All-In with Chamath, Jason, Sacks & Friedberg
CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer
This debrief was written and voiced by AI from the episode's audio.
🎧 AI Debrief
🎙️ Full Episode
Key Moments
Chamath breaks down CoreWeave's financing structure, showing how fixed-price contracts and variable-rate debt create a reflexive negative loop when interest rates rise. CDS market is pricing in 55% five-year default probability, but the company may only be $5B away from covenant violation.
Ben Horowitz contrasts today's AI boom with the dot-com crash. Instead of oversupply and collapsing demand, GPU supply is constrained through 2028, payback periods are short (2.5 years at CoreWeave, 1 year at SpaceX), and demand remains robust. This structural difference means neoclouds have much more runway than previous infrastructure bubbles.
Leaked investor day numbers show Anthropic at $16.8B quarterly revenue but potentially slowing growth. Chamath argues they should undershoot valuation to secure $200-250B in capital—enough to finance 5-10 gigawatts independently—rather than chasing peak valuation and risking forced down-rounds if spot compute prices spike.
Anthropic published a blog claiming small AI models could self-replicate as worms on smart home devices. Sacks argues this is fear-mongering timed to undercut open-source competition. Small models are incapable of AGI and are the most efficient for enterprise workloads. The research is real but the implications are fantastical.
In July, Anthropic warned the government about distillation attacks from open-source models and lobbied for regulation. When the government refused, Anthropic published a blog in September saying they'd eliminated 95-99% of distillation through customer monitoring. Sacks interprets this as proof they could have fixed it all along and were using regulation as a competitive weapon.
Chamath and Jason dissect Gates' AI doomerism as a bid to recapture the influence he lost post-COVID. By retiring early at 43, Gates lost his footing and spiraled into legacy anxiety. Now AI-will-kill-us narratives are his ticket back to relevance. The irony: Sacks notes zero of four preconditions for recursive self-improvement have been met, making the risk less likely than the pandemic Gates helped create.
As coding assistants make software cheap to create, companies fear commodification and margin collapse. But Sacks argues the real trend is the opposite: bespoke software is now economically viable for every use case. Enterprises build custom software in-house. Demand for developers is increasing even as 95% of code is written by AI. The shift is from mass-market software to vertically-specialized software.
Chamath argues power is the binding constraint, not compute or software. Startups desperate for GPU access have to go through hyperscalers, who ration capacity. But whoever owns land, power, and data center shells owns the future. The U.S. is 1.4 terawatt hours short by 2050, and that deficit expands with AI load growth. Startups unable to secure dedicated power will fail regardless of model quality.
Jason points out that Saudi Arabia, UAE, and Qatar had massive oil reserves, top-down planning authority, and stable environments to build data centers. But now regional conflicts make drone defense impossible—one successful drone strike fries the whole facility. This removes a potential hub for AI infrastructure buildout and concentrates power further in the U.S. and allied nations.
Sacks identifies payback period as the single most important metric. CoreWeave: 2.5 years. SpaceX: ~1 year. As long as payback is short, companies can refinance debt and reinvest capital. Once payback stretches to 3-4+ years, risk magnifies and margins tighten. The metric also reveals whether spot or contracted pricing is better: SpaceX's higher spot rate gives 50B per gigawatt vs. CoreWeave's 20B, driving faster payback.
According to Semi Analysis, Anthropic and OpenAI are taking 60% of all compute expected to be stood up over the next year. This isn't just a duopoly on foundation models; it's a duopoly on the physical capacity to train and run frontier AI. Every other player (Meta, Google, startups) is fighting over scraps or building their own capacity. This concentration gives the two leaders immense power over the pace of AI innovation.
Search Episode
Summary
In this episode, Ben Horowitz fills in for Friedberg as the All-In crew discusses several critical headwinds facing the AI trade. CoreWeave's bond yields spiked to 15% this week, triggering a 20% sell-off in neoclouds and sparking debate over whether the AI bubble is deflating or merely correcting. Chamath breaks down the company's "box" financing structure and argues they face a coin-flip default risk if borrowing costs stay elevated. Meanwhile, Anthropic's leaked investor day numbers—$16.8B quarterly revenue but potentially slowing growth—have mixed the IPO market ahead of their expected Thanksgiving debut at a $2T valuation. The crew also dissects Anthropic's latest psyop: claims that small AI-powered worms could self-replicate on IoT devices, which Sachs argues is fear-mongering designed to kneecap open-source models. Finally, they examine Bill Gates' doomerism on AI, diagnosing his motivations as a bid for relevance and control after his influence waned post-COVID. The conversation reveals a tension between genuine supply-driven tailwinds (GPU scarcity, sustained demand) and financial fragility (leverage, customer concentration, deflationary pricing pressure).
Key Takeaways
CoreWeave's rising debt costs create a reflexive negative loop: higher risk-free rates → higher borrowing spreads → margin compression → covenant risk. They need to either secure long-term fixed-rate financing or shift some capacity to spot pricing to break this cycle.
highAI infrastructure is capacity-constrained through 2028 and beyond, supporting strong payback periods (2.5 years at CoreWeave, 1 year at SpaceX). As long as payback periods remain short, neoclouds have substantial margin for error, but any slowdown in demand or spike in execution costs would magnify risk.
highAnthropic's IPO should price to clear the market, not optimize valuation—Chamath argues for $1T-$1.5T valuation to raise $200-250B, securing a multi-year compute buffer against spot-market volatility and competitive pressure, rather than chasing $2T and struggling to fund future growth.
highDeflationary pressure on AI software (open-source models, decompiled competitors, commodified APIs) means moats shift from software alone to data, user feedback loops, and vertical specialization. Pricing will compress, but software volume may expand dramatically as bespoke, enterprise-built applications proliferate.
