WEBVTT - Bloomberg Businessweek Weekend - August 30th, 2024 0:00:01.360 --> 0:00:05.760 This is Bloomberg business Week inside from the reporters and 0:00:05.920 --> 0:00:09.480 editors who bring you America's most trusted business magazine, plus 0:00:09.560 --> 0:00:13.680 global business finance and tech news. The Bloomberg Business Week 0:00:13.720 --> 0:00:18.680 Podcast with Carol Messer and Tim Stenebeck from Bloomberg Radio. 0:00:20.120 --> 0:00:23.000 Hi, everyone, Welcome to the weekend edition of Bloomberg Business Week. 0:00:23.280 --> 0:00:25.040 We're gonna just kind of cut to the chase here 0:00:25.079 --> 0:00:27.600 because this week has really been so much about what 0:00:27.680 --> 0:00:30.720 is one of the world's largest market cap companies. It's 0:00:30.760 --> 0:00:33.600 the chip maker at the heart of the artificial intelligence boom. 0:00:33.920 --> 0:00:37.240 We are talking about Nvidia, which wrapped up earning season 0:00:37.320 --> 0:00:39.840 for US and gave a revenue forecast that fell short 0:00:39.880 --> 0:00:43.159 of some of the most optimistic estimates, stoking concerns that 0:00:43.200 --> 0:00:46.639 its explosive growth is waning more in the report as 0:00:46.640 --> 0:00:47.640 it was coming out. 0:00:47.720 --> 0:00:50.800 In just a moment, also this hour, Dell and AI, 0:00:51.000 --> 0:00:53.920 we speak to the CTO and chief AI officer at 0:00:53.920 --> 0:00:56.600 the Computer Giant about what's coming. 0:00:56.440 --> 0:00:59.600 Up and later on. Rather than large language models of 0:00:59.720 --> 0:01:02.680 AI h, we talked simply about learning another language with 0:01:02.720 --> 0:01:05.160 the CEO of Babbel. All of that to come. We 0:01:05.240 --> 0:01:08.480 begin with the most anticipated earnings report from this past week, 0:01:08.880 --> 0:01:11.160 and Vidia, the chip maker, failed to live up to 0:01:11.280 --> 0:01:14.920 investor hopes with its latest results this past Wednesday, despite 0:01:15.280 --> 0:01:19.000 beating for the past quarter, but nonetheless, it delivered an 0:01:19.200 --> 0:01:22.319 underwhelming forecast in news of production snags with its hyped 0:01:22.400 --> 0:01:23.319 Blackwell chips. 0:01:23.640 --> 0:01:27.600 The company's quarterly report met or beat analyst estimates on 0:01:27.840 --> 0:01:31.520 nearly every measure, but in Vidia investors have grown accustomed 0:01:31.520 --> 0:01:34.880 to blowout quarters and the latest numbers didn't qualify. After 0:01:34.880 --> 0:01:38.040 the company's earnings call, CEO Jensen Wong sat down exclusively 0:01:38.080 --> 0:01:40.760 with Bloomberg TV's Ed Lovelow to talk about the results 0:01:40.800 --> 0:01:43.880 and how AI is impacting how every layer of computing 0:01:44.040 --> 0:01:44.600 is done. 0:01:44.680 --> 0:01:48.040 I think the market wanted more on Blackwell. They wanted 0:01:48.080 --> 0:01:50.320 more specifics, and I'm trying to go through all of 0:01:50.320 --> 0:01:53.960 the call and the transcript, it seems like a very 0:01:54.000 --> 0:01:56.640 clearly this was a production issue and not a fundamental 0:01:56.680 --> 0:02:00.800 design issue with Blackwell. But the deployment in the real world, 0:02:00.840 --> 0:02:03.280 what does that look like tangibly and is there sort 0:02:03.320 --> 0:02:06.440 of delay in the timeline of that deployment and thus 0:02:06.480 --> 0:02:08.960 revenue from that product. 0:02:10.160 --> 0:02:16.200 I let's see, that's just the fact that I was 0:02:16.240 --> 0:02:18.440 so clear and it wasn't clear enough. 0:02:19.120 --> 0:02:20.720 Kind of tripped me up there right away. 0:02:21.040 --> 0:02:24.280 And so so let's see, we made a mass change 0:02:24.280 --> 0:02:27.880 to improve the yield. Functionality of Blackwell is wonderful. We're 0:02:27.919 --> 0:02:32.440 sampling Blackwell all over the world today. We show people 0:02:33.760 --> 0:02:36.880 giving tours to people of the Blackwall systems that we 0:02:36.919 --> 0:02:37.680 have up and running. 