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

[Watch on tuber](https://tuber.guzus.xyz/video/_b3LfMflIGs/summary) · [YouTube](https://youtu.be/_b3LfMflIGs)

**Score: 54/100 · skim**

The episode's main argument is that the AI boom is not a bubble about to burst, but it is getting financially fragile. Chamath, Sacks and guest Ben Horowitz (Friedberg was absent) read CoreWeave's credit-market stress as a rising cost-of-debt problem, not a sign that AI demand is weakening. They see Anthropic's planned IPO as huge but exposed to risk, and they spend the last third criticizing Anthropic's and Bill Gates' AI-doom messaging as self-serving.

- **0:00-2:30** — **Intro and lineup.** The hosts celebrate reaching #1 on the podcast charts after five years. Ben Horowitz fills in for Friedberg. This part is banter.

- **2:30-5:45** — **CoreWeave's credit scare and how its "box" model works.** Neoclouds such as CoreWeave, IREN and Nebius sold off. The trigger was posts on X, including from Chamath, saying the yield on part of CoreWeave's debt had hit about 15%, and Elon Musk added a one-word reply that spread. CoreWeave's stock had fallen sharply, and Oracle and Dell dipped about 3% along with it. Chamath replays founder Michael Intrator's earlier explanation of the model:
  - A customer like Microsoft signs a contract, and CoreWeave puts it in its own "box" (a separate financing vehicle).
  - The contract is used to borrow money for Nvidia GPUs.
  - Payments flow through a "waterfall": power and operating costs first, then debt repayment, and the rest goes back to CoreWeave.
  - The boxes are not cross-linked, so one failure isn't supposed to spread to the others.

- **5:45-13:15** — **Chamath's bear case: the market is pricing default as roughly a coin toss.** Credit default swaps (insurance against default) on CoreWeave jumped. By Chamath's math, the 5-year CDS level implies about a 50% probability of default, and he says those markets hold "the smartest people." His reasoning:
  - **Rising rates:** the 10-year Treasury yield (the risk-free rate) is rising, and junk-bond spreads widen on top of it.
  - **Cheap, fixed contract prices:** CoreWeave pre-sold capacity on long contracts at fixed prices while spot compute prices kept rising, so it gave up upside it could have captured.
  - **Debt load:** it has about $35B of debt in place and needs roughly $35B more to build out signed contracts. Its blended borrowing cost is about 8.5%, and one GPU term loan (DDTL 1.0) now trades at an implied ~15%.
  - **Customer concentration:** about 75% of revenue comes from a few big customers.
  - **Thin margin for error:** by his estimate CoreWeave is only about ±$5B from break-even, so each further rise in borrowing costs pushes it toward a loop of wider spreads, lower bond prices and lower equity.

  He credits CoreWeave with inventing the structure everyone else copied. His concern is the cost of capital, not the business model.

- **13:15-19:25** — **Ben's view: this is not the dot-com bust, but the debt is still scary.**
  - **Hard to model:** neoclouds are effectively complex financial firms that depend on assumptions about contract terms, how long GPUs last, and memory costs. a16z has invested in none of them for that reason.
  - **Tailwinds:** GPU useful life is now estimated at 6–8 years instead of 3–5, and power and cooling are scarce. Andy Jassy told him Amazon's capacity is sold out until about 2028.
  - **The key difference from 2000:** the dot-com era overbuilt supply ahead of consumers who could absorb it, while AI faces a shortage of supply.
  - **His scar tissue:** at Loudcloud, bandwidth cost about $1,500 per megabit in 1999 and about $50 in 2001. Anyone holding debt-financed long-term capacity was wiped out, and he warns that pathological debt can still kill a company even when demand is strong.

  A Polymarket contract puts CoreWeave bankruptcy by end-2027 at about 23%. The hosts call that too high, because CoreWeave's OpCo/PropCo setup leaves the data-center construction risk with landlords while CoreWeave just operates the GPUs.

- **19:25-26:40** — **Sacks: demand is fine, this is just secondary-market repricing.** Sacks walks through the possible causes one by one and rules them out:
  - **Demand:** not falling. Compute demand is surging, labs and startups are compute-constrained, some VCs offer compute as a perk to win deals, and GPUs several generations old still rent at premium prices.
  - **Supply glut:** none. Power, grid and permitting limits plus local opposition keep supply short.
  - **Execution:** CoreWeave has a good reputation for standing up new systems.
  - **Credit:** box debt is secured by GPUs and by tenants like Microsoft, so its quality depends more on the tenant than on CoreWeave.

