AI is costing Microsoft, Uber more than humans: Why big tech’s AI bills are exploding

0:00 For the last 2 years, Silicon Valley has been selling one big promise. AI will make coding faster, companies leaner in terms of workforce and software development cheaper. But now, something strange is happening. The more companies use AI, the more expensive it is becoming to run. Microsoft is now reportedly canceling most of its direct Claude code licenses as well, just months after rolling the tool out across thousands of employees, including developers, designers, and product managers. According to The Verge, Microsoft's Experiences plus Devices division, the team behind Windows, Outlook, Teams, and Surface, is cutting down Claude usage before the company's new financial year begins in July. Instead, Microsoft is pushing employees towards its own GitHub Copilot CLI. And suddenly, the AI industry's biggest contradiction is starting to show. Recently, Uber reportedly burned through its entire 2026 AI budget by April. Why? Because nearly 5,000 engineers started heavily using Anthropic's Claude code faster than the company had expected, and Uber is not alone anymore.

1:03 That is how AI was sold. AI coding tools would automate repetitive work, fix bugs faster, and improve productivity. But the tools themselves are now becoming a major operating expense, and the reason comes down to something most people rarely hear about, AI tokens. Every prompt, every line of AI generated code, every chatbot response consumes computing power, which means the more employees use AI, the higher the bill becomes. And companies are actively encouraging that usage. Amazon reportedly pushed employees to token max, essentially use as many AI tokens as possible.

1:37 At Meta, workers even built an internal tracker called Claude-onomics to monitor AI usage. Goldman Sachs estimates that agentic AI systems could increase token consumption 24 times by 2030, hitting 120 [music] quadrillion tokens every month. Yes, token prices are expected to fall. Gartner says inference costs for large AI model could decline nearly 90% by 2030. But there's a catch. Agentic AI systems consume far more tokens per task. So even if prices fall, overall spending may still explode. And that changes the entire AI narrative because Silicon Valley may have solved the capability problem, but the economics problem, that still looks deeply unresolved. Follow Business Standard for more such stories.