I’m not entirely sure, it’s my understanding that later this year a lot of compute will go online with the latest hardware that will be significantly cheaper for inference.
The problem is rather, I think, that people always want to use the latest and greatest models. And that training is super expensive.
Potentially we’ll just see less new model releases.
The problem is rather, I think, that people always want to use the latest and greatest models. And that training is super expensive.
Potentially we’ll just see less new model releases.