00:00You have used M&A as a strategy.
00:04You have very interesting infrastructure where you're running H100s principally upgrading to Blackwell Generation.
00:10Yes.
00:10Those are very specific to the modalities of image and audio.
00:15That's right.
00:15Right.
00:16But outside of that infrastructure, you've gone out buying companies.
00:19Are you going to use some of the money raised for that purpose?
00:22Absolutely.
00:22Yes.
00:23Why?
00:23What's the logic here?
00:24Definitely.
00:24I mean, it's definitely a very fast race right now.
00:29We're now trying to become one of the largest companies right now in the space.
00:36And one of the ways we get talent is buying companies.
00:40So M&A is definitely a very good way of doing that.
00:44Well, Kai, I want to go back a step because you claim to have the fastest inference stack for modern generative models, multimodal.
00:51Yes.
00:52How?
00:52What are you doing to remove the technical economic barriers that you say exist when deploying GPUs?
00:57Yeah, absolutely.
00:59So what we have is we have this inference engine that is a proprietary engine that can run image and video models up to three to four times faster.
01:07And we do that by optimizing NVIDIA chips to run these workloads.
01:12So this is a proprietary technology that we've built in-house.
01:16It's software that you've built?
01:18It is software.
01:19That's right.
01:19You know, that's the interesting debate in the world we're in.
01:22If you look at some of the ASICs and custom silicon solutions, the concern is they don't have the library of software to go with it.
01:30But we're here to talk about Fold, but would you just reflect on the reality of needing to use NVIDIA for that reason?
01:36But also, you've built your own software.
01:38Absolutely, yeah.
01:39So I'll give you a quick summary of the dilemma there.
01:43So we could run our workloads on ASICs, but the scale at which we operate is a lot larger, and there's no real ASIC out there that has the scale.
01:53That's one.
01:54And then the second point is that one of our key value props is being day zero.
01:59And if you have a model and you want to run that model whenever it's available, the first day it's available, you actually have to have, like, really good software that's already ready to deploy.
02:11With ASICs, typically what you have to do is actually customizing the software a lot before you can actually launch it.
02:18So with NVIDIA, the software stack is so much more mature, so we can actually do these day zero releases.
02:25So this is one of the key reasons why we prefer NVIDIA right now.
02:28Bakai, we now have the news that maybe NVIDIA will be able to ship its H200s to China.
02:34How easily accessible is the supply that you need right now?
02:38Is that a concern for you, that sort of power going to China as well?
02:42Right.
02:43So for our workloads specifically, for image and video, H200s are a little bit of overkill for us, at least for the hopper generation.
02:52So we actually prefer H100s, and then H200s are more preferred for, like, language models.
02:58Which is not our domain.
03:00But when we're actually moving away from the hopper architecture to blackwell architecture, we're actually just going directly from H100s to B200s or B300s.
03:11Because that actually gives us more flops per watt or per dollar.
03:16So we're already rolling this out as B200s and B300s are becoming available.
03:22We're actually moving over.
03:24I think, you know, the sort of economics of this is that many of the players are actually going to move to blackwells because you just get better price performance.
03:35We're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to move to blackwells because you're going to blackwells because you're going to blackwells because you're going to blackwells because you're going to blackwells because you're going to blackwells because you're going to blackwells because you're going to blackwells because you're going to blackwells.
Comments