All-In With Ben Horowitz: CoreWeave's Debt Spiral and the Anthropic IPO

Ben Horowitz joins the All-In Podcast to ask whether CoreWeave’s debt selloff means the AI trade is cracking, and what Anthropic’s IPO numbers really say. Sitting in for David Friedberg on the October 16, 2026 episode, the Andreessen Horowitz cofounder joins Chamath Palihapitiya, David Sacks and Jason Calacanis for a long session on neocloud financing, the deflationary pressure AI puts on software, the leaked numbers from Anthropic’s investor day, a “psyop corner” on Anthropic’s research into self-replicating models on smart devices, and a closing segment on Bill Gates as the most prominent AI doomer.

TLDW

CoreWeave’s bonds sold off after the yield on one of its GPU loans hit 15%, and the panel splits on whether that is a crack in the AI trade or volatility layered on top of real, sold-out demand. Everyone agrees the danger is financing, not demand: rising long-term rates squeeze any business that lives on a spread, and CoreWeave’s fixed-price contracts are leaving money on the table while spot compute prices climb. Anthropic’s reported $16.8 billion quarter impresses the whole table, but Chamath argues it should raise $200 billion or more at a lower valuation instead of chasing $2 trillion. The back half turns on Anthropic itself, with Sacks and Horowitz calling its IoT worm research and distillation campaign a fear-based push against open source, and the episode closes on Bill Gates and the case against a fast AI takeoff.

Thoughts

Payback period is the number that explains this whole episode. Sacks put it plainly: CoreWeave says it earns back the cost of a build in about 2.5 years on a typical five-year deal, while SpaceX gets its money back in roughly one, because SpaceX sells compute close to spot on contracts either side can cancel within 90 days and collects about $50 billion per gigawatt against CoreWeave’s roughly $20 billion. That flips the usual story about safety. Long fixed-price contracts were what made CoreWeave’s financing “boxes” bankable in the first place, and in a market where spot keeps rising they have become the thing leaving, in Chamath’s estimate, $20 to $30 billion on the table while the company sits within about $5 billion of a covenant problem. Horowitz’s Loudcloud story is the mirror image: in 1999 they locked in bandwidth at $1,500 a megabit and watched it fall to $50 by 2001. A fixed contract is a bet on the direction of prices, and whichever side guesses wrong while carrying debt is the side that gets hurt. The advantage of watching payback instead of bond yields is that it moves before the crisis does. If it starts stretching, then demand, pricing or execution is slipping, whichever one the market has not started talking about yet.

Chamath’s case for pricing Anthropic’s IPO to clear the market is the most contrarian idea in the episode and may be the most correct. Raising around $80 billion at close to $2 trillion is the vanity outcome. Raising $200 to $250 billion at something nearer $1 trillion is the survival outcome, because in his words $200 billion “barely gets them a few gigawatts”, and he sees more risk in the next few years than in today’s market. Horowitz, who says he nearly ran out of cash 17 times as a CEO, backed the instinct: the thing most likely to stop the company is financing risk, so why carry it. The argument connects straight back to the CoreWeave segment. If the cheapest compute is the compute you own outright and the scarce input is power, then an AI lab’s IPO is really a compute-financing event, and the right size gets worked backwards from gigawatts, not forwards from a headline valuation. The open question is whether founders who reportedly own around 2.5% each, and who the panel describes as strikingly uninterested in money, will pick the smaller number.

The fight over Anthropic’s self-replicating worm research was framed as a story about motives, but the stronger argument Sacks made was commercial. Coinbase fine-tuned Qwen 3.5, a 9 billion parameter open-weight model, to catch fraud and, by Brian Armstrong’s account, cut fraudulent transactions by about 30% while running faster and far cheaper than Opus 4.5. If that pattern spreads, the pressure on frontier labs is not that small models are dangerous but that a lot of enterprise work does not need a frontier model at all, and tight compute makes small specialists more attractive every quarter. The panel’s reading of the timing, a warning about small models landing the day after a small-model success story, is speculation about intent and deserves to be held loosely. The technical rebuttal is more persuasive: a worm that carries its own model at about 0.4 tokens per second spreads slower, not faster, and a real attacker would use AI to write the worm rather than ride inside it. Anthropic’s own post concedes there is no full rogue replicator yet, only “signs of life”, which leaves the real question as how much weight early-warning research should carry in policy debates before the threat exists.

The doomer segment is the weakest stretch of the show and also holds its most useful tool. The weak part is the claim that Bill Gates “induced” COVID by backing gain-of-function research, which the panel offers no evidence for and which goes well beyond anything publicly established about the pandemic’s origins. Treat it as opinion. The tool is the four preconditions Horowitz borrows from Databricks CEO Ali Ghodsi for a runaway intelligence explosion: each new model would need far fewer GPUs, less time to train, and better accuracy, and the cycle would have to keep repeating. Horowitz says none of the four has been met, and the point lands because the first hour of the episode was a detailed account of why they cannot be met right now: GPUs sold out past 2028, a power shortfall Chamath puts at several Californias, and data centers financed through structures where one loan sleeve just hit 15%. You do not need to share the panel’s view of anyone’s motives to see that the physical economy of AI, the debt, the electrons and the payback periods, is the strongest argument against a fast takeoff on the table today.

