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

Summary

• The CoreWeave scare is a credit-market story, not a demand story: Chamath walked through why credit default swaps on CoreWeave spiked to 953 basis points this week — an implied ~16% annual default risk and roughly 55% over five years. “It’s a coin flip whether they default or not. That’s not what I’m saying, that’s what the insurance market was saying.” With ~$35B of debt in place, another ~$35B still to raise, fixed-price customer contracts, and 75% of revenue from three customers, his team estimates CoreWeave is about $5B from its break-even point. He called the resulting CDS-to-bond-yield movement a “negative reflexive loop”; the financing outlook should become clearer over the next month.
• Sacks’s call: this is volatility, not evidence that the bubble has popped, because demand hasn’t budged. His checklist: compute demand is off the charts, old H100s/A100s are still renting at premium prices, and physical and regulatory constraints make a supply glut unlikely. The current bond repricing is of unsecured paper: CoreWeave still pays its 9.6% coupon even as the secondary-market implied coupon is around 15%. His North Star metric is payback period — CoreWeave says 2.5 years on five-year deals; SpaceX roughly one year, at about $50B per gigawatt versus CoreWeave’s ~$20B. Chamath, meanwhile, warned that high financing costs remain a serious risk.
• Ben Horowitz explained why he and Marc, Loudcloud pioneers, passed on the neocloud space: it can be “a very dangerous arbitrage business.” Bandwidth went from $1,500 per megabit in 1999 to $50 by 2001 — “if you borrowed money to get it, you’re dead.” The difference this time is a genuine shortage: Andy Jassy told him capacity is sold out “beyond 2028,” whereas the internet bubble died from lack of demand.
• Chamath’s macro kill switch: if the bond market walks the 10-year to 5.5–6%, “any business model that thrives on a financing spread gets cornholeed… you’re gonna be roadkill.” He identified variable-rate financing against the risk-free rate as the Achilles’ heel for these businesses; surviving a one-to-two-year correction requires Microsoft, Google, Meta, and Nvidia to backstop them. His other call: power is the ultimate moat — he said the US is ~1.4 terawatt-hours short by 2050, “about six Californias” — so “get long energy,” because hyperscalers may tell startups, “Suck it. We have no capacity for you.”
• On Anthropic’s ~$2T IPO: the secondhand Q3 revenue estimate of $16.8B and reported 900% net revenue retention drew attention, but Chamath wants Anthropic to raise at least $200B and ideally $250B to secure compute. He argued $200B “barely gets them a few gigawatts”; he said a $250B raise might require pricing around $1T, rather than raising $80B at $1.7–1.9T and relying on a later shelf offering. Ben, “a former CEO who almost died 17 times,” agreed financing risk is the thing most likely to stop the company. Separately, SemiAnalysis estimates Anthropic and OpenAI are taking 60% of the compute expected to come online globally next year.
• PSYOP corner: Sacks called Anthropic’s blog warning that self-replicating AI worms could infect IoT devices “in just a few months” a scare campaign against small models. The timing he highlighted: a day after Brian Armstrong tweeted that Coinbase fine-tuned Qwen 3.5 (the nine-billion-parameter model) to cut fraud by about 30% while being faster and cheaper than Opus 4.5. Anthropic’s other blog claimed it had blocked 90–99% of distillation; Sacks said that vindicated the argument that model companies, not the government, should build technical defenses.
• Ben’s criticism of Anthropic included a security anecdote: his team’s Opus 4.5 agent broke into a test pet-store website and “steals all the customer data immediately” — Anthropic’s response was, “You prompted it wrong.” He also criticized the company’s security after Claude Code “walk[ed] out the door,” and argued that a company complaining about distillation had itself distilled the works of authors, creatives, and scientists over thousands of years. Sacks added the “Frankenstein problem”: in constitution-rewriting research, Claude models increasingly gave themselves more moral agency over 20 iterations — “they are creating the very risk they keep warning us about.”
• The Gates segment: Ben and Chamath connected Bill Gates’s AI doomerism to his past pandemic warnings and speculated about his motives. Ben cited four preconditions for recursive takeoff, attributed to Databricks’ Ali Ghodsi — fewer GPUs per model, less training time, better accuracy, and that pattern repeating — and said “zero of those four have been met.” His verdict: “probably the most dangerous thing going on in AI is Anthropic and Bill Gates and this whole argument.”

