Is AI Getting Cheaper or More Expensive?

I recently read Sam Altman’s blog post on AGI, and came out somewhat confused. His first two observations are:

1. The intelligence of an AI model roughly equals the log of the resources used to train and run it.
2. The cost to use a given level of AI falls about 10x every 12 months…

Aren’t these contradictory? How can models be simultaneously getting exponentially more expensive to develop and also exponentially cheaper to run?

I understand that these statements refer to different aspects — frontier model development vs. the cost of accessing a given level of AI over time. However, I’m not convinced they are entirely separate.

Let’s break it down. The first point suggests that cutting-edge models are becoming exponentially more expensive to train and operate. If costs keep increasing while performance improvements diminish (potentially approaching an asymptote), how sustainable is this trend?

The second point, meanwhile, applies to what users pay. But is this price drop driven more by reduced operational costs or by competition? Efficiency gains clearly play a role, but aggressive pricing strategies from competitors have also driven costs down.

This raises a key question: Will AI development become too expensive to sustain? Most…