Vipul Ved Prakash: belief

Transcript context

…By your customers. You know, either on our platform or off our platform. And we are generally seeing that, you know, that is the sort of trend where you can get better quality on your task by sort of now easily adapting these models to your data. We also have, I would say, over 20 big model builds happening on the platform, which are customer. We see a lot of training and it's also somewhat surprisingly a more continuous kind of workload. We sort of imagine that this would be more episodic. You train a model and then you do inference. But what we find is, you know, we train a model and then they train the next version and then the next version, which sort of grows in scale. I would say training is still the bigger portion. Some ways inference is super linear to model quality. And as the models are getting better, there's more and more inference. Oh, because they're more useful. Yeah, they're more useful, yeah. So, okay, so training is bigger. This is actually consistent with what we've heard from Mosaic, that, you know, people think that training is sort of like a one-time deal. You do one big run and then you're done. It's never true. And so I'm interested in, like, putting some numbers and I don't know what you have disclosed or what you want to disclose, but, like, how many GPUs do you have? What is the equivalent amount of compute that you have? Because I understand that your GPU setup is different than what people typically think of, like, a giant data center somewhere, right?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.