00:00Joining us now is SambaNova CEO Rodrigo Liang. He is a direct competitor of Cerebris. And Rodrigo, why don't you
00:08explain to us first exactly where you compete and where you differ.
00:14Because Ed Ludlow was telling us earlier that Cerebris is in a number of different businesses. They could be on
00:20the one hand a chip maker.
00:21They put together the servers and they also could be a neocloud company of sorts.
00:27Yeah, thanks for having me. Yeah, we're a chip maker. We build AI chips for inference.
00:34And so we're very focused on driving inference into the highest performance space.
00:39And so if you look at Cerebris where they're demonstrating that they can actually run really, really fast on smaller
00:45models, we run the spectrum.
00:46We think that the battleground is going to be in the large models and where the frontier models are.
00:51And even the open source models are now trillion parameter models. We're very focused on taking those big models and
00:57running them really, really fast.
00:59Rodrigo, I just wonder how much the world has woken up to this idea that inference is about to get
01:03really expensive.
01:04And it's a problem that Sambanova obviously solves for. Just think about Uber a couple weeks ago that blew through
01:10their entire inference budget for the year in just a couple months.
01:14It was something like 500 to 2,000 per engineer per month on AI API calls.
01:21I just wonder, are we about to experience this kind of second wave of disruption for these companies that are
01:27using AI that maybe the models haven't fully passed along the cost of just how expensive it is yet?
01:34Yeah, well, as we know, the AI market is going to be dominated by inference.
01:38And now you're seeing why, because the cost structure of using these models continues to rise.
01:43And so the need for more cost-effective solutions, the need for faster solutions that allow you to actually compute
01:49faster is really important.
01:51And so we're driving that change as well, that this wave of really efficient computing, fast and energy efficient, is
01:58going to be incredibly important, driving that cost down.
02:01So who gets access to your chips, right?
02:06There are now a number of smaller chip makers that solve for specific problems like Sambanova, and not everyone can
02:15have them, right?
02:16Your supply constrains.
02:17So how do you decide who gets them and who doesn't?
02:23Well, that's a key thing.
02:24Supply chain is a battleground, and we're already taking orders into next year.
02:28But the really exciting thing about this right now is as we actually make these models more efficient and we
02:34map them against chips like ours, where we can run them 10 times faster than, say, a traditional GPU, you
02:41can make more tokens available to people.
02:43And so today we're driving them into enterprises, we're driving them into software and clouds, because we think that the
02:49market is much broader than just the frontier models or just the hyperscalers.
02:54And so we love the collaboration we're making with some of these big model companies, but we actually want to
02:59see this broadly adopted into the enterprises and into many, many countries.
03:04Roger Ego, just given the supply crunch for a lot of this stuff and just the rate that you're building
03:10out, how deep are your capital needs?
03:13You obviously had recently U.S. regulators clearing the way for Intel to increase its investment in Sambanova.
03:19You have an investment from Vista, the private equity company.
03:22You see your rivals like Cerebris going to public markets in order to raise more.
03:27OpenAI CFO is saying our capital needs are really, really intense.
03:31Just how intense are yours right now?
03:34Yes, I mean, Sambanova, look, the growth is just so tremendous.
03:36And so we're in a transformative phase for the entire globe.
03:42And for all the excitement that you see, the enterprises, if you think about the banks and the retailers, they're
03:48just getting going, right?
03:50They're just getting the transitions to AI going.
03:53And so the buildup continues and it's going to actually go for many more years.
03:57We'll see you next time.
03:57We'll see you next time.
03:57We'll see you next time.
03:57We'll see you next time.
03:57We'll see you next time.
Comments