Why AI token prices are on the verge of a steep drop?
A wave of more efficient AI models and Nvidia’s next-generation chips could sharply reduce the cost of generating AI tokens
AI token prices could fall significantly as new models and infrastructure improvements dramatically increase efficiency across the industry.
Business Insider reported an AI infrastructure CEO shared that upcoming models later this year will be “a lot better and more efficient,” potentially triggering a major shift in pricing.
There is growing belief that AI token costs — the standard unit for measuring AI usage — may already be starting to decline.
OpenAI CEO Sam Altman has previously said rising AI costs are a major challenge, adding that companies will need to provide “more value for less spend.”
A token usage index tracked by Silicon Data reportedly fell from 2.06 in late May to 1.75 by June 10, suggesting early signs of easing price pressure.
Blackwell Drives Efficiency Gains
A major factor behind expected price drops is Nvidia’s new Blackwell GPU architecture, now being deployed in large-scale AI data centres.
These systems significantly outperform previous-generation chips and are expected to support faster and cheaper AI model training and deployment.
Research firm SemiAnalysis compared Nvidia’s Blackwell GB300 NVL72 system with the older Hopper HGX 200 and found major performance improvements.
Blackwell systems can generate up to 6,000 tokens per second per GPU, compared to around 90 in older systems — roughly 65 times more output.
When measured per megawatt of power, Blackwell systems also show major gains, producing about 2.8 million tokens per second versus 54,000 for Hopper.
Cost efficiency improvements are equally significant, with estimates suggesting Blackwell reduces cost per million tokens from around $4.20 to just 12 cents.
As these systems scale across data centres in 2026, analysts expect a surge in low-cost token supply, which is likely to push AI providers toward lower pricing.