OpenAI inference costs near unsustainable
- OpenAI’s cost problem is shifting from training models to serving them: as ChatGPT usage surges, inference is becoming the bigger recurring expense. - OpenAI said its weekly user base doubled to more than 800 million in seven months, while most users remain free and each query still consumes compute. - That backdrop is driving OpenAI’s push for more data centers, power and enterprise revenue. (openai.com)
Artificial intelligence models cost money twice: first to train them, then to answer every prompt users send. OpenAI’s strain now sits increasingly in that second bucket. (openai.com 1) (openai.com 2) OpenAI said in a White House submission published in late 2025 that its weekly user base had doubled in seven months to more than 800 million people, with most using its tools for free. More users means more inference, the industry term for the compute used every time a model generates an answer. (openai.com) That matters because inference is not a one-time capital hit. It is the metered cost of running live products, and it rises with every chat, image request, coding session and voice interaction. (openai.com) (cdn.openai.com) OpenAI’s own public policy papers have become unusually direct about the bottleneck. The company says AI leadership will depend on chips, power generation, transmission lines and data-center capacity, not just better model research. (cdn.openai.com) (openai.com) The company has been lining up more infrastructure accordingly. OpenAI announced a multi-year, $38 billion partnership with Amazon Web Services in November 2025 to scale core workloads, and it has separately promoted Stargate campuses in Texas and abroad. (openai.com 1) (openai.com 2) Pricing has been moving in the opposite direction from total demand. OpenAI’s current API pricing page shows cheaper small-model options alongside premium frontier models, a sign that per-token prices can fall even while total serving costs climb with volume. (openai.com) That tension helps explain why consumer scale does not automatically translate into healthy margins. If hundreds of millions of people use free or lightly priced products, a lower cost per token can still produce a larger aggregate bill. (openai.com 1) (openai.com 2) It also helps explain OpenAI’s emphasis on business customers. The company has been expanding enterprise sales and infrastructure partnerships while arguing publicly for faster buildout of American power and data-center projects. (openai.com) (openai.com) Some outside estimates put the cost burden in the billions, but many of the most detailed figures circulating online come from secondary analysis or paywalled reporting rather than audited public filings. OpenAI has not publicly broken out a clean, current line item for total inference spend. (sacra.com) (theinformation.com) The core point is less disputed than the exact number. For OpenAI and its rivals, the next fight is no longer only who can train the smartest model, but who can afford to run it at global scale. (openai.com) (cdn.openai.com)
