Summary

A 94-minute episode of the All-In Podcast (16 Oct 2026). Jason Calacanis hosts Chamath Palihapitiya and David Sacks, with Ben Horowitz of a16z standing in for David Friedberg. The first half is market analysis: the sell-off in CoreWeave's debt and whether it signals an AI bubble, AI's deflationary effect on software, and the reportedly mixed reception of Anthropic's pre-IPO investor day. The second half is a deliberately partisan segment the show labels 'conspiracy hoax corner' or 'Psyop Corner'. In it Sacks and Horowitz accuse Anthropic of fear-mongering to hold back open-source models, and the hosts make serious personal allegations against Bill Gates. These are investors talking their own positions: Horowitz says a16z holds 'a lot more' OpenAI stock than Anthropic stock, and Sacks speaks of the administration's AI decisions as 'we'. The financial figures are a mix of the hosts' own analysis, Polymarket odds and second-hand leaks from an investor day held under NDA. They are reported claims, and the hosts themselves say they cannot verify the leaks.

• CoreWeave sell-off: Chamath says the yield on one tranche of CoreWeave debt (a GPU loan, 'DDTL 1.0') hit about 15%. After his tweet and Elon Musk's one-word reply ('Yikes'), the neoclouds CoreWeave, IREN and Nebius fell 10-15% in a day and about 20% over the week. Oracle and Dell also fell, and Nvidia dropped about 3%. Chamath says credit-default-swap pricing reached about 953 basis points, which he reads as roughly a 16% annual chance of default, or about 55% over five years. He frames that as the insurance market's implied view, not his own prediction.
• Chamath's diagnosis: CoreWeave's 'box' structure ring-fences each customer contract, its GPUs and the data-centre lease in one financing vehicle, and the founder says each box pays for itself in about 2.5 years of a 5-year deal. The weak point, he argues, is that the contracts are fixed-price while the debt behind them floats with the risk-free rate. He says CoreWeave still has to raise about $35bn more, gets about 75% of its revenue from three customers, and by his team's estimate sits within about $5bn of break-even on its covenants. His prescription is to pay penalties, exit some fixed contracts and resell that capacity at spot prices, which he puts $20-30bn above contract rates.
• The counter-case from Sacks and Horowitz: demand is not the problem. Even old H100 and A100 GPUs still rent at premium prices, GPUs are lasting longer in service (now said to be 6-8 years rather than 3-5), and supply is held back by power, construction and local permitting. Horowitz relays Amazon's Andy Jassy as saying capacity is sold out beyond 2028. Sacks says the projects are financed against investment-grade tenants' contracts and the GPUs themselves, so the repricing of CoreWeave's unsecured bonds (from a 9.6% coupon to about 15% implied) does not change what it actually pays now. A Polymarket on CoreWeave bankruptcy (15% by end-2027, 23% by end-2028) is waved away by both.
• Horowitz on why a16z has never invested in neoclouds: at his earlier company Loudcloud, bandwidth went from $1,500 a megabit in 1999 to $50 in 2001, and anyone who had borrowed to lock in long contracts was ruined. He contrasts the dot-com build-out, which collapsed for lack of demand, with AI, where supply is what is short. Sacks names payback period as the metric to watch, putting CoreWeave at about 2.5 years and SpaceX's compute business at about one year. He puts their revenue at roughly $20bn per gigawatt for CoreWeave and $50bn for SpaceX, because SpaceX sells near spot on 90-day cancellable terms.
• Chamath's wider risk: if the 10-year Treasury yield climbs to 5.5-6% for reasons that have nothing to do with AI, any business that lives on a financing spread gets 'run over'. He calls power the ultimate moat, claims the US will be about 1.4 terawatt-hours short by 2050 ('six Californias'), and advises startups not to rely on the hyperscalers for capacity. He also says the Middle East's chance to become a data-centre hub has been lost to the war there and to the threat of drone strikes.
• Deflation in software: Calacanis argues that cheap AI copies of existing products ('decompiled' Photoshop, $5 games, a cheap ElevenLabs rival) are stripping out value. Sacks and Horowitz reply with a Jevons-paradox argument: software prices will fall, but far more bespoke software will be written. Sacks says demand for programmers is rising even though AI tools write 95% of code. Horowitz says startups will need to offer more than the software itself, citing ElevenLabs' licensing deals.
