Jensen Huang, chief executive of Nvidia, has warned that the rapid fall in AI token prices could undermine the lofty expectations surrounding the sector's growth. Recent figures show that American businesses are now paying considerably less for the compute that powers large language models.
Token prices tumble
Data released by corporate-spending platform Ramp indicates that the effective cost per million tokens dropped from $1.15 in March to 68 cents in September, a 41 per cent decline. The share of usage devoted to frontier models fell from 53 per cent in early August to 45 per cent by September, and the top one per cent of spenders, who generate roughly 80 per cent of enterprise revenue for firms such as OpenAI and Anthropic, cut per-employee spend by nearly ten per cent in August.
Why the shift matters
The trend suggests that AI tokens are becoming more like a commodity, comparable to basic raw materials, rather than a premium asset that could justify trillion-dollar valuations. Morgan Stanley has highlighted a potential vulnerability of up to $300 billion in bonds that finance data-centre builds for companies such as CoreWeave, should token prices continue to fall.
You have multiple metrics now starting to move in a negative direction.
Ramp chief economist Ara Khazarian explained that the decline reflects both price cuts by model providers, OpenAI reduced the cost of its GPT-5.6 Luna model by 80 per cent and Anthropic announced similar reductions, and a migration by customers toward cheaper, simpler models.
Companies turn to mid-tier models
Businesses are increasingly adopting mid-tier offerings such as Terra and Sonnet, which deliver strong performance at lower cost. Ramp's most AI-intensive firms now spend about $7,200 per employee per month on AI, roughly a third of the target set by Huang, and the figure is trending downward.
What comes next?
OpenAI's chief financial officer Sarah Friar told a Goldman Sachs conference that the firm hopes to move away from token-counting altogether, favouring pricing based on completed work. She added, "I would love, to get us away from token-counting." If the industry can shift to usage-based pricing models, it may stabilise demand and protect financing structures that rely on higher token prices.
Analysts remain divided. While some see the rise of open-source alternatives as a factor, only a small fraction of Ramp's users, 3.6 per cent, currently employ open-source or Chinese models. The competition between OpenAI and Anthropic continues, with Anthropic maintaining higher prices but seeing its advantage erode as OpenAI captures a larger share of token usage.
For now, the AI sector appears to be entering a period of cost discipline. Whether this will lead to a more sustainable growth path or signal a broader correction remains to be seen.