According to a May 22 report by Fortune, after Microsoft (マイクロソフト) rolled out Claude Code on a large scale internally, the expenses incurred by engineers using this AI programming tool surpassed the payroll costs required to hire equivalent human talent. This prompted the company to revoke some employees’ Claude Code licenses. Similar issues have surfaced at other firms: Praveen Neppalli Naga, Uber’s Chief Technology Officer, revealed in April this year that the company exhausted its entire annual budget allocated for AI programming tools in just four months — despite previously incentivizing AI tool adoption via internal leaderboards. Bryan Catanzaro, Vice President of Applied Deep Learning at NVIDIA (エヌビディア), also stated, “For my team, computing costs have long exceeded employee salaries.” Notably, Microsoft’s license revocations for Claude Code do not affect its broader Foundry partnership with Anthropic, which includes up to $5 billion in investments and access to Claude models.
The root cause of this trend lies in the combined effect of token-based billing and the scaling up of AI agents. Goldman Sachs forecasts that by 2030, widespread deployment of AI agents will drive monthly token consumption up 24-fold, reaching 12 quadrillion tokens per month. Gartner predicts that inference costs for trillion-parameter models will drop nearly 90% over the same period; however, the firm cautions that falling token prices do not necessarily equate to lower overall AI costs for enterprises. Since agent-based systems require far more tokens than standard conversational models, such volume surges could easily offset price reductions. Its analysts warn, “Do not mistake deflationary trends in bulk token pricing for broader accessibility of cutting-edge reasoning capabilities.” Meanwhile, Microsoft transitioned GitHub Copilot from request-based to usage-based billing, turning hallucinations — once merely a functional nuisance — into a significant operational expense.