Nobody can tell you what a ChatGPT answer costs.
Not the price. The cost.
Trillions of dollars are being committed on the assumption
that this number falls fast enough, and almost nobody has
done the arithmetic in public.
So I'm going to. Every week. Until someone shows me I'm wrong.
Teardowns Tuesday and Friday.
My own charts. My work shown every time.
Why I'm going to start making dated predictions.
Analysis without a prediction is unfalsifiable, and unfalsifiable analysis is entertainment.
If I think capacity will outrun demand, I should say by when. If I think inference costs will keep falling, I should say how fast and check in a year.
The point isn't to be right. The point is that being wrong in public, on a schedule, is the only mechanism that improves judgement.
Everyone in this sector has opinions. Almost nobody has a scoreboard.
A rule of thumb worth having:
In most AI products, a small minority of requests consume the majority of the compute.
Long documents. Complex agent runs. The user who pasted their entire database in.
If you haven't looked at the distribution of cost per request, you don't know your economics. You know your average, and the average is hiding the shape.
Find your top 1% of requests by cost. That's where your margin went.
What payback period would you accept for an AI project?
1) under a year
2) one to two years
3) three to five years
4) strategic, payback doesn't matter