4 ms·

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.

by minimaxir 1y ago

nojs 1y ago
Where did you get that figure from?
- minimaxir 1y ago
  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.
  - majewsky 1y ago
    Sora 1 is slightly more expensive than Veo 3, at $0.50/second: https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/ 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).
- nvr219 1y ago
  It was revealed to them in a dream.
dwohnitmok 1y ago
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 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.
- rafram 1y ago
  You also aren’t including amortized training costs, which are immense (and likely ongoing as they continue to tweak the model).
  - dwohnitmok 1y ago
    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).

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.

nojs 1y ago

Where did you get that figure from?

minimaxir 1y ago
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.
- majewsky 1y ago
  Sora 1 is slightly more expensive than Veo 3, at $0.50/second: https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/ 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).
nvr219 1y ago
It was revealed to them in a dream.

dwohnitmok 1y ago

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 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.

rafram 1y ago
You also aren’t including amortized training costs, which are immense (and likely ongoing as they continue to tweak the model).
- dwohnitmok 1y ago
  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).