Also isn't an open weight model also subsidised? Training isn't cheap and you are not paying for it.
Also isn't an open weight model also subsidised? Training isn't cheap and you are not paying for it.
We also have OpenAI numbers, where people speculate that the about 150% of the margins spent on "marketing" is a fake line used to hide operational costs.
In an interview earlier this year I remember Dario saying that the models are profitable, ie they more than pay back their inference and training over time. But because they invest in that explosive growth they have a deep negative cash burn.
And whatever markup the AI labs make, that’s on top of nvidia’s markup, micron’s markup, etc. It’s an industry where every supplier is adding a 70-80% markup!
For OpenAI, absolutely not, that's not including training costs. But I do personally accept that it can be actual marketing costs.