0:02:38.680 --> 0:02:39.560 You could find. 0:02:39.360 --> 0:02:43.400 Pictures of Blackwall systems all over the web. We have 0:02:43.520 --> 0:02:48.680 started volume production. Volume production will ship in Q four. 0:02:49.240 --> 0:02:54.880 Q four, we will have billions of dollars of Blackwell revenues, and. 0:02:56.200 --> 0:02:59.040 We were ramped from there. We were ramped from there. 0:02:59.480 --> 0:03:04.800 The demand for Blackwell far exceeds its supply, of course 0:03:04.880 --> 0:03:08.640 in the beginning, because the demand is so great. But 0:03:08.919 --> 0:03:11.560 we're going to have lots and lots of supply and 0:03:11.960 --> 0:03:15.080 we will be able to ramp. Starting in Q four, 0:03:15.800 --> 0:03:18.840 we have billions of dollars of revenues, and we'll ramp 0:03:18.880 --> 0:03:20.840 from there into Q one, into Q two and two 0:03:20.960 --> 0:03:22.840 next year. We're going to have a great next year 0:03:22.880 --> 0:03:23.240 as well. 0:03:24.440 --> 0:03:27.679 Jensen, what is the demand for accelerated computing beyond the 0:03:27.760 --> 0:03:29.519 hyperscalers and meta. 0:03:31.080 --> 0:03:35.840 Hyperscalers represent about forty five percent of our total data 0:03:35.880 --> 0:03:36.720 center business. 0:03:36.720 --> 0:03:38.560 We're relatively diversified today. 0:03:39.160 --> 0:03:44.440 We have Hyperscalers, we have Internet service providers, we have 0:03:45.040 --> 0:03:53.920 Sovereign AIS, we have industries enterprises, so it's fairly fairly diversified. 0:03:54.520 --> 0:03:59.480 A site outside of Hyperscalers is the other fifty five percent. 0:04:00.440 --> 0:04:02.840 Now, the application. 0:04:02.520 --> 0:04:05.840 Use across all of that, all of that data center 0:04:06.520 --> 0:04:11.400 starts with accelerated computing. Accelerated computing does everything, of course, 0:04:11.920 --> 0:04:16.200 from well the models the things that we know about, 0:04:16.200 --> 0:04:20.239 which is generative AI, and that gets most of the attention. 0:04:21.000 --> 0:04:22.560 But at the core we also. 0:04:22.360 --> 0:04:29.719 Do database processing, pre and post processing of data before 0:04:29.839 --> 0:04:35.920 you use it for generative AI, trans coding, scientific simulations, 0:04:36.040 --> 0:04:39.640 computer graphics of course, image processing of course. And so 0:04:39.680 --> 0:04:44.800 there's tons of applications that people use are accelerated computing for, 0:04:45.560 --> 0:04:47.640 and one of them is generative AI. 0:04:48.120 --> 0:04:51.240 And so let's see what else can I say? I 0:04:51.279 --> 0:04:52.359 think that's the. 0:04:52.520 --> 0:04:55.479 Lever jump in Jensen. Please on Sovereign AI. You and 0:04:55.520 --> 0:04:58.279 I've talked about that before, and it was so interesting 0:04:58.320 --> 0:05:02.159 to hear something behind it. In this fiscal year there 0:05:02.200 --> 0:05:04.320 will be low double digit I think you said billions 0:05:04.320 --> 0:05:07.719 of dollars in Sovereign AI sales. But to the lay person. 0:05:07.880 --> 0:05:10.880 What does that mean. It means deals with specific governments, 0:05:10.920 --> 0:05:11.840 if so, where. 0:05:13.320 --> 0:05:20.520 It's not necessarily sometimes it deals with particular regional service 0:05:20.560 --> 0:05:24.200 provider that's been funded by the government. And oftentimes that's 0:05:24.200 --> 0:05:26.839 the case in a case of in the case of Japan, 0:05:26.920 --> 0:05:33.159 for example, the Japanese government came out and offered. 0:05:34.560 --> 0:05:36.200 Subsidies of a. 0:05:36.240 --> 0:05:40.440 Couple of billion dollars I think for several different internet 0:05:40.520 --> 0:05:45.680 companies and telcos to be able to fund their AI infrastructure. 