That's all different now with agentic which was not really a big thing until the end of 2024. before they were doing 1 request, now they're doing hundreds for a given task. the reason oai/azure win over locally run models is the parallelization that you can do with a thinking agent. simultaneous processing of multiple steps.
You hit the nail on the head. Just gotta add the up to $10 billion investment from Microsoft to cover pretraining, R&D, and inference. Then, they still lost billions.
One can serve a lot if models if allowed to burn through over a billion dollars with no profit requirement. Classic, VC-style, growth-focused capitalism with an unusual, business structure.
Imagine pipelineing lots of infra-scale GPU's, naive inference would need all previous tokens to be shifted "left" or from the append-head to the end-of-memory "tail", which would require a huge amount of data flow for the whole KV cache etc. Instead of calling GPU 1 the end-of-memory and GPU N the append-head, you keep the data static and let the role rotate like a circular buffer. So now for each new token inference round, the previous rounds end-of-memory GPU becomes the new append-head GPU. The highest bandwidth is keeping data static.
Inference contributes to their losses. In January 2025, Altman admitted they are losing money on Pro subscriptions, because people are using it more than they expected (sending more inference requests per month than would be offset by the monthly revenue).
At the end of the day, until at least one of the big providers gives us balance sheet numbers, we don't know where they stand. My current bet is that they're losing money whichever way you dice it.
The hope being as usual that costs go down and the market share gained makes up for it. At which point I wouldn't be shocked by pro licenses running into the several hundred bucks per month.
Currently, they lose more money per inference than they make for Pro subscriptions, because they are essentially renting out their service each month instead of charging for usage (per token).
When an end user asks ChatGPT a question, the chatbot application sends the system prompt, user prompt, and context as input tokens to an inference API, and the LLM generates output tokens for the inference API response.
GPT API inference cost (for developers) is per token (sum of input tokens, cached input tokens, and output tokens per 1M used).
Again, this means that the average ChatGPT Pro end user's chattiness cost OpenAI too much inference (too many input and output tokens sent and received, respectively, for inference) per month than would be balanced out by OpenAI receiving $200/month in revenue from the average Pro user.
The analogy is like Netflix losing money on their subscriptions because their users watch too much streaming, so they ban account sharing, causing many users to cancel their subscriptions, but this actually helps them become profitable, because the extra users using their service too much generated more costs than revenue.
I think you maybe have misunderstood the parent (or maybe I did?). They're saying you can't compare an individual's cost to run a model against OpenAI's cost to run it + R&D. Individuals aren't paying for R&D, and that's where most of the cost is.
