WEBVTT - Digesting DeepSeek's AI Impact 0:00:00.800 --> 0:00:05.040 From the heart of where innovation, money and power collide 0:00:05.320 --> 0:00:10.600 in Silicon Valley and beyond. This is Bloomberg Technology with 0:00:10.760 --> 0:00:12.240 Caroline Hyde. 0:00:11.840 --> 0:00:13.240 And Ed Ludlow. 0:00:26.200 --> 0:00:28.520 Live from New York On Caroline Hyde and. 0:00:28.400 --> 0:00:31.760 I'm Mike Shepard in San Francisco. This is Bloomberg Technology 0:00:31.920 --> 0:00:32.400 coming up. 0:00:32.720 --> 0:00:35.479 Deep Seek Day two, we dive into the ripple effects 0:00:35.479 --> 0:00:39.080 in tech markets and the impact on US chip cubs lust. 0:00:39.120 --> 0:00:41.800 We discuss deep Seek's AI model and the future of 0:00:41.880 --> 0:00:46.040 open source with hugging faces. Chief AI scientist and President 0:00:46.040 --> 0:00:49.240 Trump says Microsoft is in talks to acquire the US 0:00:49.400 --> 0:00:52.199 unit of TikTok. Details later this ow but first we 0:00:52.280 --> 0:00:56.160 check in on the markets and we claw our way back. Mike, 0:00:56.360 --> 0:00:59.600 not nearly off setting yesterday's significant sell off. We're up 0:00:59.640 --> 0:01:02.600 just the edge point on the overall NASDAC. Look, Apple 0:01:02.640 --> 0:01:04.559 once again doing heavy lifting to the point side. 0:01:04.600 --> 0:01:07.320 Meta at a new record high, but not enough because. 0:01:07.000 --> 0:01:10.440 In video bounces back but two percent after a seventeen 0:01:10.440 --> 0:01:13.679 percent set off yesterday, but thirty billion being added from 0:01:13.680 --> 0:01:15.920 the market cap that was wiped out by six. 0:01:15.760 --> 0:01:17.160 Hundred billion dollars yesterday. 0:01:17.360 --> 0:01:21.440 This is all surrounding China coming with clearly a very powerful, 0:01:21.560 --> 0:01:24.480 very efficient, very cheap, open source model NIC. 0:01:25.560 --> 0:01:28.920 And yesterday in video called deep Seek's new model quote 0:01:29.120 --> 0:01:34.560 an excellent AI advancement that complies with US technology export controls, 0:01:34.800 --> 0:01:37.920 and that deep Seek's work illustrates how new models can 0:01:37.959 --> 0:01:40.880 be created. Let's bring in Bloomberg, Zee and King now 0:01:40.880 --> 0:01:44.120 to tell us more about this ian. This was a 0:01:44.240 --> 0:01:47.840 rough day yesterday for in Vidia. It bore the brunt 0:01:47.960 --> 0:01:50.280 of the deep Seak shock across the markets. Yet their 0:01:50.360 --> 0:01:54.720 statement really was curious. Tell us what your interpretation was. 0:01:54.800 --> 0:01:56.400 Yeah, I mean, there are a couple of things going 0:01:56.440 --> 0:01:59.040 on here. I mean they could have obviously said nothing. 0:01:59.080 --> 0:02:01.040 They could have just ignored it. They could have said, 0:02:01.240 --> 0:02:04.040 maybe it's kind of dodgy technology that we don't really 0:02:04.040 --> 0:02:06.000 believe in, but they actually endorsed it and said, no, 0:02:06.160 --> 0:02:10.280 this is good. And the sort of underpinning message we 0:02:10.320 --> 0:02:13.120 saw here was that maybe they're telling the US. 0:02:13.040 --> 0:02:15.880 Government, your export controls haven't worked. We should be allowed 0:02:15.880 --> 0:02:17.000 to ship whatever chips we. 0:02:16.919 --> 0:02:20.000 Want to China because guess what, they'll do good stuff 0:02:20.040 --> 0:02:21.320 with whatever they can get hold of. 0:02:21.960 --> 0:02:25.639 Let's just go to that narrative on chip exports potentially 0:02:25.680 --> 0:02:26.280 not working. 0:02:26.600 --> 0:02:28.519 We heard from President Trump. 