0:05:46.360 --> 0:05:51.120 India has a sovereign AI initiative going and they're building 0:05:51.160 --> 0:05:59.760 their AI infrastructure. Canada, the UK, France, Italy, I'm missing 0:05:59.760 --> 0:06:04.880 some body, Singapore, Malaysia. You know, a large number of 0:06:04.960 --> 0:06:10.600 countries are subsidizing their regional data centers so that they 0:06:10.600 --> 0:06:15.000 could become able to build out their AI infrastructure. They 0:06:15.040 --> 0:06:21.559 recognize that their countries knowledge, their countries data digital data 0:06:21.680 --> 0:06:25.640 is also their natural resource, not just the land they're 0:06:25.640 --> 0:06:27.400 sitting on, not just the air above them. 0:06:27.920 --> 0:06:31.520 But they realize now that their their digital knowledge is 0:06:31.640 --> 0:06:32.479 part of their. 0:06:32.720 --> 0:06:36.760 Natural and national resource and they had to harvest that 0:06:37.440 --> 0:06:42.960 and process that and transform it into their national digital intelligence. 0:06:43.400 --> 0:06:45.480 And so this is a this is what we call 0:06:45.560 --> 0:06:48.000 sovereign AI. You could imagine almost every single country in 0:06:48.040 --> 0:06:52.760 the world will eventually recognize this and build out their 0:06:52.760 --> 0:06:53.720 AI infrastructure. 0:06:54.480 --> 0:06:57.200 Benson, you use the word resource, and that makes me 0:06:57.279 --> 0:07:00.440 think about the energy requirements here. I think the cool 0:07:00.960 --> 0:07:03.520 it's about how the next generation models will have many 0:07:03.600 --> 0:07:06.760 orders of magnitude greater compute needs. But how will the 0:07:06.880 --> 0:07:09.560 energy needs increase And what is the advantage you feel 0:07:09.600 --> 0:07:13.960 in Nvidia has in that sense, Well. 0:07:13.880 --> 0:07:16.880 The most important thing that we do is increase the 0:07:16.920 --> 0:07:20.960 performance of and increase the performance and efficiency of our 0:07:21.000 --> 0:07:26.200 next generation. So Blackwell is many times more performance than 0:07:26.360 --> 0:07:30.000 Hopper at the same level of power used, and so 0:07:30.080 --> 0:07:34.160 that's energy efficiency, more performance with the same amount of 0:07:34.160 --> 0:07:37.520 power or same performance at a lower power And that's 0:07:37.600 --> 0:07:41.800 number one. And the second is using luca cooling. We 0:07:41.800 --> 0:07:44.800 support air cool we support air cooling, we support liquor cooling, 0:07:44.800 --> 0:07:47.120 but liqual cooling is a lot more energy efficient, and 0:07:47.200 --> 0:07:50.120 so so the combination of that, all of that, you're 0:07:50.160 --> 0:07:50.640 going to get. 0:07:50.520 --> 0:07:52.880 A pretty large, pretty large step up. 0:07:53.600 --> 0:07:56.320 But the important thing to also realize is that AI 0:07:56.400 --> 0:07:58.840 doesn't really care where it goes to school, and so 0:07:59.080 --> 0:08:03.640 increasingly going to see AI be trained somewhere else, have 0:08:03.760 --> 0:08:06.840 that model come back and be used near the population 0:08:07.160 --> 0:08:10.480 or even running on your PC or your phone. And 0:08:10.480 --> 0:08:13.280 so we're going to train large models, but the goal 0:08:13.360 --> 0:08:15.920 is not to run the large models necessarily all the time. 0:08:16.040 --> 0:08:17.880 You can, you can surely do that for some of 0:08:17.880 --> 0:08:23.160 the premium services and the very high value AIS, but 0:08:23.320 --> 0:08:25.840 it's very likely that these large models would then help 0:08:25.920 --> 0:08:28.200 to train and teach smaller models. 0:08:28.600 --> 0:08:30.760 And what we'll end up doing is have one. 0:08:30.680 --> 0:08:34.880 Large, few large models that are able to train a 0:08:34.920 --> 0:08:37.040 whole bunch of small models and they run everywhere. 0:08:38.679 --> 0:08:43.760 Jensen, you explain clearly that demand to build generative AI 0:08:43.800 --> 0:08:46.880 product on models or even at the GPU level is 0:08:46.960 --> 0:08:50.959 greater than current supply in black Vel's case in particular, 0:08:51.200 --> 0:08:54.040 explain the supply dynamics to me for your products and 0:08:54.080 --> 0:08:57.319 whether you see an improvement sequentially quarter on quarter or 0:08:57.360 --> 0:08:59.520 at some point by the end of fiscal year into 0:08:59.520 --> 0:09:00.000 next year. 0:09:01.320 --> 0:09:01.959 Well, the fact that. 0:09:01.920 --> 0:09:06.680 We're growing would suggest that our supply is improving, and 0:09:07.120 --> 0:09:10.520 our supply chain is quite large, one of the largest 0:09:10.520 --> 0:09:14.439 supply chains in the world. We have incredible partners and 0:09:14.760 --> 0:09:17.080 they're doing a great job supporting us in our growth. 