The Demand Numbers Behind AI's Trillion-Dollar Bets Are Partly Manufactured

At Amazon, clearing the company's AI targets has become a matter of running the right tool. Employees use MeshClaw, the in-house agentic assistant, to automate non-essential tasks and inflate their token-consumption scores, the better to hit a mandate that 80% of developers use AI tools weekly. Meta keeps a "Claudeonomics" leaderboard ranking staff by how much they consume; Microsoft has declared AI use "core to every role and level" ^[1]. The practice has a name — tokenmaxxing — and it is not a fringe habit but a rational response to a system that measures adoption by volume. That volume is now doing heavier work than a performance review. Jensen Huang has taken to citing per-engineer token consumption as a productivity metric, which means the person selling the GPUs and the companies buying them are both pointing at the same number as justification for hundreds of billions in capital ^[1]. It is the number employees are inflating. The financial structure leaning on it is not small, and it is being put in front of public investors right now. Anthropic filed to go public this week and is targeting a $2 trillion valuation on projected revenue that requires token consumption to grow far faster than per-token prices fall ^[2][3]. OpenAI is spending roughly thirty times its annual revenue on physical infrastructure while losing money on operations ^[4]. And prices are falling hard: DeepSeek's models run 12 to 19 times cheaper than the U.S. flagships for equivalent work ^[5]. At the same time, the cost of running these tools has begun to exceed the cost of the people they were meant to assist. Nvidia's own vice president of applied deep learning put it plainly.

For my team, the cost of compute is far beyond the costs of the employees. — Bryan Catanzaro

Microsoft cancelled thousands of Claude Code licenses and Uber exhausted its entire 2026 AI budget in four months ^[6]. Where the real revenue sits is narrower still. The Ramp AI Index, which tracks 70,000 businesses, finds the top 1% of AI spenders pay a median of $7,400 per employee each month, and that growth is "primarily driven by the use of coding agents" ^[7]. That is a base of heavy agentic-workflow users, not broad adoption. No source establishes that gamed internal consumption materially inflates lab revenue. But the demand figures investors are being asked to underwrite blend mandated adoption, gamified metrics, and concentrated heavy-user spend in ways that may not be separable from the organic productivity gains the math requires. Whether the volume can be separated from the mandates that produced it is the question the roadshow will have to answer.