Two opposite things are true about the cost of AI at the same time.
The cost of using a given level of AI has collapsed, while the cost of building the most advanced systems has soared.
Understanding both is the key to the economics of the field.
TL;DR
- The cost of running a GPT-3.5-level model dropped roughly 280-fold in 18 months, to about $0.07 per million tokens.
- AI chip performance per dollar has improved about 40 percent per year, doubling every 2.2 years.
- At the same time, algorithms have grown about 3 times more efficient each year, achieving the same performance with less compute.
The cost and efficiency figures here come from the Stanford AI Index and Epoch AI’s trends analysis.
ALSO READ: Does AI Adoption Actually Track With Productivity?
The forces pushing cost down
| wdt_ID | wdt_created_by | wdt_created_at | wdt_last_edited_by | wdt_last_edited_at | Trend | Rate of improvement |
|---|---|---|---|---|---|---|
| 1 | emmanuel-ashemiriogwa | 04/08/2026 11:43 AM | emmanuel-ashemiriogwa | 04/08/2026 11:43 AM | Inference cost (GPT-3.5-level) | 280x cheaper in 18 months |
| 2 | emmanuel-ashemiriogwa | 04/08/2026 11:43 AM | emmanuel-ashemiriogwa | 04/08/2026 11:43 AM | Chip performance per dollar | 40% per year |
| 3 | emmanuel-ashemiriogwa | 04/08/2026 11:43 AM | emmanuel-ashemiriogwa | 04/08/2026 11:43 AM | Algorithmic efficiency | 3x per year |
These three compound.
Better chips, cheaper per unit of performance, combined with smarter algorithms that need less compute for the same result, drive the price of any fixed capability down at a startling pace.
The countercurrent that keeps total cost high
The reason AI does not simply get cheaper overall is that ambition scales faster than efficiency.
Since 2010, the compute used to train notable AI models has increased about 4.5 times per year.
So while each unit of intelligence gets cheaper, labs keep buying vastly more of it to push the frontier, which is why the cost of training the most advanced models keeps rising even as the cost of yesterday’s capability falls off a cliff.
So, What?
For a business in any of the core markets, the collapsing inference cost is the practical headline.
A capability that was expensive enough to be a boardroom decision two years ago is now cheap enough to embed in a routine product feature.
That is what turns AI from a pilot project into infrastructure.
The soaring training cost, by contrast, is a concern only for the handful of labs building frontier models, not for the companies using them.
ALSO READ: Data Shows Us How Much Energy AI Actually Uses
ELI5
Using AI has gotten much, much cheaper very fast, so things that were expensive a year or two ago now cost pennies.
But building the newest, most powerful AI keeps getting more expensive, because companies keep making them bigger.
Sources
Stanford HAI | Epoch AI | Epoch AI (2)
