Foreign media reports that a key assumption underlying the high valuation and large-scale capital expenditures of AI is showing signs of weakening: the actual cost for companies to purchase the model invocation capabilities of AI is declining, rather than continuing to rise with the release of new models.
Enterprise expenditure data released on Wednesday Ramp shows that American companies are paying an effective price of token per million, which is down about 41% from the high in March, from $1.15 to $0.68. At the same time, the proportion of cutting-edge models in overall usage also decreased from around 53% at the beginning of August to 45% in September.
Corporate spending has begun to shrink.
The article states that what is most noteworthy is not just the decline in prices, but also the changes in spending patterns of high-consumption customers. For the top customers who drive the revenue of OpenAI and Anthropic companies, their average AI expenditure in August decreased by nearly 10% month-on-month. Ramp's chief economist, Ara Khazarian, believes that this does not yet constitute a bubble burst, but it has created a "crack" in the investment logic.
In his view, the market was originally betting on two things happening simultaneously: models continuing to improve, and both enterprises and users continuously increasing their usage. As long as either of these two factors was strong enough, AI could pass on the benefits to the entire industrial chain, supporting large-scale investments in chips, cloud services, and data centers. However, the latest data indicates that enterprises are not continuing to pay higher prices for the most advanced models.
Cheaper models are taking away demand.
Behind the price decline, on one hand, model manufacturers are voluntarily reducing prices. The article mentions that OpenAI has lowered the prices of GPT-5.6 and Luna by 80% compared to when they were first released, and Anthropic also announced a price cut last month. On the other hand, corporate customers are shifting from cutting-edge models to more affordable and simpler mid-range models.
Khazarian indicates that more and more enterprises are starting to default to using mid-range models such as Terra and Sonnet, as they still offer sufficient performance while being less costly. On the Ramp platform, the top 1% of enterprises with the highest usage currently spend about $7,200 per employee per month, which is significantly lower than the high usage expectations previously mentioned by NVIDIA CEO Jensen Huang. Moreover, this figure is even showing a trend of slowing down.
The article also mentioned that the "_tokenmaxxing_" sentiment, which was popular in spring, has clearly cooled down. As summer approached, companies began to emphasize cost discipline, and some tech firms stopped encouraging the use of high-intensity models internally, instead opting to restrict employees from prioritizing the use of expensive, cutting-edge models.
Price war affects computing power investment
Foreign media believes that token is increasingly becoming more like a standardized commodity rather than a scarce resource, which will weaken the market's valuation support for AI companies and computing power infrastructure. Morgan Stanley previously warned that if the price of token cannot be maintained, there may be up to $300 billion in debt financing used to support the construction of new cloud computing power and data centers that could face pressure.
This trend is not unique to the United States. The article states that the fierce price competition among Chinese companies is also driving down global model call prices. Although only 3.6% of enterprises on the Ramp platform use open-source models or Chinese models, the low-price competition in the international market still affects American manufacturers.
From a comparison of manufacturers, since August 1st, the effective price of OpenAI has dropped by 38% to $0.48; Anthropic has dropped by 22% to $0.90. The article suggests that Anthropic still retains a certain pricing power, but OpenAI is striving for more token shares through lower prices, and this advantage gap is narrowing.
Additional information:The Chief Financial Officer Sarah Friar recently stated that the company hopes to gradually move away from charging based on token in the future and shift to charging based on the results of work completed. Based on this, the article suggests that the AI of enterprises may not disappear, but the source of growth could be shifting from higher unit prices to competition for market share and increased usage.
