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

Episode Summary

Executive Summary: The conversation centers on AI as a massive but capital-intensive technology cycle with deep geopolitical, industrial, and business-model implications. The speaker argues the U.S. should avoid seeking total AI dominance over China, favors a fragile but favorable equilibrium, and worries more about timing, capital availability, and compute supply than about the technology’s eventual impact. He contrasts consumer subscription versus ad-driven monetization, highlights how commodity dynamics shape compute and chips, and sees winners in companies that can turn scale, distribution, and first-party demand into durable advantages.

Main Topics: U.S.-China AI race and geopolitical balance (Priority: 5/5): The speaker argues that U.S. total dominance in AI would be dangerous and destabilizing, and that the current state of frontier competition—where the U.S. leads and China remains months behind—is a better equilibrium. AI as a capital cycle and timing mismatch (Priority: 5/5): He emphasizes that the key risk is not whether AI matters, but whether the industry can fund the enormous buildout long enough for revenues to catch up with spending. Aggregation theory, zero marginal cost, and AI economics (Priority: 4/5): He revisits aggregation theory to explain that AI changes the economics of distribution, discovery, transaction costs, and inference pricing, but does not eliminate commodity constraints. Monetization models: subscriptions vs advertising (Priority: 5/5): The discussion contrasts consumer subscription models with ad-supported models, arguing that consumer software generally wants ads and that OpenAI and Meta should lean harder into advertising. Semiconductor supply chain and TSMC/Intel/Samsung dynamics (Priority: 5/5): A major segment focuses on chip manufacturing, fab conservatism, memory cycles, and how TSMC’s risk management has pushed supply-chain risk onto hyperscalers and accelerated onshoring efforts. Winner profiles among major tech firms (Priority: 4/5): The speaker compares Amazon, Apple, Meta, Microsoft, Google, NVIDIA, OpenAI, Anthropic, and others, arguing that scale, first-party demand, and willingness to invest through cycles will determine the winners. Power, infrastructure, and commodity markets (Priority: 4/5): He argues that AI’s ultimate bottlenecks are increasingly power and capital rather than raw compute, and that tech investors often misunderstand commodity-market behavior and cyclicality.

Key Arguments: AI’s biggest risk is not technical failure but a financing gap: if capital markets or revenues fail before the buildout pays off, the industry can suffer a major air pocket even if the technology keeps improving. U.S. AI supremacy would be geopolitically dangerous; a more stable world is one where the U.S. leads but China remains competitive enough to preserve deterrence and avoid extreme conflict. Current frontier-lag dynamics—U.S. labs ahead, China distilling models a few months behind—may persist longer than people expect because chips, iteration speed, and self-improvement compounds are hard to close. Consumer AI should likely be monetized with advertising, not subscriptions, because consumers resist paying and ad budgets can scale with usage without direct user price resistance. OpenAI’s delay in pursuing ads ceded ground to Google and Meta; a strong ad product could have created a much stronger consumer AI business model sooner. AI turns compute and intelligence into commodity-like products; that increases societal impact but compresses margins and makes cost of service the real battleground. TSMC’s conservatism and long fab lead times make the supply chain brittle; its underinvestment forces hyperscalers to bear the risk and spend aggressively on alternative capacity. Amazon is especially well positioned because it is its own first-best customer in cloud, chips, logistics, and AI services, allowing it to iterate internally and then sell externally. Apple can win by staying focused on hardware, customer access, and on-device inference, even if it does not lead frontier model development. Meta may be one of the most interesting AI setups because better personalization and ad targeting can produce enormous incremental profit from only small improvements. Microsoft’s AI strategy resembles an IBM-style middleware and enterprise-lock-in play: not necessarily frontier best, but strategically rational and cash-generative. NVIDIA’s apparent pricing power may be less secure than it looks because hyperscalers are becoming its ultimate competitors and can ultimately build or buy substitutes. Power generation, not just chips, may be the most lasting infrastructure outcome of the AI boom; if excess power is built, it could be a durable societal benefit even after bubbles burst.

Data Points: RAMP customer revenue growth: 3.2 times faster - Advertising/promo claim about Ramp customers relative to the average American business. Annual savings from Ramp: 5% annually on average - Promo claim about finance automation benefits. WorkOS enterprise readiness stack: SSO, SCIM, RBAC, and audit logs - Promo section describing core enterprise capabilities that AI companies need. Birogo/Felix output examples: PowerPoint decks, Excel models, and sourced research - Promo describing the AI finance agent’s deliverables. Vanta vendor assessment reduction: Up to 50% - Promo claim about time saved on vendor assessments. Companies using Vanta: Over 16,000 - Promo claim about customer base. Google/Meta monetization comparison: $0 content cost for Instagram - Speaker contrasts Meta’s free-content model with YouTube’s creator payments. OpenAI frontier lag vs China: About 6 to 9 months - Speaker’s estimate of China trailing frontier U.S. labs. AI infrastructure spend this year: $800 billion - Estimate mentioned for current-year capex. AI infrastructure spend next year: $1.3 trillion - Estimate mentioned for following-year capex. TSMC historical growth slowdowns: 2023, 2024, 2025 - Speaker references a growth slowdown period before a later rebound. Google ad/AI monetization issue: $100 per user per month - Microsoft pricing discussion used as a contrast for usage-based AI charging. Dropbox storage anecdote: Free storage code shared with many people - Personal anecdote illustrating consumer freemium behavior; no exact numeric quantity given. AWS external/internal servicing sequence: External customers before internal ones - Used to illustrate Amazon’s first-best-customer strategy.

Pivotal Quotes: "I think it would be very problematic for the U.S. to win." — Ben: On the geopolitical risks of a fully dominant U.S. position in AI. "Risk doesn't disappear, it just gets handed off." — Ben: On how TSMC’s conservatism shifts supply-chain risk to hyperscalers and other participants. "I am concerned that ... we just stop releasing stuff, we on the outside start to lose any sense of where is actually the frontier." — Ben: On the danger that AI safety anxieties reduce external visibility into true model progress.

Implications: AI will likely reshape every digital and physical business, but the winners will be those that manage capital intensity, distribution, and commodity constraints best. Expect more ad-driven AI, more chip and power buildout, and a prolonged strategic contest among hyperscalers, model labs, and foundries.