The latest 'crack in the thesis' for the trillion-dollar AI boom: Tokens are getting cheaper

For months, the bullish story around AI has leaned on a simple idea: better models will pull in more demand, and that demand will help justify the huge bills behind the infrastructure. Nvidia chief Jensen Huang has called it the “two exponentials” problem. One exponential is model complexity. The other is usage. Put them together, and the math is supposed to work out for the people financing the buildout.

But the data is moving the other way. Ramp, the corporate spending platform, said Wednesday that American businesses are now paying about 41% less for a million tokens than they were at the March peak. The effective price has dropped from $1.15 to 68 cents. Frontier models are also losing share inside Ramp’s data, slipping from about 53% of usage in early August to 45% in September. Even the biggest AI spenders are tightening up: the top 1% cut per-employee spend by nearly 10% in August.

Ara Khazarian, Ramp’s chief economist, calls it a “crack in the AI thesis,” and that feels about right. This isn’t a collapse. It’s more annoying than dramatic. Tokens are starting to behave like a commodity, and commodities do not usually support trillion-dollar valuations by themselves. Morgan Stanley has already pointed to as much as $300 billion in bonds tied to neocloud buildouts if token prices fail to keep pace.

The squeeze is coming from both directions. Labs are cutting prices to defend usage, while customers are sliding to cheaper models. Ramp says OpenAI cut the cost of its GPT-5.6 Luna model by 80%, and Anthropic announced price cuts last month. Businesses, meanwhile, are steering staff away from the priciest systems and toward mid-tier models such as Terra and Sonnet, which Khazarian described as strong performers that cost less.

The pattern also shows up in the numbers from the model makers themselves. Since Aug. 1, Ramp says OpenAI’s effective price has fallen 38% to 48 cents, while Anthropic’s has dropped 22% to 90 cents. Anthropic still holds a higher floor, but OpenAI is taking more share on price. OpenAI’s finance chief, Sarah Friar, told a Goldman Sachs conference on Tuesday that she would like to move away from token counting altogether and charge enterprise customers for completed work instead. That sounds less like a pricing tweak and more like a company trying to escape the scoreboard.