Big Tech is quietly discovering that AI coding agents cost more than the human workers they were supposed to replace — bad news for the "AI takes your job" story, good news for employees.

• Microsoft gave engineers Claude Code, saw usage explode, then yanked the licenses in mid-2026 because the bill got too big, pushing people to its cheaper in-house tool.
• Uber burned its entire 2026 AI budget in 4 months; one executive torched $1,200 of tokens in a single 2-hour demo, and heavy users ran up thousands a month.
• The problem is AI "agents" that work on their own — they retry failed tasks over and over, each attempt costing full price, so a single job can cost 5 to 30 times a normal chat.
• Cheaper chips don't help: every time a token gets cheaper, the agents just use far more of them, so the savings never reach the bottom line.
• Note that Microsoft still poured billions into Anthropic — it believes in the product, it just doesn't want to pay to run it internally.

Outlook: Expect companies to keep AI in small, high-value tasks with a human watching for costly runaway loops, rather than replacing workers wholesale.