Donato Riccio

I think this is not always the right question for investors because it's really impossible to predict that.

But I say that the more impactful, important question right now is that when does current level intelligence become cheap enough to be everywhere?

And I think this is happening right now.

So if we take a look at how costs evolved over the years, just two years ago, GPT-4, which was the flagship model by OpenAI available at the time, cost $30 to $60 per million tokens.

A million token is around three, four books.

And so you need $30 to $60 to process this amount of information.

But today, you have available GPT-5 Mini, which is a way better model.

It just costs $2.

So the model's got around 15 to 30 times cheaper for more intelligence in just two years.

That's what we call the intelligence per dollar curve.

And watching that curve is one of the most important indicators.

If we take a look at also how many tokens the companies are processing, Google reported that they're processing a quadrillion tokens per month.

That's a crazy number.

And so it's a 5x increase year over year.

So people are deploying this at scale.

And so how is it possible that costs are falling so fast?

It's coming from multiple directions.

So first, we have algorithmic improvements.

There are new reinforcement learning and training techniques, like GRTO by DeepSeq or RLDR by OpenAI.

It can get you better results for less compute.