How Much Will Intelligence Cost in the Future?

AI-driven intelligence costs are falling fast: token prices are dropping faster than nearly anything in tech.

We might see intelligence (compute, inference) become a renewable, fungible resource, fueled by inputs like hardware (GPUs, TPUs, ASICs), electricity, and the supply chains necessary for their creation and deployment.

But true commoditization isn’t guaranteed:

1. If hardware remains scarce, prices will rise (though the opposite keeps happening)

2. If someone develops next-level AI (e.g., GPT-6) that unlocks entirely new use cases and can charge a premium before competition catches up, scarcity will again drive prices higher. Especially if inference hardware is limited, creating a bidding war among those willing to pay. This hasn’t happened yet because current LLMs (like GPT-4) don’t deliver enough value to justify significantly higher prices. But if a GPT-6 were capable of matching or surpassing highly skilled individuals, you wouldn’t price it at $10 per million tokens. Instead, you’d charge as much as customers are willing to pay, based on the value it creates.

The result? Basic intelligence could drop to just energy and material costs, while the newest, high-value AI remains exclusive, commanding premium prices.

Long-term: Even premium AIs may eventually become commodities, with the frontier continuously advancing. This would grant everyone access to better and better intelligence, while those at the cutting edge push the boundaries even further.

*Intelligence = the ability to solve intellectual tasks (not just physical work). Here I focus only on inference, not training, because inference has the potential to scale infinitely, as demand will be infinite, unlike other resources (e.g., food, entertainment, goods), which are limited by our time and attention.