AI Compute Prices Plunge, Sparking a New Price War

In a market that has long been driven by the promise of ever‑growing AI models, a new trend is emerging: the cost of tokens is falling faster than anyone expected. According to fresh data from Ramp, the corporate spending platform, the effective price that U.S. businesses pay for a million AI tokens has dropped roughly 41% since its March peak—down from $1.15 to just $0.68.

The shift is not limited to a single vendor. OpenAI’s GPT‑5.6 Luna model was slashed by 80%, and Anthropic announced its own price cuts last month. Meanwhile, the share of token usage devoted to frontier models has slipped from about 53% in early August to 45% by September, as companies trade down to cheaper, simpler offerings.

The impact is most pronounced among the top 1% of spenders, who generate roughly 80% of enterprise revenue for OpenAI and Anthropic. In August, these heavy users cut per‑employee spend by nearly 10%, signaling a broader market recalibration.

Industry analysts warn that the trend could undermine the financial assumptions underpinning large‑scale cloud infrastructure projects. Morgan Stanley has highlighted a potential vulnerability of up to $300 billion in bonds issued by neocloud builders—companies that built data centers before securing tenants—if token prices fail to recover.

Citadel Securities noted a similar pattern in its own LLM Expenditure Index, which has begun to fall as the market bifurcates between high‑end, tech‑heavy firms that can afford frontier AI and the broader economy that relies on everyday AI.

Jensen Huang, Nvidia’s CEO, has long described AI compute growth as driven by two exponentials: the increasing complexity of models and the expanding user base. The current price decline suggests that the market may be shifting toward a more commodity‑like perception of tokens—comparable to staples such as salt or wheat—rather than a premium for cutting‑edge innovation.

"The price decline reflects a mix of labs forced to cut prices and customers moving to cheaper models," said Ara Khazarian, Ramp’s chief economist. "It poses a threat to anyone who expects the AI companies to grow unchecked.",

As the AI economy matures, the new reality may be that the next wave of competition will be defined not by breakthroughs in model architecture but by who can deliver the most cost‑effective compute at scale.