AI Eats the World? A Reality Check with Benedict Evans

What It Covers

Tech analyst Benedict Evans reviews what AI has delivered since his "AI Eats the World" presentation 18 months ago. Coding tools with product-market fit dominate early adoption, while foundational model companies face commoditization risk. Evans maps parallels to mobile, internet, and PC platform shifts to frame what remains genuinely unknown about value capture and enterprise transformation.

Key Questions Answered

• •Coding as the only proven PMF: Cursor's annualized revenue jumped from $9B to $47B run rate in months, making software development the sole AI use case with undeniable product-market fit. Everything else remains experimental. Builders and investors should treat coding as the benchmark for what "working" looks like, and apply that standard rigorously before declaring other verticals ready for scaled investment or deployment.
• •Foundation model commoditization risk: With 3–6 frontier model companies competing on identical chips, no network effects, and $1–2T in CapEx entering the market while efficiency improves 100–200x annually, pricing power erodes structurally. Chip makers, ISPs, and mobile operators all built critical infrastructure without capturing value. Model companies should urgently identify up-stack leverage before token pricing collapses toward marginal cost.
• •Value moves up the stack, not down: Mobile networks spent $200B annually on CapEx, grew data traffic 1,500–2,000x over 15 years, and still saw flat stock prices for two decades while Apple, Google, and Meta captured the returns. AI infrastructure investors should pressure-test whether their position resembles a telco or an iOS — only operating-system-layer control with network effects historically generates durable margin.
• •Daily vs. weekly usage gap signals incomplete product-market fit: Current data shows only 10% of users engage with AI tools daily, while 40% use them weekly. Weekly usage indicates the tool hasn't become habitual or essential. Product teams should diagnose whether low daily engagement reflects a workflow integration failure, a pricing mismatch, or a fundamental capability gap — each requiring a different intervention strategy.
• •Industry-specific transformation requires domain expertise, not just AI expertise: What AI means for law firms, consultancies, or advertising depends entirely on understanding internal pyramid hiring structures, client billing models, and undocumented workflows — knowledge that San Francisco rarely holds. Companies deploying AI in professional services should embed domain specialists in product design, not just engineers, because the relevant questions are industry questions, not technology questions.
• •CapEx growth has physical limits approaching: Microsoft, Meta, and Google are each on track to spend over 50% of revenue on CapEx in 2025, totaling roughly $700B across major players — comparable to the entire global oil and gas sector's annual capital spend. Growth at this rate cannot compound further without borrowing at unsustainable levels. Investors should model CapEx tapering as a base case, not an outlier scenario.