0:09:17.600 --> 0:09:20.880 As you know, We're one of the fastest growing technology 0:09:20.920 --> 0:09:23.520 companies in history, and none of that would have been 0:09:23.520 --> 0:09:27.360 possible without very strong demand but also very strong supply. 0:09:27.960 --> 0:09:30.280 We're expecting Q three to have more supply than Q two, 0:09:30.440 --> 0:09:32.760 We're expecting Q four to have more supply than Q three, 0:09:32.840 --> 0:09:34.560 and we're expecting Q one to have more supply than 0:09:34.640 --> 0:09:37.400 Q four. And so I think our supply, our supply 0:09:37.480 --> 0:09:39.920 condition going into next year will be in will be 0:09:40.400 --> 0:09:44.760 a large improvement over this last year with respect to demand. 0:09:46.640 --> 0:09:50.000 Blackwell is just such a leap and there's several things 0:09:50.040 --> 0:09:53.800 that are happening, you know, just the Foundation model makers themselves. 0:09:54.160 --> 0:09:57.600 The size of the Foundation models are growing from hundreds 0:09:57.600 --> 0:10:01.559 of billions parameters to trillions of parameters. They're also learning 0:10:01.600 --> 0:10:05.240 more languages. Instead of just learning human language, they're learning 0:10:05.240 --> 0:10:10.400 the language of images and sounds and videos, and they're 0:10:10.440 --> 0:10:14.400 even learning the language of three D graphics. And whenever 0:10:14.400 --> 0:10:16.920 they are able to learn these languages, they can understand 0:10:16.920 --> 0:10:19.520 what they see, but they can also generate what they're 0:10:19.559 --> 0:10:22.760 asked to generate. And so they're learning the language of 0:10:22.800 --> 0:10:26.880 proteins and chemicals and physics. You know, it could be fluids, 0:10:26.960 --> 0:10:30.120 and it could be particle physics, and so they're learning 0:10:30.160 --> 0:10:33.480 all kinds of different languages. They learn the meaning of 0:10:33.480 --> 0:10:36.160 what we call modalities, but basically learning the languages. 0:10:36.640 --> 0:10:39.440 And so these models are. 0:10:39.320 --> 0:10:42.800 Growing in size, they're learning from more data, and there 0:10:42.800 --> 0:10:45.920 are more model makers then there was a year ago. 0:10:46.200 --> 0:10:48.920 And so the number of model makers have grown substantially 0:10:48.960 --> 0:10:52.080 because of all these different modalities. And so that's just one, 0:10:52.200 --> 0:10:56.120 just the frontier model. The foundation model makers themselves haven't 0:10:56.160 --> 0:11:00.160 really grown tremendously. And then the generative AI mark it 0:11:00.160 --> 0:11:04.320 has really diversified, you know, beyond the Internet service makers 0:11:04.360 --> 0:11:09.880 to startups and now enterprises are jumping in. Different countries 0:11:09.920 --> 0:11:12.199 are jumping in, so the demand has really grown. 0:11:12.880 --> 0:11:15.240 Jensen, I'm sorry to cut you off. I will lose 0:11:15.440 --> 0:11:19.720 your time soon. You've also diversified. And when I said 0:11:19.720 --> 0:11:21.520 to our audience you were coming on, I got so 0:11:21.559 --> 0:11:25.200 many questions. Probably the most common one is what is 0:11:25.240 --> 0:11:28.240 in Nvidia. We talked about you as a systems vendor, 0:11:28.480 --> 0:11:31.840 but so many points on in Nvidia GPU cloud, and 0:11:31.880 --> 0:11:35.120 I want to ask, finally, do you have plans to 0:11:35.160 --> 0:11:37.440 become literally a cloud compute provider? 