The bubble argument is everywhere right now, and the data cuts against it. Our view in short: the tension between expensive tokens and corporate cost-cutting is not evidence the AI trade is breaking. It is the mechanism that drives usage, and therefore lab revenue, higher. Walls of worry are what durable trends climb.

The tension, in numbers. Tomasz Tunguz (Theory Ventures, June 29) estimates Anthropic spends roughly 2.3x its payroll on compute; the top software companies historically run about 0.4x, and the average company far below that. Yet the leading edge of adoption keeps compounding: a quarter of OpenAI’s own staff now run four or more AI agents concurrently every week (Ethan Mollick, June 30, citing OpenAI’s joint study with academic economists), and agentic workflows consume on the order of 1,000x the tokens of a simple chat query. Frontier usage is expanding fastest exactly where cost discipline is tightest. SemiAnalysis reports Uber exhausted its entire annual AI-coding budget in four months. That reads as distress to some. We read it as demand outrunning planning.

Taking the bear case seriously. The skeptics deserve a hearing, and Gary Marcus (June 28) makes the sharpest version of it: there is no durable moat. Chinese open-weight models reach near-parity six to twelve months behind the frontier, price wars push token prices toward zero, and the labs are already selling inference below cost. Reporting suggests OpenAI loses money even on $200/month subscriptions, and inference now approaches 85% of enterprise AI budgets. Unlike classic software, AI has no zero-marginal-cost leverage: when usage scales, compute scales with it. Even Sequoia’s David Cahn, hardly a pessimist, warned this week that the bottleneck trade has overshot, with Micron’s market value now above OpenAI’s.

Why the wall of worry tends to win. Every prior compute platform resolved this same squeeze the same way: unit costs collapsed and total spend rose anyway. That pattern is repeating. Frontier per-token prices are down more than 90% since 2023, yet token consumption grew roughly 1,001% from January 2025 to April 2026, and average enterprise AI budgets have moved from about $1.2M (2024) to about $7M (2026). Fortune covered this in June as Jevons’ paradox: cheaper intelligence doesn’t shrink the bill, it multiplies the workloads. Goldman Sachs’ house view is that soaring agent usage will boost tech cash flow, and Ramp’s June spend index shows the most aggressive 1% of firms spending $7,450 per employee per month on AI, a 680x gap over the median that likely reflects the adoption runway rather than the ceiling. Two caveats are worth noting: Ramp measures its own customers’ card spend, and the 1,001% figure is platform-measured usage, not revenue.

The conclusion we draw. Cost-cutting doesn’t kill AI spend. It redirects it toward whatever earns its tokens. Anthropic redeployed Fable 5 this week and moves it to paid usage credits on July 7; Polymarket currently prices a 76% chance it lists by year-end. The skeptics’ margin question is real, and it is precisely the question the coming IPOs will price. Identifying early the names that answer it well is the work we are focused on.