Key Takeaways

CoreWeave, IREN and Nebius fell roughly 20% in a week after the yield on CoreWeave’s GPU-backed DDTL 1.0 loan hit 15% and Elon Musk replied “Yikes.” Oracle, Dell and Nvidia slipped too.
Chamath says credit default swaps on CoreWeave spiked to about 953 basis points, which implies roughly a 16% chance of default in any year and about 55% over five years, a coin flip priced by the market’s sharpest money.
CoreWeave has about $35 billion of debt and needs to raise roughly another $35 billion to finance contracts already signed. Its average borrowing cost is about 8.5%, and 75% of revenue comes from three customers.
Sacks argues the selloff reprices CoreWeave’s unsecured bonds (implied coupon from 9.6% to around 15%) but not the asset-backed project loans, which rest on investment-grade tenants and the GPUs themselves.
Old H100 and A100 GPUs still rent at premium prices and GPU useful lives are stretching from three to five years toward six to eight, which undercuts the depreciation bear case.
Chamath’s real risk is the 10-year Treasury: at 5.5% to 6%, any business that lives on a financing spread “gets run over”, which makes deficits and bond markets the Achilles heel of the AI buildout.
Chamath calls power the ultimate moat and says startups may not be able to rely on hyperscalers for capacity, so owning land, power and the building shell becomes the advantage.
Reported Anthropic investor day numbers: $16.8 billion in Q3 revenue, 900% net revenue retention, positive free cash flow, 85% adjusted gross margin and 11% adjusted operating margin. ARR estimates range from $70 billion to over $100 billion.
Chamath wants Anthropic to raise $200 to $250 billion at around $1 trillion rather than $80 billion near $2 trillion. Horowitz agrees financing risk is the one most likely to stop the company.
Citing SemiAnalysis, the panel says Anthropic and OpenAI will take about 60% of the compute expected to come online worldwide next year, a compute and revenue duopoly.
Sacks calls Anthropic’s IoT worm post a scare campaign against small models, noting worms like Mirai and Kimwolf already infect smart devices and that embedding a model would slow a worm down.
Horowitz cites Ali Ghodsi’s four preconditions for a runaway AI takeoff (fewer GPUs, faster training, better accuracy, repeating) and says none of them has been met.

Chapters

2:18 CoreWeave’s Debt Selloff, “The Box” and the Coin-Flip Default Math

Calacanis opens on the neocloud selloff and replays CoreWeave founder Michael Intrator explaining “the box”: a customer contract, GPUs and a data center lease placed in one vehicle whose cash flow pays power, interest and principal before anything reaches CoreWeave, with each box isolated from the next. Chamath walks through why the market got nervous anyway. Default swap pricing implies about a 55% chance of default over five years, the risk-free rate keeps rising, half of CoreWeave’s contracted build is still unfinanced, and its contracts are fixed price. His team puts the company about $5 billion from break-even on its obligations, which feeds a reflexive loop between default swaps and bond yields.

13:43 Ben Horowitz on Loudcloud and Why a16z Skips Neoclouds

Horowitz calls neoclouds “Rube Goldberg financial machines” and says a16z has stayed out of them because he has lived through this before. At Loudcloud, bandwidth bought on three-year contracts at $1,500 a megabit in 1999 was worth $50 by 2001, a swing no entrepreneur controls. He still sees tailwinds for CoreWeave: GPU prices have held up even as token prices fell, data center construction is getting more expensive, and Andy Jassy told him Amazon is sold out beyond 2028. The dot-com bust lacked demand; AI does not.

20:11 Sacks Says Volatility, Chamath Says the 10-Year Is the Achilles Heel

Sacks works through the fundamentals one at a time: demand is still off the charts, there is no supply glut because powered capacity is hard to build and harder to permit, CoreWeave’s execution reputation is intact, and project loans are backed by tenant credit and GPUs. What moved was the price of its unsecured bonds. Horowitz adds that the Anthropic and OpenAI IPOs will find buyers because almost every other holding in a portfolio is threatened by them. Chamath agrees on the asset value but says the one real weakness is financing at a spread to a rising risk-free rate, and that deficits, not AI, will decide whether the party continues.

31:14 Decompiled Software, the Jevons Paradox and New Moats

Calacanis asks whether AI that can clone a product for a few dollars wrecks software margins and venture returns. Sacks says the bigger story is a new category of bespoke software for use cases that were never big enough to serve, and notes demand for coders is rising. Horowitz calls it the Jevons paradox of software and points to ElevenLabs adding licensing deals and better models on top of the code. Sacks names data and feedback loops as moats, with Harvey as the example, and Chamath answers that the ultimate moat will be power.

