Token

There’s been a lot of debate about token pricing lately. Some people think chip costs are going up, that the pace of new models is too aggressive, and that token prices will keep climbing. Then someone replied on X with a take that stuck with me: tokens are at the stage mobile data was 15-20 years ago. AT&T, China Mobile, everyone had these tiny data plans. 200MB, 500KB, 1.2MB per month. Opening a webpage, you’d think about whether it was worth the data. People literally counted their kilobytes.

Now everything runs on basically unlimited data. Nobody opens an app and worries about bandwidth. Consumers pay for it willingly and don’t even think about it. The entire app economy we have now couldn’t exist when data was metered in kilobytes.

Tokens are where mobile data was back then.

Right now people ration tokens. I catch myself checking how much a single agent run costs before I kick it off, same way I used to check my remaining data before loading a video. Everyone building with AI thinks about token budgets the same way we used to think about data plans.

But the cost of a token is basically electricity. And if compute efficiency keeps improving at the rate it has been, exponentially, then in 3-5 years tokens that match or beat today’s best models could be nearly free. And maybe 100x tps faster than what we have now.

That changes what tokens are for. Right now most people still think of tokens as the thing behind a chatbot, a conversation interface. But tokens are really the unit of intelligence that drives everything an agent does. When an agent researches a company or a founder, writes code, runs a search, all of that is tokens being spent. Tokens aren’t chat. They’re the fuel that powers all intelligent operations.

When mobile data got cheap enough, nobody was thinking about “how do I use my data plan better.” They were building Uber, Instagram, TikTok. The constraint disappeared and entirely new things showed up that nobody was planning for when they were counting kilobytes.

I think the same thing is about to happen with tokens.