As OpenAI negotiates an unprecedented funding round of up to $100 billion, it is proud of a number of key optimizations at the business operations level. Although the total cost of computing power remains high as a proportion of revenue, the company is achieving higher revenue returns for every cent of server operating costs invested in supporting the ChatGPT subscription business and selling model usage rights to enterprise customers.
OpenAI’s computing power margin for paying users — the share of revenue left after deducting the cost of running artificial intelligence models for paying users — has jumped from about 52% at the end of last year and about 35% in January 2024 to about 70% in October this year, according to a person familiar with the company’s financials.
In contrast, competitor Anthropic’s computing power profit margin last year was about negative 90%, according to an analysis of its financial data by The Information. The analysis shows Anthropic is on track to lift that margin to about 53% by the end of the year, and its most optimistic forecast suggests it could hit 68% next year.
However, overall, Anthropic's forecast data shows that if the operating costs of artificial intelligence models for non-paying users and the cost of model training are taken into account, the company's overall server operating efficiency will surpass OpenAI. OpenAI has hundreds of millions of free chatbot users. In the future, it will need to commercialize these users through advertising or shopping rebates to make up for this efficiency gap. The number of free users of Anthropic’s chatbot is much smaller.
Anthropic estimated this summer that the company’s total investment in computing power may reach as high as $60 billion between 2025 and 2028, including the cost of research and development of new artificial intelligence models. This forecast does not take into account recent server leasing agreements with Google and Microsoft. Over the same period, OpenAI expects to spend $220 billion on servers.
OpenAI’s huge investment in servers reflects the core judgment of the company’s leadership: the shortage of server supply is the biggest challenge restricting the company’s development and the road to general artificial intelligence. General artificial intelligence refers to artificial intelligence that has the ability to automatically complete most tasks of high economic value.
"Without sufficient computing power support, we simply cannot advance related research and development. We are currently facing a serious computing power bottleneck." The company's CEO Sam Altman said in a podcast interview released on Thursday. "If the computing power scale is doubled, I think the company's current revenue can also be doubled." Altman did not elaborate further on this.
For potential investors interested in investing in OpenAI, they may be eager to pay attention to any signs of improvement in the company's profit margins related to paying users, because the improvement in the profit margin of paid packages is expected to subsidize the operating costs of free users. Even so, the computing power profit margin of 70% is still far lower than the similar profit margin level of listed software companies-these companies can provide services to new users (including free users) at very low cost.
OpenAI also faces another challenge, its biggest chatbot competitor is none other than Google. Google uses a self-developed tensor processing unit, a customized server chip, which effectively reduces operating costs; while OpenAI uses expensive Nvidia server chips. According to "The Information", OpenAI leadership believes that Google's artificial intelligence operation efficiency is much higher than its own. This means that Google will be under much less pressure to monetize free users.
At the beginning of this year, OpenAI prioritized reducing the server running costs (i.e., model inference costs) of artificial intelligence products. According to a person familiar with the matter, in February this year, after DeepSeek launched a new model, OpenAI activated a "red alert" internal emergency mechanism. DeepSeek claims that the training cost of its new model is much lower than proprietary models developed by companies such as OpenAI.
Another person familiar with the matter said that the improvement in OpenAI's computing power profit margin is due to two major factors: first, the continued decline in computing power rental costs throughout the year, and second, the company has optimized and adjusted the artificial intelligence model to make it more efficient. In addition, OpenAI has launched higher-priced subscription service packages to generate more revenue from some customer groups.