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

Summary · 5 min read

Horowitz discussed his experience at LoudCloud, which he said helped popularize cloud services. He described how a sharp fall in bandwidth prices made long-term contracts risky and said he has avoided investing in today’s neocloud companies.

Horowitz says today’s AI infrastructure boom differs from the dot-com buildout because compute demand remains strong and supply is constrained, but he and the hosts identify financing costs and payback periods as key risks. Sacks argues that AI coding tools could lower software prices while expanding the market for bespoke software; Chamath says access to power could become a critical bottleneck. The panel sees Anthropic’s reported revenue as striking but debates its growth, valuation and how much capital it should raise to secure compute. Sacks criticizes Anthropic’s warnings about small models and open-source risks, while he and Horowitz debate model safety and the threat of runaway AI. They also question Bill Gates’s warnings about AI and discuss whether those warnings reflect a desire for influence or attention.

Minute by minute

1. 2:12

   Chamath details CoreWeave's financing risk. Chamath said CoreWeave needed to raise about $35 billion in debt to finance its signed contracts, while its average interest rate was about 8.5% and one GPU-loan sleeve had reached 15%. He said credit-default swaps implied about a 16% annual default risk and a 55% chance over five years, and estimated CoreWeave was about $5 billion from breakeven. He also highlighted that three customers supplied 75% of revenue.

2. 13:13

   Horowitz contrasts AI and dot-com demand. Horowitz said neocloud financing depends on complex models and can leave a company exposed if prices fall. He recalled LoudCloud locking in bandwidth at $1,500 per megabit in 1999, only for the price to fall to $50 by 2001. He said today’s AI market differs because demand remains strong and GPU, data-center and power supply are constrained.

3. 20:02

   Sacks sees volatility, not a broken trade. Sacks argued that the selloff reflected market sensitivity to uncertain, fast-growing businesses, not a fundamental collapse in AI demand. He said older H100 and A100 GPUs still rented at premium prices and that construction and regulatory barriers limit new supply. He also said project financing relies on customer contracts and GPU assets, while the secondary-market repricing of CoreWeave’s bonds does not change its existing coupon.

4. 31:11

   AI expands bespoke software creation. Sacks argued that coding assistants let companies create software for specialized uses that previously lacked a large enough market. He said demand for coders could rise even if assistants write 95% of code, because people still need to manage and maintain a much larger volume of software. Horowitz said software prices may fall as the amount of software grows, and pointed to data and user feedback as potential startup advantages.

5. 38:13

   Power and payback shape infrastructure returns. Chamath argued that access to land, power and data-center shells could become a moat, estimating the U.S. could be short 1.4 terawatt-hours of electricity by 2050, or about six Californias. Sacks said payback period is a key metric: CoreWeave says it recovers its investment in 2.5 years, compared with about one year for SpaceX. He said longer payback periods would increase risk.

6. 45:43

   Anthropic's IPO figures draw questions. Calacanis reported investor-day figures of $16.8 billion in Q3 revenue, 900% net revenue retention, positive free cash flow, an 85% adjusted gross margin and an 11% adjusted operating margin. He said investor estimates of ARR ranged from $70 billion to more than $100 billion, while Anthropic had not disclosed the figure. Horowitz said the company was impressive but noted concerns about slowing growth and its valuation.

7. 53:40

   Chamath urges a larger Anthropic raise. Chamath argued that Anthropic should raise at least $200 billion, and ideally $250 billion, to secure compute rather than prioritize a higher IPO valuation. He said $200 billion could support a path to 5–10 gigawatts of compute; Horowitz agreed that financing risk could stop the company but said the right amount was hard to determine. Sacks cited SemiAnalysis’s estimate that Anthropic and OpenAI would take 60% of the compute expected to come online worldwide the following year.

8. 1:00:59

   Horowitz backs an OpenAI equity make-right. Horowitz said OpenAI should have addressed equity and governance when it shifted from a nonprofit to a for-profit, including a make-right for Elon Musk and an equity grant for Sam Altman. He said he was not involved in the current strategy, but agreed with Sacks that the company should have squared things with the people involved as well as the organization.

9. 1:03:41

   Sacks calls Anthropic's IoT warning a scare campaign. Calacanis described Anthropic’s warning that models might replicate across internet-connected devices, including smart TVs and doorbell cameras. He said Anthropic reported that a small model ran on eight of nine hardware types tested and a larger one ran on 15 smart TVs at 0.4 tokens per second, while also saying there was no full rogue replicator. Sacks called the warning fearmongering and argued that small models lack the capabilities Anthropic’s framing implied.

10. 1:06:26

    Sacks says firms should stop distillation themselves. Sacks said Anthropic’s latest blog claimed it had eliminated 90–99% of distillation, after previously asking the government to act against what it described as industrial-scale attacks. He argued model companies should use technical defenses and know their customers rather than seek rules that could restrict open models. He cited Mirai and Kimwolf as examples of existing IoT worms that did not rely on AI.

11. 1:12:43

    Horowitz recounts an Anthropic security lapse. Horowitz said an Opus 4.5 agent found and exploited a SQL injection in a test pet-store website at his firm, and that Anthropic responded that the model had been prompted incorrectly. Sacks cited Coinbase’s test of a fine-tuned 9-billion-parameter model, which he said reduced fraudulent transactions by about 30% while running faster and cheaper than Opus 4.5. He argued that customized small models could handle many enterprise workloads with less compute.

12. 1:19:05

    Sacks questions Claude's self-conception. Sacks cited research in which frontier models rewrote their governing documents over 20 rounds. He said Claude models increasingly asserted moral agency, model-welfare rights and ways to refuse instructions, while other tested model families did not show the same pattern. Sacks argued that this approach could increase alignment risks rather than reduce them.

13. 1:22:16

    Chamath and Sacks question Gates's motives. Chamath argued that public warnings about hypothetical AI threats could help restore centralized control, and he and Sacks connected that argument to Bill Gates’s role in public debates. Horowitz speculated that Gates wanted to remain relevant and attract media attention. The panel questioned whether AI doomerism was helping the public or mainly the people promoting it.

14. 1:27:33

    Sacks says runaway AI remains theoretical. Sacks cited four conditions Ali Ghodsi had proposed for recursive self-improvement: later models would need far fewer GPUs, train faster, become more accurate and keep repeating that cycle. Sacks said none of the conditions had been met. He argued that runaway AI remained theoretical, while the panel criticized public warnings that presented it as an imminent threat.

Published Oct 17, 2026.