Pricing ChatGPT Pro at $200/mo filters it to only power users/enterprise, and given the cost of the GPT-o1 API, it wouldn't surprise me if those power users burn through $200 worth of compute very, very quickly.
> We have guardrails in place to help prevent misuse and are always working to improve our systems. This may occasionally involve a temporary restriction on your usage. We will inform you when this happens, and if you think this might be a mistake, please don’t hesitate to reach out to our support team at help.openai.com using the widget at the bottom-right of this page. If policy-violating behavior is not found, your access will be restored.
Source: https://help.openai.com/en/articles/9793128-what-is-chatgpt-...
Especially when the base line profit margin is negative to begin with
I find this difficult believe, although I don't doubt leaks could have implied it. The challenge is that "the cost of compute" can vary greatly based on how it's accounted for (things like amortization, revenue recognition, capex vs opex, IP attribution, leasing, etc). Sort of like how Hollywood studio accounting can show a movie as profitable or unprofitable, depending on how "profit" is defined and how expenses are treated.
Given how much all those details can impact the outcome, to be credible I'd need a lot more specifics than a typical leak includes.
I can't find any sources _not_ mentioning billions of loss for 2024 and for the foreseeable future
And the 'search' part of it could use many of these clusters in parallel, and then pick the best answer to return to the user.
...they probably quantise a bit, but not loads, as they don't want to sacrifice performance. FP8 seems like a possible middle ground. o1 is just a bunch of GPT-4o in a trenchcoat strung together with some advanced prompting. GPT-4o is theorised to be 200B parameters. If you wanted to run 5 parallel generation tasks at peak during the o1 inference process, that's 5x 200B, at FP8, or about 12 H100s. 12 H100s takes about one full rack of kit to run.