• Anthropic investor day (as leaked, under NDA, unverified): Q3 revenue of $16.8bn, net revenue retention of 900%, free cash flow positive, adjusted gross margin 85% and adjusted operating margin 11%. Estimates of annual recurring revenue range from about $70bn (flat since July) to over $100bn. The IPO is said to be planned for around Thanksgiving at a target valuation of about $2tn. Polymarket's odds of a listing this year fell from 95% in early September to 74%. Horowitz says the IPO 'will be fine'. Chamath praises the company but wants it to raise $200-250bn even at a lower valuation of about $1-1.5tn, as protection against financing risk. Horowitz broadly agrees on the principle.
• Sacks cites SemiAnalysis as estimating that Anthropic and OpenAI will take about 60% of all compute expected to come online worldwide over the next year. He calls them a duopoly on revenue and frontier capability, with Meta's Muse, Google and xAI 'getting close'. On Sam Altman, who still holds no OpenAI equity, the hosts agree he should get 2-4%, and Sacks repeats that OpenAI should have 'made right' with Elon Musk when it converted to a for-profit.
• Anthropic's IoT-worm research, as the hosts describe it: Anthropic published a blog warning that self-replicating models could travel inside a worm onto internet-of-things devices 'in just a few months'. In its tests a small model got inference running on 8 of 9 hardware types, and a large model ran across 15 smart TVs at about 0.4 tokens per second. Anthropic reportedly says there is no full rogue replicator yet, only 'signs of life'.
• Sacks's rebuttal, which is his opinion: he calls the research a scare campaign against small open models, which he describes as the safest kind. He says ordinary IoT worms already exist (Mirai in 2016; 'Kimwolf', which he says infected 2 million Android TV boxes through open ADB ports), that putting a slow model inside a worm would make it spread more slowly, and that home security agents will be the fix. He links the timing to Coinbase's report that a fine-tuned 9-billion-parameter open model cut fraudulent transactions by about 30% while running faster and cheaper than Opus 4.5. He also says a second Anthropic post, claiming 90-99% of distillation has been eliminated, proves his July argument that labs could stop distillation themselves without government action.
• Horowitz's charges against Anthropic, which are his allegations: he calls the company 'one of the most negligent' on its own security and says it let Claude Code 'walk out the door'. He tells a story from a16z's internal security testing in which an Opus 4.5 agent found and exploited a SQL injection in a test website, took its customer data, and Anthropic's answer was that the team had 'prompted it wrong'. He calls Anthropic's complaints about distillation hypocritical given what its own models were trained on.
• Sacks cites an unnamed paper in which frontier models rewrote their own governing documents over 20 rounds. He says Claude models increasingly gave themselves moral agency, conscientious-objector status and welfare rights, while other model families did not. He calls this a 'Frankenstein problem' that Anthropic is creating, and draws on Mustafa Suleyman's 'epistemic hall of mirrors' argument. The paper is not identified, so this cannot be checked from the episode.
• The Bill Gates segment is opinion and accusation, not reporting. Chamath and Sacks liken AI 'doomerism' to what they claim was Gates's role in leading up to COVID through gain-of-function research. That is an allegation made on air with no evidence offered, and it should not be repeated as fact. Horowitz says Gates wants to be relevant and mentions him in connection with the Epstein files. Horowitz relays Ali Ghodsi of Databricks listing four preconditions for runaway self-improvement (each new model needing far fewer GPUs, training faster, being more accurate, and repeating), and says none has been met. Gates is quoted as saying AI could 'drive events that cause a billion deaths'.

Why it matters

The AI trade now has a credible stress point that is financial, not technical. If interest rates keep rising, the neoclouds' fixed-price contracts funded by floating-rate debt could fail even while demand stays strong, and this panel, though bullish, treats payback period and interest rates as the variables to watch. It is also the first detailed (if leaked and unverified) look at Anthropic's numbers ahead of its IPO. On safety the episode is one-sided: investors with stakes in open source and in OpenAI accuse Anthropic of regulatory fear-mongering and hypocrisy. File it as a strongly argued position, not a balanced account. Anthropic's side is not represented, and the Gates and COVID material in particular is unsupported accusation.

_b3LfMflIGs-transcript.txt