experienced that at Google multiple times. >> Yeah. So Nvidia GPUs are great for training, but at inference time they're memory bottlenecked. How does Grock change the memory architecture to address that? >> Well, first it's important to think about the trade-offs. Um there is no such thing as a free launch. And um what you're trying to do is you're trying to get the lowest cost per token because the cost actually determines your capacity, right? Everyone's in a race. people would pay more for more capacity, but if I'm spending twice as much um or if I'm spending uh you know twice as much, I'm getting about half as much of the capacity per dollar, right? That's that's really what I care about. Can I get this many tokens per dollar? Um the thing is you also need speed. And so the the trade-off is if you want the absolute best cost per, you know, token, you're just going to use a GPU as it is. you're going to use a very large batch size and it's not going to be as fast as it otherwise could be. Uh what we did with the LPU was we were able to scale to multiple chips without using um any external memory and spread the model out across those chips. So we could use much faster SRAMM which allowed us to generate tokens faster but at a lower cost. And so if you know what a paro curve is, uh, and you'll have to explain that to the audience potentially. If you look at the paro curve of GPUs versus LPUs, they're quite different. And there are portions of of the curve where the GPU is better economically, and there are portions where the LPU is better economically, usually at the the faster end of it. When you put the two together, it fills out that middle zone. And so between the GPU, the GPU plus LPU and the LPU, you now have the best cost per token. You have the most capacity at any speed that you want to run at.