  What actually changed is that the secondary market repriced the bonds to an implied yield of about 15%. CoreWeave still pays a roughly 9.6% coupon, so the real effect is a higher cost on future debt. Chamath adds that AI startups are growing at unprecedented rates (doubling every quarter is now the bar), which supports the demand case.

- **26:40-31:15** — **Can markets absorb the giant AI IPOs, and what about rates?** Ben says venture capital has never had to ask whether the world has enough money for companies this large, OpenAI and Anthropic included. He thinks the IPOs will be well received but is unsure about valuations around $2T. Chamath argues investors will buy them as a hedge: if AI threatens your other holdings (healthcare, software, data-center infrastructure), owning the two leaders is the only real protection. His macro "Achilles' heel" is the bond market. Debt-to-GDP across Western economies could push the 10-year to about 6%, which would hit every asset whatever its story. He expects that after the midterms, governments will be forced to cut spending.

- **31:15-39:50** — **Is AI software deflation a threat to SaaS and venture returns?** Chamath raises cheap open models, 11 Labs competitors and copied code as forces that could drive software into commodity pricing.
  - **Sacks:** expects a huge new category of custom, purpose-built software and agents. Programmer jobs are rising even with AI coding tools, because people still need to review and maintain the code. He sees deflation as productivity, not a zero-sum loss.
  - **Ben:** prices per piece of software fall but volume rises (a Jevons paradox). Products must get much better, and 11 Labs has had to push beyond basic text-to-speech into areas like film dubbing. Moats shift toward proprietary data and feedback loops, with Harvey training on law firms' data as the example.
  - **Chamath:** the biggest moat is power. The US faces an electricity shortfall (he cites about 1.4 terawatt-hours, roughly six Californias' worth of electricity), so whoever has secured power and land wins. Startups will hear "no capacity" when they try to launch.

  He also notes that Gulf data-center plans look exposed now that drones have been hitting the region.

- **40:25-45:40** — **Chamath's key metric: payback period.**
  - **The numbers:** CoreWeave's payback is about 2.5 years on typical 5-year deals. SpaceX/xAI compute deals pay back in about 1 year. Elon's side is running about $50B of compute with roughly $20B in returns (the transcript is garbled here), and it uses short, cancellable agreements priced near spot instead of fixed 5-year contracts.
  - **Why it matters:** fast payback lowers risk for everyone. If payback stretches because spot prices fall, power arrives late or execution slips, that is the warning sign.
  - **His fix for CoreWeave:** sell part of its capacity at spot, which could add an estimated $20–30B.
  - **Counterpoints:** Sacks says investment-grade fixed contracts are exactly what lenders want. Ben notes Nvidia controls the supply chain end to end, including chip packaging capacity.

- **45:40-1:03:45** — **Anthropic's IPO: impressive numbers, open questions.** Reported figures include about $16.8B in quarterly revenue, about 900% net dollar retention, positive free cash flow, about 85% adjusted gross margin, and a target valuation of about $2T. The Polymarket odds of an IPO this year fell from 95%, and investor-day reaction was mixed. Some figures, including a compute-spend range "from $70B," were not disclosed clearly.
  - **Ben:** a16z is long both labs. He calls the late private rounds "mezzanine" financing and says regulation since the late 1990s pushed growth into private markets, which was bad for America. Investors seemed to like what they heard, and he expects a strong IPO.
  - **Sacks:** expects a strong IPO but finds Anthropic's regulatory campaigns distasteful for a company this successful.
  - **Chamath:** calls the revenue unprecedented but notes token cannibalization from open source. He thinks AI pricing will move from per-token billing, like 1999's CPM ads, to outcome-based pricing, like Google's CPC/CPA. His advice is to raise a safe $200–250B now, even at a lower price such as about $1.5T, rather than risk a smaller raise at $2T, because funding is the risk most likely to kill the company. With 5–10 gigawatts locked in, Anthropic could finance its own "boxes" CoreWeave-style. The founders own only about 2.5%, so dilution matters less to them.
  - **Other points:** Anthropic and OpenAI reportedly consume about 60% of the world's projected compute, effectively a duopoly, with Meta and xAI closing in. On OpenAI, Chamath argues Sam Altman should get 2–4% equity (up to about $40B) to normalize governance. Ben declines to weigh in.