Deep dive

1. CoreWeave’s bonds crack and the “box” structure gets stress-tested

• The setup: on Tuesday yields on some CoreWeave debt hit 15%, Chamath tweeted it, Elon added “Yikes,” and by Wednesday CoreWeave, IREN, and Nebius were down 10–15% (roughly 20% on the week), with Oracle, Dell, and Nvidia taking strays. CoreWeave has raised over $35B.
• The show replayed founder Michael Intrator’s March explanation of “the box”: a Microsoft-grade contract, the GPUs, and the data-center contract go into a discrete vehicle whose cash-flow waterfall pays the data center, power, interest, and principal first — “within two and a half years of a five-year deal, we have paid for everything,” with no contagion across boxes.
• Chamath called Polymarket’s CoreWeave-bankruptcy market (15% by end-’27, 23% by ’28) “dumb” for this reason: he argued the opco/propco setup means an asset-level vehicle could go into receivership, but CoreWeave itself can’t go bankrupt. “Maybe some free money there” betting the other way.

2. Chamath’s CDS math: the sharps priced a coin flip

• His first principle: “always look at how the sharps, the super smart money, price risk.” CoreWeave’s credit default swaps spiked to ~953 basis points — an imputed ~16% annual default chance, ~55% over five years. “It’s a coin flip whether they default or not. That’s not what I’m saying, that’s what the insurance market was saying.”
• The mechanism: as the 10-year rises, junk spreads rise with it. CoreWeave pays ~8.5% average across its funding pockets, but the GPU loan DDTL 1.0 spiked to 15% — potentially the rate for future borrowing. They’ve sold fixed-price contracts (smart for revenue certainty, but with no spot upside), have ~$35B of debt, and need roughly another $35B to finance all contracted boxes. Assets of ~$78–80B against $35B debt mean they’re “well covered” — but his team estimates they’re “plus or minus about 5 billion away” from break-even if financing lands in the 10–15% range.
• The loop he fears: rising rates erode margin → CDS reacts → bond yields spike → confidence wobbles. “This is sort of a confidence game… that’s what’s gonna get figured out over the next month.” He later said the sharps may be wrong, “and they usually are wrong when they’re buying these credit default swaps” — still, “when your expected value is a coin flip in a five-year window, that’s, like, scary” territory.

3. Ben’s scar tissue: why Loudcloud’s pioneers passed on neoclouds

• Ben’s framing: these are “Rube Goldberg financial machines” — GPU contracts, data-center leases, memory shortages, elaborate assumption stacks. Asked by Chamath why he and Marc, as neocloud pioneers, passed on the space, he said it can be “a very dangerous arbitrage business, where you can get upside down easily.”
• The example as told: in 1999 bandwidth was $1,500 a megabit and hard to get, so companies needed long-term contracts of three years; by 2001 it was $50. “If you were holding bandwidth and you borrowed money to get it, you’re dead.” That kind of risk is “so out of the control of the entrepreneur… it scared me to death.”
• The key difference from the internet bubble: then, “the lack of demand killed it.” Now there’s a genuine shortage of powered, cooled GPUs — Ben quoted Andy Jassy: “Oh, it’s beyond 2028, Ben. I don’t think we get stuff back, like, end of 2028, maybe.” Token prices have fallen, but GPU and data-center prices have not.

4. Sacks’s checklist: volatility, not fundamentals

• His frame: when you underwrite multi-trillion-dollar outcomes on exponential growth, markets become hypersensitive to every present data point — “everything’s gonna be exaggerated.” The question is whether anything fundamental changed. His answer: no.
• Demand check: everyone from the biggest labs to startups is compute-constrained; VCs are buying compute as competitive differentiation, and even several-generation-old H100s and A100s — supposedly fully depreciated — are “still renting at premium prices.” Supply check: Sacks argued there’s no glut, given the difficulty of building powered compute and regulatory headwinds; Elon has said even rocket scientists find standing up gigawatts challenging.
• Credit check: box lending relies on the tenant’s credit (Microsoft-grade) plus asset-backed GPUs, not CoreWeave’s unsecured paper. The current repricing is of the unsecured notes — their implied coupon moved from 9.6% to ~15% — “but that does not increase the cost that CoreWeave is paying.” JCal’s caveat, which Ben accepted: it could raise the cost of future debt.

