Just last month, the most important metric in Silicon Valley was the burning of tokens, the units of measurement of computing power used by AI models. CEOs were giving employees the Matthew McConaughey”Those are rookie numbers, you have to increase those numbers.”speech The Wolf of Wall Street. Now, they’re asking their staff to hit the brakes.
According to a Wall Street Journal reportCorporate leaders have realized that it actually costs money to burn AI tokens, and doing it almost exclusively for the sake of doing it with no other goal in mind isn’t actually a great business strategy. Good thing these guys get paid tens of millions of dollars a year to figure this stuff out.
Earlier this week, Uber CEO Dara Khosrowshahi said it was “increasingly difficult to justify“The cost of AI initiatives within the company because production was not keeping up with the rate of token burn, while acknowledging that part of the reason they went so crazy with token burn in the first place is that it “may seem” like AI is free.
Turns out it’s not.
An anonymous AI consultant told Axios that one of his clients accidentally spent 500 million dollars in a single month because he never bothered to put a usage limit on employee access to Anthropic’s Claude. That’s… a lot. I like it to the point of straining credulity. The Journal didn’t find anything quite as egregious, but it did hear about one financial institution that saw its employees spend hundreds of thousands of dollars on tokens per month, with employees using premium-tier models to ask basic questions and have inane back-and-forth conversations.
That’s almost always going to be how this silly era of justifying spending on AI was going to be. Corporations have already spent tons of money to adopt these systems and need to justify the expense. To do this, they encourage their employees to use it as much as possible. In turn, employees do it, even when it doesn’t make sense to use AI for a task.
Goal killed your chip burning leaderboard last month after it was leaked, revealing that the main “Token Legend” managed to burn 281 billion tokens in one month, more than the amount of computing it would take to replay the entirety of Wikipedia 33 times. Amazon joined that retreat this week, according to the Financial Timesremoving their marker of employees who used the company’s internal AI tools the most, a decision that was allegedly made after it became obvious that employees were giving AI agents useless tasks just to maintain their position on the leaderboard.
It is clear that the corporate world is willing to frivolously set money on fire in an effort to justify its current cash-burning pits. However, it turns out they have limits. You can only use “tokens burned” as a metric in so many quarterly earnings calls before shareholders start wondering what the price of all those tokens is.
