Editing Openai/69408f2e-3acc-800c-a0c5-7ab1f3f9f65c (section)

==== Because OpenAI keeps detailed costs private, analysts use industry math and usage patterns: ====

===== Independent estimates suggest OpenAI’s core real-time inference infrastructure (e.g., Sora/ChatGPT service) might cost on the order of tens of millions of dollars per day: =====
* A popular independent analysis put ongoing infrastructure cost (GPUs + overheads like networking, power, storage) around ~$12–$17 million per day — roughly $4–$6 billion per year just to keep large models running. Jeff Bruchado<ref>{{cite web|title=Jeff Bruchado|url=https://jeffbruchado.com.br/en/blog/openai-sora-cost-15-million-day-ai-infrastructure|publisher=Jeff Bruchado|access-date=2025-12-16}}</ref>

===== Another estimate projects daily inference expenditure based on usage could be hundreds of thousands to millions per day for supporting billions of tokens served. Economía Digital<ref>{{cite web|title=Economía Digital|url=https://www.digital-economy.org/modules/3/ai-infrastructure-llm-economics|publisher=Economía Digital|access-date=2025-12-16}}</ref> =====

===== Training a flagship large model is very expensive: =====
* Industry data shows training a frontier LLM like GPT-4 could cost tens to hundreds of millions of dollars, often >$100 M per major run. Hakia<ref>{{cite web|title=Hakia|url=https://www.hakia.com/tech-insights/cost-of-ai/|publisher=hakia.com|access-date=2025-12-16}}</ref>
* Larger next-gen models being developed now could run into hundreds of millions or more per training cycle as scale grows. arXiv<ref>{{cite web|title=arXiv|url=https://arxiv.org/abs/2405.21015|publisher=arxiv.org|access-date=2025-12-16}}</ref>