Microsoft Tells Its Own Engineers to Ease Off the AI

There is a nice irony in the memo that went round Microsoft this week. The company whose pitch to the rest of us is that every developer should have an AI assistant running has just told its own engineers to slow down. The message, from executive vice president Jay Parikh and first reported by 404 Media, introduces division-level AI token budget targets and lets individual engineers watch their own usage climb.

The line people have latched onto is the one that names the problem. "Tokenmaxxing is not what we are optimizing for," Parikh wrote, adding that he wants everyone focused on outcomes that move the needle for customers and the business rather than on running up their consumption.

What tokenmaxxing means

Tokens are the units a language model reads and writes, and they are what you pay for. Tokenmaxxing is the habit of firing off more queries, spinning up more agents and testing more models than the task really needs, partly because it is easy and partly because heavy usage had started to look like a badge of how hard someone was working. According to the reporting, some engineers were running from a few hundred dollars to a few thousand a month in tokens.

Alongside the budgets, Microsoft is making OpenAI's GPT-5.6, which is cheaper to run, the default model for internal work. Staff can reach for something heavier, but the quiet nudge is to reach for it only when the job warrants it.

It would be easy to read this as Microsoft going cool on AI. I do not think that is it. The more interesting reading is that the bill has arrived. For two years the industry talked about AI usage as if more was always better, and a company running these tools at the scale of tens of thousands of engineers is now watching what unmetered enthusiasm actually costs. Setting a budget is not a retreat from AI. It is what happens when a technology stops being a demo and becomes a line item, one that shows up on the same balance sheet as everything else Microsoft is building.