0:11:39.520 --> 0:11:39.600 No. 0:11:41.200 --> 0:11:45.640 Our GPU cloud was designed to be the best version 0:11:46.559 --> 0:11:51.160 of Nvidia cloud that's built within each cloud. Nvidio DGX 0:11:51.200 --> 0:11:58.760 cloud is built inside GCP, inside Azure, inside AWS, inside OCI, 0:11:59.480 --> 0:12:02.360 and so we build our clouds within theirs so that 0:12:02.440 --> 0:12:03.839 we can implement. 0:12:03.400 --> 0:12:05.320 Our best version of our cloud. 0:12:05.679 --> 0:12:09.880 Work with them to make that cloud, that infrastructure, that 0:12:09.880 --> 0:12:14.680 AI infrastructure and video infrastructure has performance as great TCO 0:12:14.760 --> 0:12:20.600 as possible, and so that strategy has worked incredibly well. 0:12:20.960 --> 0:12:24.440 That's in Nvidia CEO Jensen Wong speaking exclusively with Bloomberg 0:12:24.440 --> 0:12:27.440 Technology co host ed Ludbow. Here The full conversation on 0:12:27.480 --> 0:12:31.440 the Bloomberg Talks podcast feed, available everywhere you download your podcasts. 0:12:31.559 --> 0:12:34.680 Coming up, another tech company writing the AI wave, Dell 0:12:34.960 --> 0:12:37.640 recheck in with the CTO and Chief AI Officer at 0:12:37.679 --> 0:12:40.160 Dell Technologies. Is that on the other side, you're. 0:12:40.000 --> 0:12:42.760 Listening to Bloomberg Business Week. This is Bloomberg. 0:12:46.640 --> 0:12:50.160 You're listening to the Bloomberg Business Week podcast. Catch us 0:12:50.200 --> 0:12:53.439 live weekday afternoons from two to five pm Easter listen 0:12:53.480 --> 0:12:56.560 on Applecarplay and then brout Auto with a Bloomberg Business 0:12:56.559 --> 0:13:00.120 act or watch us live on YouTube. 0:13:00.360 --> 0:13:02.680 So a few weeks ago, Dell announced they are getting 0:13:02.760 --> 0:13:05.520 leaner cutting jobs as part of a reorganization of its 0:13:05.600 --> 0:13:08.120 sales team to focus on a new group centered on 0:13:08.360 --> 0:13:12.120 AI products and services. You may recall earlier this year 0:13:12.440 --> 0:13:16.760 that CEO Michael Dell talked about aipcs being pretty standard 0:13:17.000 --> 0:13:18.120 in twenty twenty five. 0:13:18.600 --> 0:13:21.440 We wanted an update about what Dell is doing for that, 0:13:21.520 --> 0:13:25.000 Bloomberg News reporter Katie Graifeld and I turned to John Rose. 0:13:25.040 --> 0:13:29.760 He's CTO and Chief AI Officer at Dell Technologies. We 0:13:29.800 --> 0:13:33.079 spoke to him before the release of Dell's latest earnings report. 0:13:33.280 --> 0:13:35.640 CTOs have been around forever, but chief AI officers, at 0:13:35.720 --> 0:13:38.719 least at Dell and really in federal government and a 0:13:38.760 --> 0:13:40.320 few other places started about a year ago. 0:13:40.640 --> 0:13:40.800 You know. 0:13:40.840 --> 0:13:44.080 It was a recognition that this is complex, it's bigger 0:13:44.120 --> 0:13:46.760 than an IT project. It has to have somebody caring 0:13:46.760 --> 0:13:48.600 and feeding it because it was a lot of governance 0:13:48.760 --> 0:13:51.480 and technology coming together. So, you know, for us, it 0:13:51.520 --> 0:13:53.480 was a year ago that we put the position in place. 0:13:53.600 --> 0:13:55.760 We've talked about the Chief AI Officer in the context 0:13:55.760 --> 0:13:58.960 of the federal government and the new regulations around federal 0:13:59.000 --> 0:14:02.400 agencies needing a chief AI officer. Do you think that 0:14:02.440 --> 0:14:04.480 every company needs a Chief AI Officer? 0:14:05.480 --> 0:14:08.400 I don't I think some companies need chief AI officers 0:14:08.440 --> 0:14:11.280 if you're doing an at scale kind of AI transformation 0:14:11.320 --> 0:14:14.119 of a large organization, But if if you're a smaller organization, 0:14:14.360 --> 0:14:17.679 you need someone who's driving it. But do you need 0:14:17.679 --> 0:14:20.520 to change your organizational structure? For a company like Dell, 0:14:20.760 --> 0:14:23.200 we have literally hundreds of not thousands, of projects that 0:14:23.240 --> 0:14:25.240 are going to happen, and a thirty percent of our 0:14:25.280 --> 0:14:26.800 work is probably going to shift below. 