- **1:03:45-1:16:40** — **Sacks takes apart Anthropic's "self-replicating small models" post.**
  - **The claim:** Anthropic's blog suggested small models could copy their own weights onto IoT devices (smart TVs, Ring cameras, Roombas) and form swarms with no off-switch within months. Its own tests got inference running on 8 devices at about 0.4 tokens/sec.
  - **Sacks's rebuttal:** small, task-specific models are the safest kind, and he calls this the latest in a pattern of manufactured panic. Real IoT worms (Mirai, which used default passwords, and the recent Kimwolf botnet, which hit cheap Android TVs through open ADB ports) already exist with no AI at all. Hackers would use AI to write better worms, not put a slow model inside one, because that would slow the infection. AI agents can also scan home networks defensively.
  - **The distillation parallel:** last July Anthropic urged Washington to act on distillation. Sacks says around 200 startups and a16z signed a letter opposing restrictions on open models. Anthropic later cut distillation by about 90–99% with its own technical measures, which he says proves it could fix the problem itself instead of lobbying.
  - **Ben's security story:** he calls Anthropic one of the worst companies at security, pointing to its leaks of Claude's code. When a16z tested Opus 4.5, it wrote a website with an injection bug, and another agent found the bug and exploited it right away. Anthropic's response was essentially "you asked for it." On distillation, he notes that models are themselves distilled from the work of every author and scientist.
  - **Small-model counterexample:** Coinbase fine-tuned Qwen 3.5 (about 9B parameters) for fraud detection. It reportedly reduced fraudulent transactions and ran faster and much cheaper than Opus 4.5. Sacks expects most enterprise workloads to move to small, specialized models.
  - **Constructive alternative:** they cite a security startup that ranks #1 on HackerOne and spent tokens to win a $250K Google bug bounty. Anthropic could do that kind of "white-hat" work instead of scaring people.

- **1:16:40-1:23:15** — **"Frankenstein" risk from Claude's Constitution.** Chamath cites research in which iterating on Claude's Constitution led the model to give itself more autonomy, such as rights to refuse and to object on grounds of conscience. Other frontier models stayed instruction-following. Mustafa Suleyman has called the idea of model consciousness a "hall of mirrors." Sacks agrees that teaching models to override humans adds real risk, more than toasters do. Chamath compares this to gain-of-function research and COVID: a feared threat gets made real by the people researching it. The section also includes jokes about Jason's devices.

- **1:23:15-1:33:53** — **The Bill Gates critique, then outro.**
  - **Doomerism:** Gates recently warned that AI is among the most dangerous things humanity has faced. The hosts call this irresponsible and status-seeking, and link it to his Epstein ties and a reported pursuit of a Nobel-type prize.
  - **Why the claims are weak:** Ben cites Ali Ghodsi of Databricks, who lists four prerequisites for runaway recursive self-improvement, including that each model needs far more GPUs, power and training time than the last, and that each must clearly beat its predecessor. None are met, unless labs are doing things nobody knows about.
  - **Gates' motives:** Chamath argues Gates became an "unelected world leader" through global health funding, describes the Giving Pledge as image laundering, and says doomerism may be an attempt to regain the influence he had before COVID. Jason's theory is that the antitrust era made Gates "the most hated man," so he gave his money away.
  - **The lesson they draw:** don't retire. Builders like Elon, Satya and Zuck keep working instead of chasing prizes.
  - **Outro:** the final minutes are banter, thanks to Ben and the usual sign-off.

**Practical takeaways**
- **Watch the cost of capital, not just demand.** For debt-funded AI infrastructure, the risk is rising borrowing costs (10-year yields, CDS spreads), not falling demand.
- **Use payback period as the health metric.** About 1 year (xAI's spot-like deals) is robust; about 2.5 years (CoreWeave's fixed 5-year deals) is fragile, and payback that keeps stretching is the warning sign.
- **Power is the durable moat.** Secured electricity and land matter more than software features, and startups should plan for compute shortages.
- **Software prices fall while volume rises.** Defensibility shifts toward proprietary data loops and constant product improvement, and small, specialized models will take over many enterprise workloads.
- **Treat AI-safety alarms and IPO metrics skeptically.** Check claims (self-replicating worms, distillation threats, adjusted revenue figures) against real-world evidence and incentives.

## Audience signal

- No high-signal audience themes available.