SemiAnalysis: enterprise AI caps span $250 to $2,000 a month

The headline story of the last few months was Meta and Uber burning through their AI budgets in ways that made great copy, with one Meta employee alone reportedly running through roughly 280 billion tokens in a month. A new SemiAnalysis note from Crystal Huang, Joey Brookhart and Dylan Patel, drawn from conversations with over 50 enterprise customers via Slack, phone and the Databricks AI Summit, argues those stories were the exception rather than the pattern.

The actual budget picture the authors describe is less lurid and more revealing. An aerospace and defense manufacturer set a $250 a month cap per employee. A pharmaceutical company set $500. Workday and Stripe employees get $2,000. A cybersecurity company splits it by seniority, roughly $800 for juniors and up to $4,000 for senior staff. Uber, after burning through its annual Claude Code budget in four months, moved to $1,500 a month per employee. The spread is enormous and there is clearly no market consensus on what the right number is yet.

Data the authors cite from Ramp puts the shape of the market in sharper focus. The 99th percentile customer spends around $90,000 per employee per year on AI. The 90th percentile spends about $7,300. The median Ramp customer spends $136. That is a market where a thin sliver of firms is doing the tokenmaxxing everyone writes about while most companies are barely engaged, and it means the interesting policy questions live in the middle of the distribution, not at the top.

Conservation is the part worth watching. Companies are turning off premium models, disabling Opus 4.8 and defaulting employees to Sonnet, and employees are pushing work into Microsoft 365 Copilot's unlimited seat before touching metered budgets. That is a natural consequence of a pricing structure that rewards gaming, and it means Microsoft may be quietly absorbing a chunk of demand that Anthropic and OpenAI are trying to meter.

The honest caveat is that the piece is built from vendor conversations and Ramp expense data, not audited spend reports, so take the specifics as reported rather than settled. What the reporting does not give you is whether these caps actually change output or just shove consumption sideways into the unlimited seat, and that is the question the next round of enterprise conversations will need to answer.