SemiAnalysis: enterprise AI token budgets replace the free-for-all as spending scales into the billions

A new SemiAnalysis report, based on conversations with enterprises across industries, finds that the era of unlimited employee AI token consumption is ending — companies that spent early 2026 encouraging staff to burn through tokens are now imposing hard monthly spending caps as costs scale into real money.

What's new

SemiAnalysis writes that "tokenmaxxing started earlier this year when companies like Meta and Salesforce began encouraging their employees to consume as many AI tokens as possible," turning heavy usage into an internal status symbol rather than a cost to manage. That approach broke down once the bills came due. The report's clearest example: "Uber made headlines for burning through their Claude Code and Codex annual budget in four months. In response, the company imposed a $1,500/month/employee limit."

Uber isn't alone. SemiAnalysis details a range of caps now in place across the companies it spoke with:

• An aerospace and defense manufacturer caps employees at $250 a month.
• A pharmaceutical company caps usage at $500 a month.
• Workday and Stripe run employee budgets around $2,000 a month.
• A cybersecurity company sets tiered limits from $800/month for juniors up to $1,600–$4,000/month for senior staff.
• A travel-tech company defaults to $200/month per employee, extendable to "tens of thousands of dollars" for teams that need it.

At Meta specifically, the report notes an internal "Claudeconomics" dashboard employees built to rank the company's top 250 token consumers, after usage reportedly hit over 60 trillion tokens in a 30-day period, with the single heaviest user accounting for roughly 280 billion tokens. The dashboard was shut down two days after The Information reported on the spending. SemiAnalysis puts Meta's overall burn at roughly 70 trillion tokens per month as of February, translating to close to $50,000 per employee per year at list price.

Context

The first half of 2026 saw AI coding tools like Claude Code and Codex become default developer infrastructure at large tech and non-tech companies alike, with usage frequently treated as an unambiguous good — more tokens burned was read as more AI adoption, and some companies explicitly tied "AI-driven impact" to performance expectations. SemiAnalysis's reporting suggests that framing is now colliding with the actual economics: at list price, heavy individual usage can run into tens of thousands of dollars a year per employee, and annual budgets set before usage patterns were understood are getting blown through in months rather than a full year, as happened at Uber.

Why it matters

This is one of the clearer data points yet on what AI coding-agent adoption actually costs at scale inside large organizations, as opposed to per-seat SaaS pricing headlines. The shift from "consume as much as possible" to explicit per-employee dollar caps — some as low as $200-250/month — signals that enterprises are starting to treat token spend the way they treat any other metered infrastructure cost, with budgets, tiers by seniority, and monitoring, rather than as an unlimited perk. For AI vendors like Anthropic and OpenAI, whose coding tools are named directly in the Uber example, it's a signal that enterprise revenue growth from token consumption may face a ceiling as customers impose their own governance, even as usage of the tools themselves remains high.