0:14:26.640 --> 0:14:27.360 The machine line. 0:14:27.520 --> 0:14:31.160 That is a gigantic effort, and having someone in the 0:14:31.200 --> 0:14:34.880 position from an executive perspective is necessary. But smaller companies 0:14:34.880 --> 0:14:37.760 are going to depend on companies helping them navigate this 0:14:37.880 --> 0:14:39.680 rather than creating organizational structures. 0:14:39.840 --> 0:14:42.480 Well, I'm curious in your role how you juggle timelines 0:14:42.560 --> 0:14:45.600 because you think about the here and now feels like 0:14:45.680 --> 0:14:48.240 AI has to be in every single conversation. You have 0:14:48.280 --> 0:14:51.960 a ton of companies trying to gain a foothold or 0:14:52.040 --> 0:14:53.960 get up to speed or play catch up. But then 0:14:54.000 --> 0:14:58.200 you also have this fantastic long term opportunity that we 0:14:58.240 --> 0:15:00.800 talk about all the time that no one really knows 0:15:00.840 --> 0:15:03.480 what that holds. So I mean, in your day to day, 0:15:03.480 --> 0:15:05.560 what does it look like trying to juggle those two. 0:15:06.000 --> 0:15:06.200 Yeah. 0:15:06.240 --> 0:15:09.080 You know, as a CTO, my job is to make 0:15:09.120 --> 0:15:11.600 sure we never miss a technology inflection. So things like 0:15:11.680 --> 0:15:13.920 quantum live under me, and these are long term places. 0:15:14.080 --> 0:15:16.360 One of the interesting things about AI, though, is that 0:15:16.480 --> 0:15:20.640 the future is happening now. There is no five year 0:15:20.760 --> 0:15:23.359 journey that you're going to get to eventually. The disruptions 0:15:23.400 --> 0:15:28.240 are happening almost weekly, and so this intersection between execution 0:15:28.560 --> 0:15:32.080 and vision is more pronounced than any other technology I 0:15:32.120 --> 0:15:34.880 have ever seen. So, you know, good chief AI officer 0:15:34.960 --> 0:15:37.880 is not just an operational manager. They actually have to 0:15:37.920 --> 0:15:40.640 be extremely current with the technology and what is emerging 0:15:40.680 --> 0:15:42.800 so that they don't miss an inflection which could happen 0:15:42.840 --> 0:15:44.400 in the span of like a matter of weeks. 0:15:44.560 --> 0:15:46.000 I'm trying to figure out the way to ask this, 0:15:46.080 --> 0:15:48.320 to contextualize it in the right way. But when we 0:15:48.360 --> 0:15:51.960 talk to people about AI, John, we oftentimes hear them 0:15:52.080 --> 0:15:54.760 say it's just like the Internet in the nineteen nineties. 0:15:55.080 --> 0:15:58.960 You're saying no, no, on the scale of impact it is. 0:15:59.120 --> 0:16:01.760 But I was in creating that that year. 0:16:01.880 --> 0:16:03.680 I was the CTL a bunch of networking companies in 0:16:03.720 --> 0:16:07.440 the nineties, and we had a decade and a half 0:16:07.640 --> 0:16:10.800 to kind of get there, and it happened slowly over time, 0:16:11.000 --> 0:16:13.040 it did have a profound impact. I think from a 0:16:13.040 --> 0:16:16.200 scale perspective it's similar. But the speed in which this 0:16:16.240 --> 0:16:18.080 is happening, think about it, you know, let me give 0:16:18.080 --> 0:16:19.880 you an example. It would be technical for a second. 0:16:20.000 --> 0:16:23.000 There's a term out there called retrieval augmented generation RAT, 0:16:23.040 --> 0:16:25.200 which now is in the normal of vernaculars. 0:16:24.880 --> 0:16:27.160 Adie and Univer say that to me. I say it 0:16:27.200 --> 0:16:27.760 all the time. 0:16:28.280 --> 0:16:30.200 Yeah, we were just talking about. 0:16:29.960 --> 0:16:33.600 In the commercial ok so common term very important. 0:16:33.600 --> 0:16:36.400 Technology allows you to take your external data and use 0:16:36.400 --> 0:16:38.360 it with a large language model without the large language 0:16:38.360 --> 0:16:39.080 model absorbing it. 0:16:39.120 --> 0:16:40.120 That's pretty important, right. 