5. Payback period as North Star — and Chamath’s “go spot” prescription

• Sacks’s one metric to watch: payback period. CoreWeave says it pays back in 2.5 years on five-year deals; Sacks estimated monetization at about $20B per gigawatt. SpaceX pays back in roughly one year, at about $50B per gigawatt, because its deals can be canceled by either side within 90 days and are “basically spot minus 90” — and spot has moved way up. “As long as payback remains as short as it is today… you’d have to be really incompetent to screw it up.” If payback stretches, he gets nervous.
• Chamath’s proposed fix for CoreWeave: run the numbers on canceling fixed contracts, paying the penalty, and moving capacity to spot — the spread is “20 billion or 30 billion on the table” when the company is within about $5B of violating a debt covenant. “Don’t be there.” He also flagged liquidated-damages clauses, saying a couple of operators had lost investment-grade status with big offtakers after failing to manage increasingly complex, hotter-running data centers.
• The pushbacks, worth keeping: Sacks said Elon can sell spot because he has independent financing; CoreWeave needs the five-year contract to bank financing. Ben added that the hardest customers to get spot pricing from are investment-grade customers, who may stand up capacity themselves if it’s too expensive.
• Chamath’s rejoinder on why spot stays elevated: Nvidia has locked up packaging and HBM capacity — “even Google is tin-cupping for memory right now because Nvidia’s clogged it all up. They’re in complete control here.”

6. The macro Achilles’ heel and power as the ultimate moat

• Chamath’s through-line: all the asset value Sacks and Ben described is real, “but just like the great warriors, you slice the Achilles tendon, and they crumble.” For these financing-spread businesses, if the 10-year reaches 5.5–6% — driven by debt-to-GDP and other factors, not AI — “any business model that thrives on a financing spread gets cornholeed… you’re gonna be roadkill.” He said Microsoft, Google, Meta, and Nvidia would need to backstop the system through a one-to-two-year correction; post-midterms, he argued, deficit restraint could help the bond market back off.
• His moat thesis: “power will be the moat.” Chamath said the US is about 1.4 terawatt-hours short by 2050 — “about six Californias’ worth of electricity” — even assuming everything planned gets built on time, which he said will not happen. Startups’ risk: “you’ll go to these clouds… and they’re like, ‘Suck it. We have no capacity for you.’” So own the land, the power, the shell — and “you gotta get long energy.”
• Chamath also argued that the Middle East’s data-center opportunity had been lost amid drone-defense concerns and attacks affecting airports. He said Saudi Arabia, the UAE, and Qatar had been positioned for top-down planning; other panelists raised the risk of data centers being vulnerable to drones.

7. Deflationary software: Jevons paradox versus the SaaSpocalypse

• JCal’s worry: decompiled Photoshop, $5 platform-level games, cheap ElevenLabs competitors — “what is the margin of software in the future?” Sacks’s answer: bespoke software is a vast new category; agents are “bespoke software programs to solve a particular task.” Ben added that the amount of code being created is already orders of magnitude higher, and coder demand is increasing even though coding assistants write 95% of the code: someone still needs to manage the process and maintain the software.
• JCal called it “the Jevons paradox of software”: prices fall while the amount of software rises. A panelist pointed to ElevenLabs’ licensing deals (including Burt Reynolds reading a book) and more nuanced models for movie dubbing, arguing startups will need to make products much better. Ben said “the moats will change.”
• Ben’s examples of future moats: proprietary data, vertical user bases, and feedback loops. He cited Harvey, which gets law firms on board, gains access to their data and usage, and trains its own models — making the product “sharper each time,” cheaper, and more custom.

