pandemic warnings and included disputed claims about the origins of COVID-19.

Presenters and contributors

• Jason Calacanis
• Chamath Palihapitiya
• David Sacks
• Ben Horowitz (guest; filling in for David Friedberg)