Hacker News Reader

Comment by minimaxir | Hacker News Reader

Sam says Sora 2 has to make money but there is no revenue model that can feasibly offset a $4-5 compute cost per video.

With some back of the napkin math, I am pretty sure you're off by at least two orders of magnitude, conceivably 4. I think 2 cents per video is an upper limit.

https://news.ycombinator.com/item?id=45434298

Generally speaking, API costs that the consumer sees are way higher than compute costs that the provider pays.

EDIT: Upper limit on pure on-going compute cost. If you factor in chip capital costs as another commentator on the other thread pointed out, you might add another order of magnitude.

You also aren’t including amortized training costs, which are immense (and likely ongoing as they continue to tweak the model).

I suspect amortized training costs are only a relatively small fraction of the amortized hardware costs (i.e. counting amortized hardware costs already accounts for the large fraction of the cost of training and pulling out training as a completely separate category double counts a lot of the cost).

Where did you get that figure from?

It’s more a ballpark since exact numbers vary and OpenAI could be employing shenanigans to cut costs, but in comparison, Veo 3 which has similar quality 720p video costs $0.40/second for the user, and Sora’s videos are 10 seconds each. Although Veo 3 could cost more or less to Google than what is charged to the user.

I suspect OpenAI’s costs to be higher if anything since Google’s infra is more cost-efficient.

Sora 1 is slightly more expensive than Veo 3, at $0.50/second: https://azure.microsoft.com/en-us/pricing/details/cognitive-... (select "US East 2" as region; not all regions have Sora).

Though again, caveat: Those are API prices. The real prices could be both lower (if there is a profit margin) or higher (if they are subsidized).

It was revealed to them in a dream.

Reply on news.ycombinator.com