Microsoft Just Revealed AI Coding’s Biggest Problem

追加: 2026-05-30

[00:00:00]For the last couple years, everyone in tech kept saying the same thing. AI is going to replace human programmers. The logic sounded simple. Humans cost money.

[00:00:10]They take breaks. They need sleep. But AI writes code in seconds, costs pennies, and works 24/7.

[00:00:17]So tech companies started pushing their employees to use as much AI as possible.

[00:00:22]They called it token maxing. A token is just a tiny piece of data. Think of it like a syllable in a sentence. At Meta, workers made a leaderboard called Clodonomics to see who could use the most AI. At Amazon and Uber, managers forced employees to run AI tools all day just to keep their internal scores high.

[00:00:43]So what did the bosses actually want?

[00:00:46]Massive productivity for almost no cost.

[00:00:50]They dreamed of a world where one human worker sits next to 100 AI assistants and does the work of an entire department. And for a while, the math looked great. Experts said that by 2030, running a top AI system would cost 90% less. Companies thought they were about to save billions by replacing people with machines.

[00:01:11]But then the math broke, and Microsoft just exposed the biggest problem with AI coding. Here it is. AI is so hungry for power that using it widely is just too expensive. The tech industry ran straight into an old economic trap from the 1800s called Jevons paradox. The rule is simple. When you make something cheaper and more efficient, people don't use less of it, they use way, way more.

[00:01:39]Instead of asking AI one simple question, companies started using something called agentic AI. These are AI agents that work like tiny robots. To fix one single bug in a program, an AI agent might secretly create 10 other AI sub agents in the background. One writes code, another tests it, another looks for mistakes, another rewrites it. This loop happens thousands of times before you ever see the answer.

[00:02:07]Because of this, experts now say AI use will explode 24 times higher by 2030.

[00:02:13]And when you use something 24 times more, a 90% discount doesn't matter anymore. Your bill still goes through the roof.

[00:02:22]Then the turning point came. The bills arrived. Uber's tech chief made a shocking admission. Because they turned AI into a game, Uber burned through its entire 2026 AI coding budget in just 4 months.

[00:02:37]At Meta, the Claudeonomics leaderboard showed workers using 60 trillion tokens in 30 days. That created millions of dollars in pure waste. But here's the biggest one. Just 6 months after giving thousands of employees access to a powerful AI tool called Claude Code, Microsoft panicked. According to reports from The Verge, Microsoft started canceling most of its Claude Code licenses. They're pulling back the tool their own engineers loved and forcing them onto cheaper, more limited internal tools.

[00:03:09]An AI vice president at Nvidia put it bluntly, and I'm quoting here, "For my team, the cost of the computers is far beyond the cost of the employees."

[00:03:19]Read that again. The electricity and computer chips now cost more than human salaries.

[00:03:26]So, here's the new reality. The big plan to replace human workers with cheap AI has hit a massive wall.