All-In PodcastCoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer
CHAPTERS
- 0:00 – 2:17
Ben Horowitz subs in: #1 podcast banter and setting the agenda
The hosts celebrate hitting #1 on Apple Podcasts and welcome Ben Horowitz as a fill-in for David Friedberg. They preview a packed AI-focused docket spanning CoreWeave’s debt, AI deflation dynamics, Anthropic’s IPO, and doomer narratives.
- •All-In hits #1; playful credit-taking among hosts
- •Ben Horowitz joins as guest bestie
- •Friedberg is away; show tees up a heavy AI/markets episode
- 2:17 – 3:54
CoreWeave bond selloff sparks “AI bubble” chatter across neo-clouds
Jason introduces the market shock: CoreWeave debt yields spike and neo-cloud equities (CoreWeave, AIREN, Nebius) drop ~20%. The segment frames the key question—normal volatility or a canary in the coal mine for the AI trade.
- •CoreWeave bond yields jump (15% referenced) and stocks sell off
- •Elon’s “Yikes” amplifies headlines and sentiment
- •Contagion hits adjacent names (Oracle, Dell, Nvidia)
- •Debate framed: localized credit repricing vs broader AI bubble risk
- 3:54 – 6:04
Inside CoreWeave’s ‘box’ financing model (asset-backed cashflow waterfalls)
A replayed clip from CoreWeave founder Michael Intrator explains how contracts, GPUs, and datacenter inputs get packaged into isolated “boxes” with waterfall payments. The structure is pitched as non-contagious, self-amortizing project finance that improves borrowing rates over time.
- •Customer contract + GPUs + datacenter contracts placed into a ‘box’
- •Customer pays the box; box pays power/datacenter/debt before equity flows out
- •Claimed payback: ~2.5 years on a 5-year deal
- •Boxes are intended to be independent to prevent cross-contagion
- 6:04 – 13:43
Chamath’s credit-market lens: CDS spikes, fixed-price contracts, and refinancing math
Chamath argues the ‘smart money’ signal is in credit default swaps and secondary debt pricing, not equities. He outlines how rising risk-free rates, fixed-price revenue contracts, customer concentration, and a large incremental debt need create a reflexive confidence loop.
- •CDS implies sharply higher default risk (market-implied)
- •Risk-free rate rise pressures junk debt costs broadly, dragging CoreWeave financing higher
- •CoreWeave’s fixed-price contracts limit upside if spot pricing rises
- •Large incremental funding need (another ~$35B) collides with higher borrowing costs
- •Customer concentration: ~75% revenue from three customers magnifies perceived risk
- 13:43 – 20:11
Ben’s bubble comparison: why neo-cloud arbitrage can break (LoudCloud scars)
Ben calls neo-clouds “Rube Goldberg financial machines” with many moving parts (power, cooling, memory, build complexity). He contrasts today’s AI compute shortage with the internet bubble’s demand shortfall, while explaining why asset-price/contract arbitrage can still implode.
- •Neo-clouds depend on complex physical supply chains beyond GPUs
- •AI differs from the dot-com bandwidth glut: demand is extremely strong today
- •LoudCloud lesson: input costs can collapse fast (bandwidth $1500 → $50) and kill leveraged operators
- •Why a16z avoids neo-cloud investing: downside from uncontrollable pricing/financing dynamics
- 20:11 – 31:14
Sacks’ fundamentals checklist: demand, supply constraints, and unsecured vs asset-backed debt
Sacks argues CoreWeave’s move looks like exaggerated volatility rather than a fundamental break. He walks through demand strength, supply scarcity, execution risk, and why secondary pricing of unsecured notes doesn’t directly raise current coupons—though it may affect future financing.
- •Demand remains ‘compute constrained’ across startups and major labs
- •Supply is hard to bring online (construction, labor, politics, regulation)
- •Asset-backed project finance depends on tenant credit + GPU collateral
- •Unsecured bond repricing changes market perception and future cost of capital more than current payments
- •Metric to watch emerges: are economics (and risk) actually deteriorating?
- 31:14 – 38:19
AI deflation & ‘decompiled software’: Jevons paradox, moats, and what survives
Jason raises the fear that software margins collapse as models clone features and competitors ship instantly. Sacks and Ben counter with a Jevons-paradox view: lower cost creates vastly more software (especially bespoke/agentic software), shifting moats toward data, workflows, and feedback loops.