0:02:28.280 --> 0:02:30.080 Just yesterday about what the impact could be. 0:02:30.160 --> 0:02:30.760 Just take a listening. 0:02:31.800 --> 0:02:34.280 In a very near future, we're going to be placing 0:02:34.360 --> 0:02:38.880 tariffs and farm production of computer chips, semiconductors, and pharmaceuticals 0:02:38.880 --> 0:02:42.480 to return production of these essential goods to the United 0:02:42.520 --> 0:02:46.239 States of America. They left US and they went to Taiwan, 0:02:46.400 --> 0:02:49.720 which is about ninety eight percent of the chip business 0:02:49.720 --> 0:02:52.720 by the way, and we want them to come back. 0:02:54.240 --> 0:02:56.240 Of course, a lot of in video chips are manufactured 0:02:56.280 --> 0:02:58.560 in Taiwan, and so there's going to be an impact there. 0:02:58.840 --> 0:03:00.640 What do you think the regulatory and pat will be 0:03:01.520 --> 0:03:03.520 let alone where the inferencing is really going to be 0:03:03.560 --> 0:03:04.920 the future for GPU demand? 0:03:05.880 --> 0:03:08.760 Yeah, I mean this is the big question here. And 0:03:08.800 --> 0:03:12.160 the expectation I think before the inauguration was that this 0:03:12.240 --> 0:03:15.720 administration would, if anything, double down on this kind of 0:03:15.840 --> 0:03:19.520 crackdown of chips and the flow to China. And here 0:03:19.560 --> 0:03:21.760 we're getting a different aspect of it. So I think 0:03:21.800 --> 0:03:24.720 we'll have to see how much attention the new administration 0:03:24.880 --> 0:03:27.799 is placed is pacing on this issue. But the expectation 0:03:27.960 --> 0:03:29.120 is that they're not going to let. 0:03:29.040 --> 0:03:30.800 Up em back to Nvidia. 0:03:30.919 --> 0:03:34.920 How much is this complicating their efforts to get their 0:03:34.960 --> 0:03:38.640 customers to keep buying all those premium chips like Blackwell 0:03:38.800 --> 0:03:41.200 and the next generations that you've been writing about so much. 0:03:41.280 --> 0:03:42.840 Yeah, well, it creates a massive question. 0:03:42.880 --> 0:03:45.760 It's like, well, if a Chinese startup can fine tune 0:03:45.760 --> 0:03:48.720 their models to this extent using sort of two three 0:03:48.800 --> 0:03:51.520 year old technology on the cheap, why can't you write 0:03:51.640 --> 0:03:55.000 so if you're Microsoft, if you're a WUS and you're 0:03:55.200 --> 0:03:57.960 running all of these systems, and why wouldn't you do 0:03:58.040 --> 0:04:01.120 it cheaper if you can? So obviously has implications for 0:04:01.200 --> 0:04:03.160 what in video is trying to sell. So we'll see 0:04:03.160 --> 0:04:04.400 how they answer those questions. 0:04:04.680 --> 0:04:06.560 Thank you so far the statement is all we get 0:04:06.600 --> 0:04:09.640 in king, We thank you. Now let's just talk about 0:04:09.680 --> 0:04:12.680 how the tech markets, of course roiled by china seemingly cheap, 0:04:12.720 --> 0:04:15.840 powerful AI model. Here's what some Bloomberg Television guests had 0:04:15.840 --> 0:04:16.800 to say about deep Seek. 0:04:17.839 --> 0:04:20.680 I don't think it's about China versus US, right, It's 0:04:20.760 --> 0:04:26.080 really about closed source versus open source and more more, 0:04:26.160 --> 0:04:30.200 American developers should leverage the open source and what deep 0:04:30.240 --> 0:04:31.960 Seak had done and built on top of it. 0:04:31.880 --> 0:04:34.240 And make it greater. Really amazing what they have done. 0:04:34.400 --> 0:04:38.279 The fraction of training cost, fraction of interference costs to 0:04:38.279 --> 0:04:41.000 produce a model like that in deep Seek, but it 0:04:41.120 --> 0:04:43.279 will unravel basically the path of spending. 