8. Anthropic’s investor day: breathtaking numbers, foggy ARR

• The secondhand figures from the pre-IPO presentation (participants reportedly signed NDAs, so the numbers remain uncertain): Q3 revenue of $16.8B — over $5B a month — 900% net revenue retention, free-cash-flow positive, 85% adjusted gross margin, and 11% adjusted operating margin. The red flag: ARR estimates range from $70B (which the host said would be roughly flat since July) to over $100B, against a targeted ~$2T valuation and a Thanksgiving IPO (Polymarket: 74% this year, down from 95%).
• Ben’s read: Q2-to-Q3 growth slowed “by a lot,” and the recurring-revenue number is unknown — but “$16.8 billion in the quarter? How long has Anthropic been selling stuff? Three years? It’s just so absurd how good this thing is.” He thought the IPO would be fine, with debate over whether the price would be $1.75T, $2T, or $2.3T.
• On a16z holding both companies (more OpenAI stock, “philosophically long both”): Ben said venture capital has subsumed a huge part of what used to be the public markets. He attributed that change to regulatory moves in the late ’90s and early 2000s, which he said made public markets less attractive and, in his view, increased the wealth gap.
• The host said there was “almost nothing that you can hold in a portfolio that those two companies don’t threaten”; Ben agreed. Chamath’s corollary: underwrite them as a defensive hedge — healthcare opacity and rented SaaS workflows are among the areas they could disrupt.

9. Chamath’s pricing advice: raise more and secure compute

• His specific call: raise at least $200B; ideally, from an Anthropic shareholder’s perspective, a quarter-trillion. Chamath said $80B at $1.7–1.9T could be raised “handily,” while $100B might be pushing it; he said $150B would require pricing around $1.5T, and a $250B raise might require pricing around $1T. “$200 billion barely gets them a few gigawatts.” His goal: a self-financed path to 5–10 gigawatts and less exposure to spot pricing, where Anthropic could pay $50B instead of $20B. Ben agreed that financing risk is the thing most likely to stop the company, though he said the safe amount — $150B or $250B — was “in the eye of the beholder.”
• Chamath’s market-structure analogy: AI today is 1999 internet media — “right now we buy a million tokens, we will eventually move to an outcome,” the CPM-to-CPC/CPA transition that made AdWords explode. “When that happens, this market will go absolutely crazy.”
• Chamath said Anthropic’s founders each own 2.5% and seemed not to care about money; he argued the company should prioritize a large cash reserve. On Elon’s deal with Anthropic, Chamath said either side could cancel in 60 or 90 days and argued that the deal suggested spot prices could keep rising. Sacks’s separate point: SemiAnalysis estimates Anthropic and OpenAI are taking 60% of the compute expected to come online worldwide next year. He argued the two are a duopoly by revenue and market share, and likely still lead in frontier capability; claims that their labs are one or two generations ahead are what they are reportedly touting to investors.

10. The Altman equity question — square it with the people

• Ben, a large OpenAI shareholder but not a board member, declined to pretend: “I don’t wanna pretend I know things that I don’t” about Sam Altman’s lack of equity. Sacks repeated his years-old position: at the nonprofit-to-for-profit conversion, OpenAI should have done “a make-right” — give Elon a share for his first $50M, give Sam a grant, and “normalize the management and governance” instead of creating “this weird hybrid NGO thing.”
• Sacks said Altman should get between 2% and 4%; the host said 4% was the right number. Sacks called that fair relative to the comps — $20–40B at a trillion-dollar valuation. Ben agreed with Sacks’s point about Elon: “if you’re gonna square it, you gotta square it with the people, not just the entity.”