- •Deflation examples: cheap game creation, commoditized voice, rapid open-source competition
- •Sacks: bespoke enterprise software and agents explode total software output
- •Coding jobs may rise despite assistants writing most code (maintenance/orchestration still needed)
- •Ben: franchises survive by evolving beyond ‘just software’ (distribution, licensing, nuanced training)
- •Moats shift: proprietary data + user feedback loops (e.g., vertical AI like legal)
- 38:19 – 45:40
The real constraint: power, land, and datacenter capacity as the ultimate moat
Chamath argues energy and datacenter capacity—not models—become the limiting reagent. The group discusses an impending “electron deficit,” hyperscaler capacity limits, geopolitics as a datacenter risk, and why payback period and liquidated damages matter operationally.
- •Chamath: power is the ultimate moat; US faces major electricity shortfall by 2050
- •Enterprises may be blocked from launching products due to cloud capacity constraints
- •Middle East datacenter opportunity undermined by security/drone threats
- •Sacks: payback period is the North Star metric (CoreWeave ~2.5 years cited)
- •Chamath: liquidated damages + operational complexity (hotter chips, harder cooling) increase risk
- 45:40 – 53:39
Anthropic investor day leaks: massive revenue, questions on growth, and IPO setup
The conversation shifts to Anthropic’s investor day and mixed market reactions. Reported revenue figures are enormous for a young company, but uncertainty around ARR and growth deceleration sparks debate about valuation, open-source headwinds, and what the IPO means for the broader AI trade.
- •Anthropic IPO timing chatter; Polymarket odds referenced
- •Reported Q3 revenue presented as astonishing; retention figures cited as extreme
- •ARR estimates disputed (70B–100B range mentioned) and growth slowdown becomes central concern
- •Valuation target discussed as unprecedented (2T floated)
- •Sacks: investor-day sentiment seems positive, but numbers will be known at filing
- 53:39 – 1:03:36
Chamath’s IPO playbook: raise a ‘war chest’ now (200–250B) vs maximize valuation
Chamath argues Anthropic should prioritize financing resilience—compute security and survivability through a downturn—over headline valuation. Ben largely agrees with the ‘avoid financing risk’ philosophy, debating what “safe” cash levels really are given gigawatt-scale capex.
- •Chamath: raise $200B minimum (prefer ~$250B) to fund compute and weather corrections
- •Logic: compute scarcity + spot pricing volatility make future funding riskier
- •Tradeoff: lower valuation may be needed to clear that fundraising size
- •Ben: financing risk is a major failure mode; better to remove it while markets are open
- •Discussion ties back to CoreWeave-style project finance and offtake dynamics
- 1:03:36 – 1:24:22
Psyop Corner: Anthropic’s ‘self-replicating IoT model worms’ and distillation reversal
Sacks frames Anthropic’s safety posts as regulatory-capture “panic campaigns,” especially against open-source and small models. They dissect the IoT worm scenario as speculative and note existing non-AI worms already exploit insecure devices; the distillation post is cited as proof Anthropic could solve the issue without government intervention.
- •Anthropic blog warns about models propagating to IoT devices; panel calls it fearmongering
- •Sacks: embedding a model in a worm payload is slow/inefficient vs using AI at design time
- •Existing IoT botnets (e.g., Mirai-style patterns) show the core problem is device security
- •Distillation controversy: Anthropic now claims big improvements after previously pushing policy crackdowns
- •Argument: enterprises increasingly want small, fine-tuned, sovereign models; Anthropic reacts defensively
- 1:24:22 – 1:30:59
Bill Gates ‘AI doomer’ narrative: relevance, power, and the risks of manufactured panic
The panel debates Gates’ warnings that AI could cause catastrophic death tolls, tying the rhetoric to incentives around influence and institutional control. Ben critiques “theoretical runaway” scenarios as lacking prerequisites today, while the group argues panic narratives can be socially and politically destabilizing.
- •Motivations discussed: relevance, credibility leverage, political influence, legacy-building
- •Parallels drawn to pandemic-era ‘manufacture the risk’ dynamics (gain-of-function analogy)
- •Ben cites prerequisites for recursive self-improvement (fewer GPUs, faster training, better accuracy, repeatability) as unmet
- •Argument that sensational narratives fit media incentives (jobs, environment, extinction)
- •Call for builders to focus on practical progress rather than doomer status games
- 1:30:59 – 1:34:05
Closing riffs: don’t retire into doom—keep building
The episode winds down with a broader cultural critique: stepping away from building can lead to legacy obsession and fear-based influence campaigns. The hosts end with invitations (election-night livestream) and playful wrap-up banter.
- •Jason’s thesis: retirement/legacy-chasing fuels spirals and public fear narratives
- •Sacks: caring about elite approval makes you controllable by institutions/media
- •Suggestion: Gates should ‘start another company’ and engage constructively
- •Final banter and sign-off as the #1 podcast
