Foreign media report that a key assumption underpinning the high valuations and massive capital expenditures in AI is showing signs of weakening: the actual price companies pay for AI model inference capabilities is declining, rather than continuing to rise with each new model release.
Ramp’s enterprise spending data released on Wednesday showed that U.S. businesses paid an effective price of $0.68 per million tokens, a roughly 41% decline from the March high of $1.15. Meanwhile, the share of overall usage attributed to frontier models dropped from about 53% in early August to 45% in September.
Corporate spending is beginning to contract.
The article notes that what deserves the most attention is not just the price correction, but also the shift in spending by high-end customers. Top clients driving revenue for OpenAI and Anthropic saw their per-person AI spending decline by nearly 10% month-over-month in August. Ara Khazarian, Chief Economist at Ramp, believes this does not yet constitute a bubble burst, but it has created a “crack” in the logic of AI investment.
In his view, the market had originally bet on two things happening simultaneously: models continuing to grow stronger, and businesses and users steadily increasing their adoption. As long as one of these factors was strong enough, AI companies could pass on their profits throughout the supply chain, supporting large-scale investments in chips, cloud services, and data centers. However, the latest data shows that businesses have not continued to pay higher prices for the most advanced models.
Cheaper models are stealing demand.
Behind the price decline, on one hand, model providers are proactively lowering prices. The article notes that OpenAI has reduced the price of GPT-5.6 Luna by 80% since its launch, and Anthropic also announced a price cut last month. On the other hand, enterprise customers are shifting from frontier models to more affordable, simpler mid-tier models.
Khazarian said an increasing number of companies are defaulting to mid-tier models like Terra and Sonnet because they offer sufficient performance at lower costs. On the Ramp platform, the top 1% of companies with the most intensive AI usage currently spend about $7,200 per employee per month on AI—significantly lower than the high usage levels previously projected by NVIDIA CEO Jensen Huang—and this figure is still slowing down.
The article also noted that the once-popular "tokenmaxxing" sentiment in spring has significantly cooled. As summer arrived, companies began emphasizing cost discipline, and some tech firms have discontinued internal practices that encouraged heavy model usage, instead restricting employees to prioritize lower-cost models over expensive, cutting-edge ones.
The price war impacts hash power investment
Foreign media believe that tokens are increasingly resembling standardized commodities rather than scarce resources, which undermines the market's valuation support for AI companies and computing infrastructure. Morgan Stanley previously warned that if token prices cannot be sustained, up to $300 billion in debt financing used to support the construction of new cloud computing and data centers could come under pressure.
This trend is not unique to the United States. The article states that intense price competition among Chinese AI companies is also driving down global model invocation prices. Although only 3.6% of businesses on the Ramp platform use open-source or Chinese models, low-price competition in international markets still exerts downward pressure on U.S. providers.
From a vendor comparison perspective, since August 1, OpenAI’s effective price has dropped 38% to $0.48, while Anthropic’s price has fallen 22% to $0.90. The article suggests that Anthropic still retains some pricing power, but OpenAI is gaining market share in tokens through lower pricing, narrowing the advantage gap.
Additional information: OpenAI’s Chief Financial Officer, Sarah Friar, recently stated that the company hopes to gradually move away from token-based pricing toward charging based on the outcomes of completed tasks. The article concludes that while AI companies’ revenues may not disappear, their growth drivers are shifting from price increases to competition for market share and higher usage volumes.