Skip to content
#AI

How high are OpenAI’s compute costs? Possibly a lot higher than we thought (AI column)

#AI #Economics #Management #Future #ML

It is always interesting to count money in someone else’s pocket, but even more interesting when these calculations allow you to assess the convergence of the current wave of hype around AI. In fact, the Financial Times did this by writing. article motively parsing Ed Zitron, who collected data on OpenAI accounts in Azure. In the FT article, the authors tried to estimate the real costs of inference and get comments from both companies: OpenAI and Microsoft.

If you describe the main points of the article, We are talking only about bones on inference, and the cost of LLM training is not taken into account here. This means that it is about the operating costs of processing user requests through the UI and through the API. And capex is just about training. Cheers. 7 OpenAI could spend more than $ on inference in Azure12,4 Billions in revenue estimates of about $6,8 billion Every dollar earned is worth approximately 2 dollar

  • To the first. 6 months 2025 Inference reportedly ate almost $US5 billion, on revenue of about $4,3 Billion and previously announced “total” cash burn in $2,5 Billions in the same period. It turns out that the work of models alone is more expensive than the entire bone is officially recognized. The complexity of the assessment is due to the financial relationship between OpenAI and Microsoft, which resembles a recursion. OpenAI reportedly gives up to 20Microsoft shares a percentage of Microsoft's revenue, and Microsoft shares a portion of Azure/Bing's revenue with OpenAI. It turns out a closed circuit where money runs in a circle, and the outside observer can only guess where the real margin is. Microsoft and OpenAI numbers are not hard to deny. The FT showed them flattened estimates: the answer was in the style of “the numbers are not exactly accurate”, but without alternative calculations. If you just extrapolate the trend, somewhere to 2033 This year, the total revenue will only begin to cover one reference cost. And if you take into account the share of Microsoft, in some scenarios it does not cover them at all.

As a result, in the current conditions, the convergence of the business model is questionable, and when the situation will change is not yet clear.

#AI #Economics #Management #Future #ML