>The cost to use tokens is already zero or deeply negative, depending on your accounting
???
Sure, there's zero cost tokens in the form free chatgpt, but negative cost implies that you're getting paid to consume tokens, which I'm not aware of?
In the sense that OpenAI had ~$2.5 billion of revenue 2024 but likely spent ~$3.8 billion just on inference.
Meaning they lose 50 cents for every dollar they spend to send you a token, more if you are a free user. This discounts employees, CapEx, real and imagined (1.4T), overhead, and training (subsidized by M$oft).
Thus, if you have a way to turn every token they give you back into USD, even at a discounted rate, you can and should rob them blind before the well dries up.
That doesn't mean the cost is negative, it means the profit is negative for openai.
I mean, sure, sign me up, and I'll use a bunch of negative-cost tokens...
In economics, the Jevons paradox occurs when technological advancements make a resource more efficient *to use*; however, as the cost of *using the resource drops*, if demand is highly price elastic [PED], this results in overall demand [for the resource] increasing.
The cost to use tokens is already zero or deeply negative, depending on your accounting. And the author is mistaking the cost of 'creating/extracting the resource' (token generation economics) with cost of 'using the resource' (calling an API). I don't think we'll understand the true cost of, and demand for AI services for quite some time.