0:16:40.440 --> 0:16:43.720 That term, as a discussion in our industry is about 0:16:43.760 --> 0:16:47.320 a year old. It's foundational to almost every enterprise AI 0:16:47.480 --> 0:16:49.920 story that you can imagine. The technology didn't exists three 0:16:50.000 --> 0:16:53.040 years ago. So the speed in which these things are happening, 0:16:53.200 --> 0:16:55.640 you know, it's not like we've developed the Internet, then 0:16:55.640 --> 0:16:57.320 we have ten years to figure it out and mature 0:16:57.320 --> 0:16:59.080 it and then scale it and twenty years later everybody 0:16:59.120 --> 0:17:02.440 has to be there. This is if you're not there now, 0:17:02.520 --> 0:17:05.440 or if you're not moving towards getting there, you're already late. 0:17:05.960 --> 0:17:08.080 And I haven't seen that happen in any other big 0:17:08.119 --> 0:17:09.240 technology and FI is it. 0:17:09.160 --> 0:17:14.680 More steam engine like Jamie Diamond talks about Industrial revolution. 0:17:15.800 --> 0:17:19.600 Yeah, it's after the steam engine existed and was commercially available. 0:17:19.760 --> 0:17:22.600 What you saw was this massive uptick you had. And 0:17:22.640 --> 0:17:24.960 it's actually a good analogy because the steam engine was 0:17:25.000 --> 0:17:28.520 a way to transfer energy, and you need to transfer 0:17:28.600 --> 0:17:30.960 energy to do all kinds of things, from producing goods 0:17:30.960 --> 0:17:33.960 and services to doing transportation, and so it became incredibly pervasive. 0:17:34.000 --> 0:17:35.040 AI is very similar. 0:17:35.040 --> 0:17:38.600 Anybody who tries to piginol that it only does chatbots for, 0:17:38.720 --> 0:17:42.720 you know, searching staff for it only does enterprise applications X, Y, 0:17:42.760 --> 0:17:46.040 and Z is wrong because it's a foundational technology. It's 0:17:46.080 --> 0:17:49.280 a different way to distill information into knowledge and wisdom, 0:17:49.320 --> 0:17:51.439 and you're going to apply it everywhere, much like we 0:17:51.480 --> 0:17:53.359 did when we moved to the internal combustion engine. 0:17:54.040 --> 0:17:57.600 Well, let's talk about AI as it relates to Dell's business, 0:17:57.640 --> 0:17:59.800 because of course you've made an aggressive pivot when it 0:17:59.840 --> 0:18:03.320 comes to the AI server space. How different is the 0:18:03.400 --> 0:18:07.960 manufacturing and sales versus the traditional server manufacturing. 0:18:08.080 --> 0:18:12.119 Yeah, first thing, maybe to Dispelomith, we didn't actually do 0:18:12.200 --> 0:18:15.760 an aggressive pivot, meaning about four or five years ago, 0:18:15.840 --> 0:18:18.280 we built our first AI service Wait before Jenai, and 0:18:18.320 --> 0:18:20.639 the reason for it was that there were deep neural networks, 0:18:20.680 --> 0:18:24.080 there were early AI systems. Nobody understood them except maybe 0:18:24.080 --> 0:18:27.040 one hundred thousand people or smart enough, and so we 0:18:27.080 --> 0:18:29.119 had been building these systems for quite a while and 0:18:29.200 --> 0:18:31.720 predicting that there would be a time where this would explode, 0:18:31.720 --> 0:18:33.880 this would be something that would be bigger. So when 0:18:34.440 --> 0:18:37.240 CHATGPD happened and suddenly Jennai was available, we had this 0:18:37.400 --> 0:18:40.960 large language model explosion. We actually already had products that 0:18:41.000 --> 0:18:42.560 were ready to go into the market, and that's one 0:18:42.560 --> 0:18:44.639 of the advantages we had as this market formed. We 0:18:44.640 --> 0:18:47.560 were actually already there. And then what we had to 0:18:47.640 --> 0:18:50.280 do is adapt a little bit. But our supply chain 0:18:50.359 --> 0:18:52.439 was in place, our service organization was in place. We 0:18:52.480 --> 0:18:54.119 had to kind of optimize who we were selling to 0:18:54.160 --> 0:18:57.119 because it was new customers, but it wasn't really a 0:18:57.119 --> 0:19:00.520 gigantic surprise. The other thing that we have now included 0:19:00.760 --> 0:19:06.360 is that there isn't significant difference between the two markets 0:19:06.359 --> 0:19:07.600 in AI that we participate in. 0:19:07.640 --> 0:19:09.159 The first is the training market. 