00:00First off, just your comments on the recent sell-off in the AI plays in Korea.
00:04What do you make of that? Was there an overreaction, you think?
00:07I don't think it's not an overreaction.
00:09So basically, the demand is rare.
00:12So right now, the new market is looking for the new market, especially the inference domain.
00:18NVIDIA is doing very well.
00:20Basically, NVIDIA built for training workload and done that brilliantly.
00:24I respect that.
00:25However, the gravity of AI compute has shifted from training to inference,
00:29where AI model is actually running energy efficient and cost effectively.
00:34So NVIDIA GPU is not designed for it.
00:36So market is looking for the new market winners.
00:39So when you talk about new market winners, Korea possibly stands a good chance of being those new market winners.
00:46Yet, is there a sense that there's an underappreciation of the role Korea is playing in the global AI infrastructure?
00:53I agree with that 100% because especially the LLM and current models,
00:57it's highly, highly optimized for the memory.
01:00It requires a lot of memory, HBM, as you may know.
01:03So future architects are going to be the memory-centric.
01:07Basically, Rebellion is so part of a unique combination of strategic investors that include SK Hynik, Samsung, ARM, and Aramco.
01:15Right now, we are collaborating with the memory vendor and the Samsung founder to fully leverage Korea's excellent semi-econ
01:23ecosystem in the Manpower.
01:24As you mentioned, it's underestimated.
01:26Right now, it's going up.
01:27As you can see, the Korean market, especially SK Hynik and Samsung, went up, went up.
01:31That's what I think.
01:33How do you see the picture changing in perhaps two to three years?
01:36In two to three years, definitely, we're going to have each sector have each market winner.
01:41Training and automation is NVIDIA still there.
01:43However, in the influence market, the new market winner is Korean, is a strong candidate.
01:47Rebellion is one of them, I totally believe.
01:50Why?
01:51Again, influence requires fundamentally different engineering trade-off between the latency, throughput, and energy efficiency.
01:58Rebellion is a second generation of AI chip companies.
02:01It's built for inference alone, but the previous first generation in GPU built for training and trying to retrofit training
02:11hardware for inference.
02:12It's a really important time because the market is looking for the new players.
02:16You talked about how you're in a lot of partnerships, including Samsung, ARM, and so on and so forth.
02:22Why do you need these partnerships?
02:25Is that the only way startups can survive when they're rivaling the big guys?
02:30That's the right question to ask.
02:32Basically, it's more than partnership.
02:34Our strategic investors, they invest in us thanks to our low-power feature and architecture.
02:39As you may know, currently, the supply chain of the semiconductors is, you know, screw up.
02:43It's not very easy to get the HBM, not very easy to get just the DMP for the server.
02:49Also, semiconductor industry requires huge capital and the manpower, not survive as long, as many.
02:55It's just the ecosystem, not a single player.
02:58Even for the United States, family startup needs a huge support from the SK Hynix and Samsung.
03:03That's why we are very focused on the ecosystem, the part of the ecosystem, especially in Korea.
03:09But a lot of startups actually do not want those partnerships to maintain margins.
03:14Is that a consideration for you?
03:16No, not at all.
03:17So our strategy goal is going to be the part of the ecosystem.
03:22Now, we talk about your chip.
03:24They say that your chip can deliver three times power efficiency versus NVIDIA's H100.
03:31Talk to us about that.
03:32And if that's really the case, why aren't the hyperscalers converting?
03:37Very good question.
03:38Basically, I'm not saying that we are better than NVIDIA.
03:40In terms of functional diversity and the flexibility, NVIDIA's GPU, general purpose GPU is better than us.
03:47However, when it comes to inference alone, inference alone, we can offer much higher energy efficiency and performance at the
03:53same time.
03:54That's why we are now collaborating with hyperscale and big lab even in the United States.
03:59That's why we became the unicorn just in four years.
04:02So you are in talks with hyperscalers for their conversion to use your chip.
04:07Who are some of these hyperscalers that you're in talks with?
04:10Yeah, I'm very sure this is because it's strictly limited.
04:13But maybe next time you're going to be the first one to talk about our big name later.
04:17But again, it's a really huge demand, especially in code assistance or something like that.
04:22So within the next, what, three to five months, you'll be making an announcement about the conversion of these hyperscalers?
04:28Is there a time frame that we can be looking at?
