TokenBudgeting: Our Conversations with Enterprises on Token Spend
- Source
- newsletter.semianalysis.com
- Author
- Crystal Huang
- Date
Context
An employee's AI allowance can function as a spending cutoff or a trigger for discussion. In conversations with more than 50 enterprise customers, research publisher SemiAnalysis found both approaches. One cybersecurity company reportedly assigned different allowances by seniority and role, then alerted managers when employees exceeded them instead of cutting off access. That arrangement lets managers examine unusually expensive work while keeping it moving. The interviews illustrate why a single company-wide allowance can conceal different needs across jobs; they do not establish a representative industry budget or measure the productivity gained from each dollar. Enterprise-spending report.
- Spend is extremely top-heavy. In Ramp's data, 99th-percentile customers run about $90,000 per employee per year, 90th-percentile about $7,300, and the median customer $136. Even Meta, near $50,000 per employee, is only a 3-5% customer for Anthropic.
- Caps are converging on nothing: $250 a month at a top-three aerospace and defense manufacturer, $500 at a giant pharma, roughly $2,000 at Workday and Stripe. There is no emerging industry number, only each company's own tolerance.
- The more mature pattern is a soft limit rather than a hard cut-off. One public cybersecurity company sets $800 a month for juniors and $1,600-$4,000 for seniors; going over alerts a manager for a conversation instead of shutting access off.
- The cheapest lever is the default model, not the cap. One travel-tech firm with 800 engineers flipped everyone's default from Opus to Sonnet, leaving Opus reachable but requiring a deliberate choice to use it.
- SemiAnalysis attributes over 70% of OpenAI's and Anthropic's ARR to coding use cases, with Anthropic's share higher because its revenue mix is 90%+ B2B against OpenAI's roughly 60%.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
“The results showed that Meta employees consumed over 60T tokens over a 30-day period, with the single highest individual accounting for roughly 280B tokens.”
“Most recently, Uber made headlines for burning through their Claude Code and Codex annual budget in four months.”
“We believe that over 70% of ARR today across OpenAI and Anthropic can be attributed to coding use cases with Anthropic spend higher than OpenAI’s given the difference in B2B vs B2C mix (Anthropic 90%+ B2B vs OpenAI 60%).”
“In contrast to the high numbers of token spend in the news flow, Anthropic’s own documentation says the average Claude Code usage per developer is between $150-$250 per month, and only 10% of users spend over $30 per day.”
“Across over 50 conversations with medium to large enterprises, one thing became clear: the headline Uber, Meta, and other Fortune 500 tokenmaxxing stories were a result of poor incentives and lax oversight rather than an absence of high ROI activities/projects to spend those tokens on.”
Checking sign-in…
Loading comments…