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

What It Covers

Chamath, Sacks, Jason, and guest Ben Horowitz analyze CoreWeave's bond market stress — with yields spiking to 15% and credit default swaps implying a 55% five-year default probability — alongside Anthropic's IPO signals, AI sector deflation risks, Anthropic's regulatory fear campaigns against open-source models, and Bill Gates's AI doomerism motivations.

Key Questions Answered

• •CoreWeave Debt Structure: CoreWeave's "box" financing model isolates each GPU deployment contract — pairing a customer commitment (e.g., Microsoft) with asset-backed debt — so corporate-level bankruptcy is structurally unlikely. However, the company still needs to raise an additional $35B to finance remaining contracted capacity, and rising risk-free rates are pushing borrowing costs from 8.5% toward 15%, compressing margins to within roughly $5B of debt covenant violations.
• •Payback Period as North Star Metric: For neo-cloud infrastructure businesses, payback period is the single most predictive risk metric. CoreWeave's current payback is 2.5 years on 5-year contracts; SpaceX's xAI cluster achieves roughly 1 year by selling at spot pricing (~$50B per gigawatt vs. CoreWeave's ~$20B). Investors should monitor payback period elongation as the earliest warning signal of deteriorating unit economics before revenue or EBITDA metrics reflect the problem.
• •Spot vs. Fixed Contract Arbitrage: CoreWeave locked in fixed-price, long-term GPU contracts while spot GPU pricing has surged dramatically due to scarcity. This creates an estimated $20–30B revenue gap between contracted rates and current spot market value. Chamath argues CoreWeave should pay contract termination penalties now, convert available capacity to spot pricing, and capture that spread — potentially adding billions directly to the bottom line and reducing covenant breach risk.
• •Anthropic IPO Sizing Strategy: Anthropic is targeting a $2T IPO valuation to raise roughly $80B. Chamath argues this is structurally insufficient: the company needs $200–250B in capital to self-finance 5–10 gigawatts of compute through a potential 2–3 year market correction, avoiding dependence on expensive spot pricing (which runs ~$50B per gigawatt vs. ~$20B via structured financing). Raising less at a higher valuation trades short-term optics for existential financing risk.
• •AI Bespoke Software Expansion (Jevons Paradox): Coding assistants are not eliminating software developer demand — they are expanding the total addressable market for software itself. Enterprises previously unable to justify custom software development are now building bespoke tools, vertical agents, and internal applications. Demand for developers who can manage AI-generated code is rising even as 95% of code is AI-written. The net effect is deflationary on per-unit software pricing but expansive on total software volume and developer employment.
• •Anthropic's Open-Source Fear Campaign Pattern: Anthropic has run two coordinated regulatory pressure campaigns in 2025: first claiming Chinese model distillation constituted IP theft requiring government intervention, then publishing research suggesting small AI models could self-replicate across IoT devices. In both cases, Anthropic's own technical fixes (KYC-based distillation detection, eliminating 90–99% of distillation) proved the government intervention unnecessary. Investors and policymakers should treat Anthropic safety blog posts timed near competitive or regulatory inflection points with skepticism.
• •Energy as the Defining AI Moat: The binding constraint on AI infrastructure through 2028 and beyond is not GPUs, capital, or talent — it is power. Amazon's Andy Jassy confirmed compute availability is sold out past 2028. The US faces a 1.4 terawatt-hour electricity deficit by 2050 even assuming all planned generation comes online on schedule. Startups and enterprises that secure land, power contracts, and physical data center shells now will hold a structural competitive advantage that capital alone cannot replicate later.