Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487]

What It Covers

Ben Thompson joins Patrick O'Shaughnessy to analyze AI's geopolitical stakes, capital structure risks, and competitive dynamics across Big Tech. Thompson applies aggregation theory to the AI era, examines TSMC's capacity constraints, evaluates OpenAI, Anthropic, Meta, Google, and Amazon's strategic positions, and warns that AI investment may face a railroad-era capital gap before revenue catches up.

Key Questions Answered

• •AI Capital Stack Risk: AI investment has burned through free cash flow, then debt markets, and now equity issuance — Google issuing equity being the signal. The next funding source is pension funds and insurance floats via vehicles like NVIDIA's $500B structure. If revenue doesn't scale fast enough to return to free cash flow funding, a railroad-era capital gap could trigger a significant blowup before AI's economic impact fully materializes.
• •Inference Cost Stratification: Treat AI inference costs as a spectrum, not a single number. A user asking basic recipe questions costs near zero to serve — comparable to loading a webpage. A researcher running extended reasoning tasks costs orders of magnitude more, as every additional second of compute is direct marginal cost. Microsoft's shift to usage-based enterprise pricing above $100/user/month reflects this reality and risks breaking the "thoughtless revenue" headcount-linked billing model.
• •Consumer Monetization Error: OpenAI replicated Dropbox's mistake at 100x scale — building a subscription consumer product when advertising was the correct model. Consumers don't pay for software and don't want productivity tools. Had OpenAI launched an ad-supported model immediately after ChatGPT's breakout, they would now have a mature ad flywheel, putting Google and Meta under far greater pressure. They are pivoting to advertising now, but years late.
• •TSMC Risk Transfer Mechanics: TSMC deliberately constrained capacity growth in 2023–2025 to avoid overcapacity risk, transferring that risk to hyperscalers as foregone revenue. Risk doesn't disappear — it relocates. The resulting compute scarcity is now so acute that Amazon and Google are economically incentivized to bring Intel and Samsung logic fabs up to speed, providing geopolitical diversification away from TSMC as a byproduct of pure profit motive, not policy.
• •Meta's Advertising Structural Advantage: Meta's ad marketplace functions as a verification machine for AI-generated content at global scale — AB testing billions of ad variations with human purchase decisions as ground truth. Moving from embedding-based ad matching to LLM-based predictive ad placement requires only a few percentage points of conversion improvement to generate billions in incremental revenue. Meta pays zero for user-generated content, giving it structurally better margins than YouTube's revenue-share model.
• •Amazon's First-Best-Customer Flywheel: Amazon's compounding advantage comes from using its own retail operation as the launch customer for every new technology — AWS, Graviton chips, Trainium, logistics, and now AI call center software. Internal scale makes mediocre early products viable long enough to improve. Trainium now powers Anthropic. This model means Amazon's core physical retail and logistics business is largely insulated from AI disruption while simultaneously generating the demand signal that makes its AI infrastructure competitive.