Missive #239 (Tokens, tokens, everywhere).

Friends:

Tokens are the unit economics of the AI era. Each query you type into your favorite AI chat platform consumes X tokens to process the input, and another Y tokens to process the output.

But what’s the cost of one token? We know what OpenAI is charging for tokens, but what’s their cost of producing them? And what will the company’s unit economics look like a few years from now when they are operating at steady state?

This is the zillion-dollar question.

Right now, OpenAI, Anthropic, and the rest are just struggling to keep up with demand. OpenAI says that in May 2025 they processed 97.4 billion tokens in a single day. Were their unit economics positive? Who knows? We know that overall they are losing money, still spending large buckets of cash on building capacity.

In a normal world, you would expect unit economics would improve with scale, and you’d expect that production efficiency gains would eventually help them catch up with demand.

But this is where “Jevons’ Paradox” enters the group chat.

William Jevons was an economist who in 1865 wrote a book called The Coal Question. He looked at data indicating that while James Watt’s steam engine should have reduced the demand for coal (because it used coal so much more efficiently than previous machines), Watt’s invention actually massively increased demand for coal in Britain because it made coal-powered industries far more profitable and widespread. An economic paradox.

Two centuries later, it’s happening again. Last year Google announced a 33-fold reduction in energy consumption for the median Gemini prompt. Seemed like a good sustainability milestone. But no, it has just made us all burn through even more AI tokens, causing data center demand to increase even faster.

The number of AI data centers being built right now is staggering. Global data center energy use is expected to double by 2030, reaching 945 TWh and representing nearly 3% of global electricity consumption.

And then there’s the cost of the environmental damage. Or the “negative externalities” as the economists say.

If you add all that up, what will the true economic cost of AI tokens look like in five years? Will these companies be able to operate profitably (and not destroy the planet with their data centers)?

The very smart people at Google, Anthropic, and OpenAI are betting that scale and efficiency will eventually solve the economics.

But Jevons is still lurking in the group chat, pointing out that more efficient AI may just mean we use a lot more of it, creating endless cycles of demand and capital investment. We’ll see.

Have a good week, all.

-Bret