38:19 Power Is the Moat, Payback Is the Metric

Chamath says the US is heading for a large electricity deficit by 2050 and that banks and startups may find their clouds simply cannot add capacity. Sacks proposes payback period as the North Star metric: CoreWeave pays back in about 2.5 years on five-year deals, SpaceX in about one because it sells near spot. Chamath adds liquidated damages clauses as another risk and says CoreWeave should run the numbers on paying to exit some contracts and selling that capacity on spot, since Nvidia’s grip on packaging and memory should keep spot elevated.

45:44 Anthropic’s Investor Day Gets a Mixed Reaction

Calacanis runs the reported numbers from Anthropic’s investor day ahead of a planned IPO around Thanksgiving: $16.8 billion in Q3 revenue, 900% net revenue retention, positive free cash flow and a target valuation near $2 trillion, with Polymarket odds of a 2026 listing down to 74%. Horowitz says a16z holds more OpenAI but is long both, and that venture capital has swallowed what used to be the public markets. He reads the concern as slower growth from Q2 to Q3, while Sacks says investors at the event liked what they heard and calls Anthropic a strong business.

53:57 Price It to Clear: Chamath’s $200 Billion Case

Chamath calls the revenue breathtaking and predicts AI pricing will move from tokens to outcomes, the way online ads moved from CPM to cost per click. His specific advice is to raise at least $200 billion, ideally a quarter trillion, even if that means pricing near $1 trillion. Horowitz agrees financing risk is the one to remove. The panel then calls Anthropic and OpenAI a compute and revenue duopoly, and agrees Sam Altman should get roughly 2% to 4% of OpenAI.

63:45 Psyop Corner: Self-Replicating Models and Distillation

Calacanis summarizes Anthropic’s post on small models that could copy themselves onto smart TVs, cameras and robot vacuums: inference ran on eight of nine hardware types tested, and a large model ran across 15 smart TVs at about 0.4 tokens per second. Sacks calls it the latest panic against open source and links it to Anthropic’s new distillation post, which he says proves his July argument that labs should stop distillation themselves through know-your-customer checks. He points to Mirai and Kimwolf as proof IoT worms already exist and argues a model inside the payload would only slow one down.

72:49 Horowitz on Security, Sacks on Small Models

Horowitz criticizes Anthropic’s own security, citing the Claude Code leak and an a16z test in which an Opus 4.5 agent exploited a test web store and took its customer data. He also notes distillation breaks terms of service, not the law. Sacks ties the timing to Coinbase’s fine-tuned Qwen 3.5 fraud model and argues every enterprise workload is heading toward its own small model. Calacanis suggests Anthropic prove itself as a white hat on HackerOne, the way XBOW did by winning a $250,000 Google Chrome bounty.

79:05 The Claude Constitution Study and Bill Gates the Doomer

Sacks cites a study in which frontier models rewrote their own governing documents over 20 rounds, and only Claude models gave themselves steadily more moral agency, which he calls the risk Anthropic warns about. The panel then turns to Bill Gates, who has called AI “the most dangerous thing that humans have ever gone near”, accusing him of seeking relevance and central control and drawing a comparison to gain-of-function research. Horowitz lays out Ali Ghodsi’s four preconditions for a runaway takeoff, says none has been met, and Calacanis closes with his advice to Gates: don’t retire, keep building.

Notable Quotes

“And over a 5-year duration, what that math implies is about a 55% chance of a default. So, it’s a coin flip whether they default or not.”
Chamath Palihapitiya, reading CoreWeave’s credit default swap pricing

“The lack of demand killed it. We don’t really have a lack of demand in AI.”
Ben Horowitz, comparing today’s buildout to the dot-com bandwidth glut

“The one metric I would really keep an eye on is payback period. To me, that’s like the North Star metric here.”
David Sacks, on how to tell whether neocloud economics are breaking

“Right now we buy a million tokens. We will eventually move to an outcome.”
Chamath Palihapitiya, comparing AI pricing to the shift from CPM to cost per click

“The thing most likely to stop the company is financing risk and so why take it?”
Ben Horowitz, on how much Anthropic should raise in its IPO

“You know we’ve gone from the Terminator is coming to your toaster is going to kill you.”
David Sacks, on Anthropic’s research into self-replicating models on IoT devices

“Zero of those four have been met. We’re not on track for any of those four to be met unless something is going on the big labs that we don’t know about.”
Ben Horowitz, on Ali Ghodsi’s preconditions for a runaway AI takeoff

For the full back-and-forth, including the charts Chamath walks through on CoreWeave’s financing, watch the full episode of the All-In Podcast here.

All-In With Ben Horowitz: CoreWeave’s Debt Spiral and the Anthropic IPO