Tesla told employees this week it will impose a $200 weekly cap on AI spending starting July 6. Software engineers had been consuming thousands of dollars in tokens each week. The cap requires sign-off for anything above the limit. It applies to every AI tool the company has approved for internal use.
Every tool except one. Beta versions of xAI products, including Grok, are exempt. The exemption is notable because Tesla engineers largely prefer Claude, Anthropic's model, according to four people familiar with internal usage patterns. Grok is not popular. The cap does not restrict the unpopular tool. It restricts the popular one.
Elon Musk is CEO of both companies. The expense policy is not a cost-control measure. It is a demand-routing mechanism. Revenue that would flow to Anthropic now flows to xAI, redirected by the employer's authority over the employee's toolchain.
Tesla is not the first. Uber burned through its entire 2026 artificial intelligence budget by April. The company had rolled Claude Code out to roughly 5,000 engineers, and 70% of committed code was AI-generated. Uber had ranked engineers on internal leaderboards by Claude Code usage. The leaderboard that gamified adoption became the budget line that required a quarterly explanation.
Monthly API costs for power users ranged from $500 to $2,000. Multiply by thousands of engineers, and the math overwhelms any annual budget modeled on pilot-phase run rates.
Uber's COO Andrew Macdonald offered the clearest executive admission of the disconnect. The link between rising AI tool usage and useful consumer-facing innovation, he said, is not there yet. It is very hard to draw a line.
Microsoft's Experiences and Devices division canceled Claude Code licenses entirely by June 30. Engineers will use GitHub Copilot CLI instead. The pattern was identical: engineers adopted Claude Code rapidly over six months, preferred it to Microsoft's own tool, and the per-token pricing produced costs that exceeded every budget projection. June 30 is the last day of Microsoft's fiscal year. The cancellation is a balance-sheet decision timed to a quarterly close.
An unnamed company spent $500 million on Claude in a single month after it failed to set usage limits on employee licenses. Walmart placed usage caps on Code Puppy, its internal AI coding assistant, and implemented token-based rationing tied to demonstrated business value.
The pattern is uniform. The companies that pushed hardest for AI adoption are the first to impose spending caps. The tool worked. The tool was expensive. The tool was expensive because it worked.
Software had zero marginal cost after the license. A company could deploy Slack or Jira to 10,000 employees and the incremental cost of the 10,000th user was rounding error. AI tokens are metered. Every query has a price. Every response has a price. The more useful the tool, the more an engineer uses it, the higher the bill. The value proposition and the cost function point in the same direction. This is the structure of a utility, not a software product.
Gartner projects worldwide AI spending of approximately $2.59 trillion in 2026. It simultaneously describes the technology as in the Trough of Disillusionment. Only one in four S&P 500 companies reported quantifiable AI impact in Q1 2026, according to Morgan Stanley. The spend is rising. The evidence of return is not.
Tesla's cap lands on July 6. Uber's leaderboards are gone. Microsoft's Claude Code licenses expired yesterday. And the company that spent $500 million in a month learned what every utility customer eventually learns. The meter was running the entire time.