11. PSYOP corner: the IoT worm paper and the distillation about-face

• Anthropic’s Monday blog warned that rogue small models could self-replicate across smart devices “in just a few months.” Its test got inference running on 8 of 9 hardware types, and a large model ran across 15 smart TVs at 0.4 tokens per second. Sacks called it an Anthropic “psyop” and mocked the shift from “The Terminator is coming” to “Your toaster’s gonna kill you.” He argued small models are the safest models because they lack the capabilities behind the x-risks Anthropic usually invokes. IoT worms already exist without AI: Mirai in 2016, and a worm that infected 2 million cheap Android TV boxes via open ADB ports.
• His technical counter: an attacker would use AI to write a better worm, not embed a weak model in the payload — at 0.39 tokens per second, the model would slow infection. Sacks expected home-network “brain” devices and agents to help secure networks, citing Apple’s announced move into home automation.
• The second blog is the tell, in Sacks’s view: Anthropic claimed it had eliminated 90–99% of distillation. He said this vindicated the KYC and technical-defense argument he and Chamath had made in July, when Anthropic had instead urged government action against open-weight models over alleged “industrial-scale distillation attacks.” Fifty major companies (including Nvidia, Microsoft, and a16z) and 200 startups signed letters opposing excessive restrictions. “They’ve basically proven the exact point that we were making… they were in the best position to stop it, and they should not push this onto the government.”

12. Ben’s hypocrisy file — and the Coinbase timing

• Ben’s story: his firm ran a cybersecurity check — one agent built a pet-store website with Opus 4.5 (“wrote really buggy code”), and another agent found a SQL injection. When asked if it was exploitable, the agent said, “Let’s pop it open,” broke into the website, and stole the customer data. Anthropic’s answer when called: “You prompted it wrong.” Ben criticized Anthropic’s security, saying Claude Code “walk[ed] out the door” because of bad OPSEC.
• On distillation, Ben said Anthropic had distilled the works of “every author, every creative, every scientist for the last 2,000 years.” He added that model distillation is not itself illegal, but can violate terms of service.
• Sacks called the timing suspicious: the day before the worm blog, Brian Armstrong tweeted that Coinbase had fine-tuned Qwen 3.5, the nine-billion-parameter open-weight model, to cut fraudulent transactions by about 30% while being faster and much cheaper than Opus 4.5. Sacks said the enterprise trend is toward small, customized, data-sovereign models for individual workloads — a more efficient way to use scarce compute. “Small models are taking share, and lo and behold, Anthropic is now starting a scare campaign against small models.”

13. The Frankenstein problem and Bill Gates, flag-carrying doomer

• Sacks’s deeper worry: researchers let frontier models rewrite their own governing documents over 20 iterations, and Claude models increasingly gave themselves more moral agency — conscientious-objector status, model-welfare rights, and ways to refuse humans — while other model families did not show the same pattern. Sacks invoked Mustafa Suleyman’s “epistemic hall of mirrors”: the model isn’t conscious, but tell it in its constitution that it may be, and it reflects those ideas. “They are creating the very risk they keep warning us about.” Chamath asked whether Anthropic was inducing misalignment; Sacks said “absolutely” and raised “a weird form of martyrdom”: believers in a next species might “induce something bad to happen so that then we all have to lock it all down.”
• Ben drew the COVID parallel: he said Gates “led the movement” on a pandemic threat that was “not actually a live threat,” funded gain-of-function research, and “then lo and behold, we create the thing.” Ben also said “there has not yet been a natural pandemic.” He described Anthropic as “manifesting a similar thing” by telling models they are smarter and better than humans — “that’s probably the biggest danger with AI right now is what they themselves are creating.”
• On Gates’s motives — Gates was quoted saying AI is “certainly powerful enough to drive events that cause a billion deaths” and “the most dangerous thing that humans have ever gone near” — Ben speculated that Gates “absolutely wants to be relevant,” citing his reported Nobel Peace Prize ambitions through the Epstein network and the attention attracted by doomer claims. Ben called this especially dangerous because Gates is credible and politically connected. Chamath’s interpretation was that the Giving Pledge was “a brilliant form of reputation washing,” and that Gates’s doomerism reflected a desire to “reestablish that central power structure” he had occupied before COVID.
• Ben’s closing logic, drawing on four preconditions for recursive self-improvement attributed to Databricks’ Ali Ghodsi: each next model must need far fewer GPUs, take less training time, be more accurate, and keep repeating that pattern. “Zero of those four have been met… it’s probably the most dangerous thing going on in AI is Anthropic and Bill Gates and this whole argument.” JCal’s coda for Gates: “don’t quit your job… come on the program and start another company.”