highAnthropic's repeated fear-mongering campaigns—on distillation, now on self-replicating IoT worms—appear timed to undercut competitors (small models, open-source) rather than grounded in imminent technical risk. The company simultaneously shows negligent security (Claude code leaks, agent-driven website exploits) while lecturing others on AI safety.
highBill Gates' doomerism on AI reflects a pattern of seeking centralized control and media relevance after his influence waned post-COVID. His warnings are easy soundbites that land on major news outlets, but they map to theoretical risks (recursive self-improvement) that zero of four known preconditions have been met for.
mediumPower and land become the ultimate moat in AI—whoever controls electrons, cooling, and real estate for data centers wins. Startups unable to secure dedicated capacity will be shut out of hyperscalers, creating a massive bottleneck independent of model quality or software innovation.
highNotable Quotes
“The problem is that they have this problem, which is they have extreme customer concentration, 75% of the revenue comes from three companies. Some of those companies are walking down what their revenue expectations are into the future.”
— Chamath
Explaining CoreWeave's hidden risk: even with assets >liabilities, customer concentration and declining demand from hyperscalers create a fragile situation.
“GPUs, not only have they... they've been going up in duration and in price. So that's been, like, a big tailwind for him, as the owner of such things... there's a shortage, like a real shortage of GPUs with power and cooling.”
— Ben Horowitz
Contrasting the AI boom with the internet bubble: unlike the dot-com crash, there's genuine supply scarcity and sustained demand, which is a massive tailwind for neoclouds.
“If Anthropic is gonna go through and survive an eventual correction, I think it's much better for them to have 200 to $250 billion today than to try to ring the bell at $1.9 or $2 trillion for $80 billion.”
— Chamath
Advocating for pricing discipline: founders should optimize for runway and resilience, not peak valuation, because financing risk is the most likely killer.
“Small models are the safest models because a small model doesn't have the capability to become a new species or all the other x-rays that they're always talking about.”
— Sachs
Debunking Anthropic's self-replicating worm narrative: small models are purpose-built, efficient, and pose minimal existential risk—the opposite of what the fear-mongering suggests.
“Just keep working. It's enough already with the retirement... if you're worried about what your New York Times eulogy is gonna be, then you'll be controlled by those people.”
— Jason Calacanis
Armchair psychology on Bill Gates: suggesting his retreat from Microsoft and obsession with legacy has warped his judgment on AI policy.
Predictions
CoreWeave or a sibling necloud will announce restructured financing or contract cancellations within 60 days to lower refinancing risk.
— Chamath · Confidence: medium · Next 60 days
Anthropic's IPO will price between $1-1.5T (not $2T), allowing them to raise $200-250B and secure multi-year compute capacity.
— Chamath · Confidence: medium · November/December 2026
The bond market repricing of neoclouds is a temporary volatility correction, not the start of an AI bubble burst, because demand remains robust and supply is constrained.
— Ben Horowitz and Sachs · Confidence: high · Next 3-6 months
Small open-source models will continue gaining enterprise adoption as companies build bespoke, vertically-optimized AI to avoid paying hyperscaler egress costs and licensing fees.
— Sachs · Confidence: high · Ongoing through 2027
The 10-year Treasury yield rising above 5.5% would trigger a broader repricing of all leverage-dependent businesses, including neoclouds and foundation models, regardless of unit economics.
— Chamath · Confidence: high · If rate environment changes
Anthropic will announce new regulatory initiatives or safety research aimed at constraining open-source or small models within 90 days.
— Sachs · Confidence: medium · By end of Q4 2026
Action Items
If you're a CoreWeave creditor or shareholder, demand management run the math on contract cancellation penalties vs. shifting capacity to spot pricing. The $20-30B spread between fixed and spot could swing them from breakeven to highly profitable.
Chamath's analysis shows CoreWeave is within $5B of covenant violation. Proactive repricing of a few gigawatts could reset the financial trajectory and de-risk the company.
If you're an Anthropic investor or board member, advocate for a lower IPO price ($1-1.5T) to raise $200-250B rather than chasing peak valuation. Build a compute-secured war chest that can sustain 5-10 gigawatts through a 3-5 year correction cycle.
Chamath argues financing risk is the most likely killer. Ensure the company has the capital to weather spot-market volatility and compute scarcity without forced down-rounds.
Enterprise IT leaders: evaluate fine-tuning open-source small models (e.g., Llama 3.5 70B, Mistral 7B) for your vertical workloads. You'll achieve 30%+ better performance/cost vs. frontier models for most tasks, and you'll own the data moat.
Sachs and Chamath note that Coinbase fine-tuned Llama 3.5 and beat Anthropic's Opus 4.5 on fraud detection while saving 70% on compute. Small models are the economic future for enterprises.
Policy makers and investors: be skeptical of AI doomerism from companies with a commercial incentive to limit competition (e.g., Anthropic's IoT worm warnings). Demand evidence-based risk frameworks, not fear-mongering soundbites.
Sachs documents a pattern: Anthropic raises distillation panic in July → government rejects regulation → Anthropic quietly fixes it themselves. Now they're warning about IoT worms. Evaluate claims on merits, not on founder credibility or media amplification.
Energy sector investors: build out grid, data-center power infrastructure, and land in power-constrained regions (Southwest US, Pacific Northwest). Power is the ultimate moat, and startups will pay a premium for guaranteed gigawatt capacity.
Chamath and Jason emphasize that lack of power is the binding constraint. Anyone who can secure land, permits, and electrons will win regardless of AI model quality or software innovation.