Shawn Wang: prediction

Transcript context

…Yeah, I think that's kind of like the main data war story developing. GPU poor versus GPU rich. Yeah. Where do you think we are? I think there was, again, going back to like the small model thing, there was like a time in which the GPU poor were kind of like the rebel faction working on like these models that were like open and small and cheap. And I think today people don't really care as much about GPUs anymore. You also see it in the price of the GPUs. Like, you know, that market is kind of like plummeted because there's people don't want to be, they want to be GPU free. They don't even want to be poor. They just want to be, you know, completely without them. Yeah. How do you think about this war? You can tell me about this, but like, I feel like the, the appetite for GPU rich startups, like the, you know, the, the funding plan is we will raise 60 million and [00:39:00] we'll give 50 of that to NVIDIA. That is gone, right? Like, no one's, no one's pitching that. This was literally the plan, the exact plan of like, I can name like four or five startups, you know, this time last year. So yeah, GPU rich startups gone. But I think like, The GPU ultra rich, the GPU ultra high net worth is still going. So, um, now we're, you know, we had Leopold's essay on the trillion dollar cluster. We're not quite there yet. We have multiple labs, um, you know, XAI very famously, you know, Jensen Huang praising them for being. Best boy number one in spinning up 100, 000 GPU cluster in like 12 days or something. So likewise at Meta, likewise at OpenAI, likewise at the other labs as well. So like the GPU ultra rich are going to keep doing that because I think partially it's an article of faith now that you just need it. Like you don't even know what it's going to, what you're going to use it for. You just, you just need it. And it makes sense that if, especially if we're going into. More researchy territory than we are. So let's say 2020 to 2023 was [00:40:00] let's scale big models territory because we had GPT 3 in 2020 and we were like, okay, we'll go from 1. 75b to 1. 8b, 1. 8t. And that was GPT 3 to GPT 4. Okay, that's done. As far as everyone is concerned, Opus 3. 5 is not coming out, GPT 4. 5 is not coming out, and Gemini 2, we don't have Pro, whatever. We've hit that wall. Maybe I'll call it the 2 trillion perimeter wall. We're not going to 10 trillion. No one thinks it's a good idea, at least from training costs, from the amount of data, or at least the inference. Would you pay 10x the price of GPT Probably not. Like, like you want something else that, that is at least more useful. So it makes sense that people are pivoting in terms of their inference paradigm. And so when it's more researchy, then you actually need more just general purpose compute to mess around with, uh, at the exact same time that production deployments of the old, the previous paradigm is still ramping up, um, uh, pretty aggressively. So [00:41:00] it makes sense that the GPU rich are growing. We have now interviewed both together and fireworks and replicates. Uh, we haven't done any scale yet. But I think Amazon, maybe kind of a sleeper one, Amazon, in a sense of like they, at reInvent, I wasn't expecting them to do so well, but they are now a foundation model lab. It's kind of interesting. Um, I think, uh, you know, David went over there and started just creating models. Yeah, I mean, that's the power of prepaid contracts. I think like a lot of AWS customers, you know, they do this big reserve instance contracts and now they got to use their money. That's why so many startups. Get bought through the AWS marketplace so they can kind of bundle them together and prefer pricing.…

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