0:19:09.160 --> 0:19:11.360 These large scale deals that we tend to be winning 0:19:11.400 --> 0:19:14.320 a lot of and are very interesting. And then the enterprise, 0:19:14.359 --> 0:19:16.359 which is actually not a training business, it's more of 0:19:16.359 --> 0:19:19.280 an inference business. It turns out that the products that 0:19:19.280 --> 0:19:22.080 we're building in the training site at scale transfer almost 0:19:22.119 --> 0:19:24.760 literally into the enterprise unmodified, which is a very good 0:19:24.760 --> 0:19:27.080 thing because we can address two markets with a portfolio. 0:19:27.359 --> 0:19:29.199 Hey, John, one thing that we're really interested in is 0:19:29.320 --> 0:19:33.000 understanding who your AI server customers are today. We did 0:19:33.040 --> 0:19:37.720 see Michael Dell's tweet on shipping too Musk's XAI initiative. 0:19:38.119 --> 0:19:40.399 How would you define your AI server customers? Is its 0:19:40.440 --> 0:19:43.080 second tier cloud that's accounting for the growth. Where are 0:19:43.160 --> 0:19:46.320 the traditional enterprise when it's coming to AI adoption. 0:19:46.600 --> 0:19:48.840 Well, first, there's two very distinct markets. There is the 0:19:48.880 --> 0:19:52.880 traditional enterprise, which is still an early market. It's mostly experimentation. 0:19:52.960 --> 0:19:55.680 Those experiments are now graduating and giving an example internally, 0:19:55.680 --> 0:19:58.160 we have a project that was first an experiment around 0:19:58.200 --> 0:20:00.879 our service organization, creating some that would allow people to 0:20:00.920 --> 0:20:04.800 make intelligent choices about troubleshooting a problem. We then very 0:20:04.880 --> 0:20:07.399 quickly move from a prototype to thousands of people, and 0:20:07.440 --> 0:20:09.600 we're on a path the potentially forty thousand of our 0:20:09.600 --> 0:20:12.320 service technicians using this technology and a span of literally 0:20:12.359 --> 0:20:13.320 measured in six. 0:20:13.160 --> 0:20:16.200 Months or so. So the enterprise is on that journey. 0:20:16.240 --> 0:20:18.800 We haven't really hit scale in any meaningful way, but 0:20:18.840 --> 0:20:20.679 we have a lot of things in the pipeline as 0:20:20.680 --> 0:20:23.760 people are figuring it out. That market, i would say, 0:20:23.800 --> 0:20:26.800 is my view, is a bigger market because there are 0:20:26.840 --> 0:20:30.040 millions of enterprises and every process and every enterprise is 0:20:30.080 --> 0:20:32.879 a target for this technology. But it takes time to 0:20:32.920 --> 0:20:35.119 get to a point of stability, to transfer the knowledge 0:20:35.160 --> 0:20:38.600 and too basically get moving great future. The today market 0:20:38.640 --> 0:20:41.040 is mostly characterized by the ones you mentioned. It's companies 0:20:41.040 --> 0:20:43.960 that are building their own foundation models. They are training 0:20:44.040 --> 0:20:46.800 large scale models. That's not what enterprises do. Enterprises, you 0:20:46.920 --> 0:20:50.520 use those models to do inference tasks. The training business 0:20:50.680 --> 0:20:55.359 is growing dramatically because quite frankly, that's the underlying engine 0:20:55.359 --> 0:20:55.919 of all layout. 0:20:55.960 --> 0:20:57.560 But is it only a handful of companies that are 0:20:57.560 --> 0:20:57.760 doing it. 0:20:57.880 --> 0:20:59.719 It's not a very big market in terms of compared 0:20:59.760 --> 0:21:02.159 to the tens of millions of enterprises that are going 0:21:02.200 --> 0:21:03.560 to use size that are going to use it. The 0:21:03.560 --> 0:21:05.760 people that are going to produce these models is not 0:21:05.880 --> 0:21:08.960 anywhere near that number. There's more than a handful of them. 0:21:09.520 --> 0:21:11.679 There are certain tiers of them, as you mentioned, there 0:21:11.720 --> 0:21:15.159 are even certain enterprise, very large enterprises, Xtra being a 0:21:15.160 --> 0:21:18.480 good example. You're talking about the Yeah, they're also training 0:21:18.480 --> 0:21:20.600