04:30Yeah, definitely less than four months or something like that.
04:34Our POC is successfully gone.
04:36Yes, huge demand again.
04:38Right now, our bottleneck is supply chain, not the demand from the customer.
04:42Talk to us about the roadmap for Rebellions to be beyond a Korean player.
04:48What are your ambitions and what can be expected in the next several quarters?
04:53Definitely.
04:54We are basically originally built for being a global player.
04:58Basically, I spent 11 years in the United States, five years for MIT PhD.
05:01Another six years in SpaceX, Morgan Stanley.
05:04We originally built for global player.
05:06But Korea gave us fundamentally proof points, our reference, you know, it validated our technology at scale.
05:15Basically, our two customers, SK Telecom and Korea Telecom, top two data center companies in Korea gave us a reference.
05:22Based on that value of reference, we are going global actively.
05:26One of the examples is Saudi Arabia, Inhumane, and Aramco.
05:30Where are the other markets which will be imported for you?
05:32Basically, we're not able to everywhere at once, but we're looking for the sovereign market.
05:37So, we're looking for the sovereign AI is very important.
05:38One example is Japan and Saudi Arabia.
05:40They are trying to build heterogeneous data center where NVIDIA and non-NVIDIA co-exist together.
05:47And United States, non-United States supplies co-exist together due to geopolitical landscape and the contention.
05:54So, we are looking for Saudi Arabia, for example.
05:57In Japan, also looking for the emerging countries in Asia.
06:01That's why I'm here and I attended the meeting yesterday and I am at the event.
06:05What's the messaging that you are sending to the people here as you take part in this conference?
06:10I was saying that the emerging countries face real risk, just become, could become just consumer of AI.
06:19You know, with very sensitive data flowing beyond the border and sovereignty compromise.
06:24Each nation, each organization deserves the ability to build and control the AI, especially in the infrastructure side.
06:31That's why you need to build, you know, heterogeneous compute platform, NVIDIA and NVIDIA together, especially GPGP for training and
06:40influence for another purpose, build custom, right, rebellion, in order to diversify the supply chain.
06:46That's the beginning of a sovereign AI infrastructure.
06:49We talk about how Rebellion is trying to break the dominance of NVIDIA.
06:52How can that be done?
06:53I mean, this is one company that controls 90% of the AI accelerated market.
06:58How do you break that control?
06:59It's not healthy at all, right?
07:01Because it's the beauty of the market.
07:03Market demand is huge.
07:04Even though NVIDIA is doing very well, again, it's the first generation GPUs optimized for training at the same time.
07:10It's not 100% optimized for inference.
07:13Again, the gravity of AI confidence shifted.
07:16So the market looking for the inference again.
07:19So, yeah, that's our go-to-market strategy.
07:21It's the beauty of the market AI.
07:23And, of course, when you talk about Korea, there's so much focus now and a lot of excitement when it
07:28comes to the AI space.
07:28You have the likes of Samsung, Hynix, Naver, Kakao.
07:32Yet, when you take a look at the AI platform, something's missing.
07:37Why is that?
07:38Why is it not happening?
07:39Oh, AI platform in Korea is pretty excellent, but I'm not that expert.
07:43I'm not a hardware guy.
07:44But right now, a lot of government commitment to Korean AI ecosystem.
07:49So, foundation model through the silicon.
07:52I think in the future, I'm going to share more detail, but I'm not that much expert.
07:56Okay.
07:56This, I'm sure you can answer.
07:58Are you planning for an IPO?
07:59When might that be?
08:00What are your thoughts on fundraising?
08:02So, basically, we recently completed a series of C. It's raising almost half billion USD.
08:09So far, at the valuation, 1.5 billion USD, we have enough resource to serve our customer.
08:14But, as you may know, Bloomberg report saying that we just got a contract with JP Morgan two days ago.
08:20We kicked off the preparation.
08:22I'm very sorry to nothing specific to share with you today.
08:25But, in the future, please invite me again.
08:27I'm going to be happy to share all the details and valuation timeline.
08:30So, no further fundraising plans for the moment.
08:31Is that it?
08:32Oh, we will plan one more project, but not very sure yet.
08:35But, again, yeah.
08:36In the market, in public, it's really good.
08:39So, no reason to delay or something like that.
08:42But, again, no specific plan yet.
Comments