0:04:43.400 --> 0:04:47.279 Of course, I think a healthy dose of skepticism is 0:04:47.480 --> 0:04:52.520 very very good, it's needed. But even if you multiply 0:04:52.960 --> 0:04:57.200 their spend and their cost by ten times two D time, 0:04:57.240 --> 0:05:00.560 this is still a order and maltitude in terms of 0:05:00.640 --> 0:05:01.440 theotocopy action. 0:05:01.680 --> 0:05:04.920 They're not really a company that's out there trying to 0:05:04.960 --> 0:05:08.520 compete directly with the big US players. So in that sense, 0:05:08.560 --> 0:05:11.680 I think it was an overreaction to the release of 0:05:11.720 --> 0:05:12.240 this model. 0:05:13.440 --> 0:05:14.320 Let's get more reaction. 0:05:14.440 --> 0:05:18.040 Mardin Alton, now chief investment strategist at Empower. The question 0:05:18.160 --> 0:05:22.440 for many is will capital expenditure remain for in video chips, 0:05:22.600 --> 0:05:24.320 for AI hardware writ large. 0:05:24.520 --> 0:05:26.040 Is that what you were tackling yesterday. 0:05:27.080 --> 0:05:29.760 Well, I think that is one of the critical questions, right, 0:05:29.800 --> 0:05:33.000 because this was a very surgical sell off, a very 0:05:33.080 --> 0:05:36.320 surgical development in the sense that it's attacking really not 0:05:36.480 --> 0:05:40.880 the case for using AI, but really the supply chain 0:05:40.960 --> 0:05:44.080 around AI and all roads on supply chain lean back 0:05:44.320 --> 0:05:47.280 to Nvidia, and so this question of whether companies really 0:05:47.320 --> 0:05:49.279 need to spend this much. I think that is the 0:05:49.360 --> 0:05:51.680 question that markets are grappling with. But it's hard to 0:05:51.720 --> 0:05:54.239 imagine that we're going to see, and of course we'll 0:05:54.279 --> 0:05:56.240 see in earnings this week, but it's hard to imagine 0:05:56.240 --> 0:05:59.880 that we're going to see companies massively change their spending 0:06:00.320 --> 0:06:04.640 patterns simply because of this news. I think there's still 0:06:04.680 --> 0:06:07.599 this arms race going on, So of course that's something 0:06:07.600 --> 0:06:10.520 that we're going to have to watch especially closely. But 0:06:10.680 --> 0:06:13.760 we may need more than just this news to unravel that. 0:06:14.279 --> 0:06:18.760 But the unraveling could start when you have earnings expectations 0:06:18.800 --> 0:06:22.040 as they are and valuations at forty one times future 0:06:22.040 --> 0:06:25.560 earnings for a namelike in video. How much you worried 0:06:25.560 --> 0:06:28.520 that this could have a similarity to what we saw 0:06:28.520 --> 0:06:29.360 in the dot com era. 0:06:30.400 --> 0:06:33.520 I mean, that's really something that we're watching really closely. 0:06:33.560 --> 0:06:36.560 When you're looking at earnings expectations for the broad market overall, 0:06:36.800 --> 0:06:39.480 you're seeing an acceleration of earnings and of course maybe 0:06:39.520 --> 0:06:42.480 not the same for MAGS seven, but certainly a continuation 0:06:42.640 --> 0:06:45.120 of the powerful earnings that we've seen there. And then 0:06:45.160 --> 0:06:48.280 you couple that with valuations that are extreme. When we 0:06:48.360 --> 0:06:51.440 do our analysis of valuations and we break the time 0:06:51.520 --> 0:06:54.320 series for our price to expected earnings down into death siles. 0:06:54.600 --> 0:06:56.880 We're looking at the tech sector that's in the ninetieth 0:06:56.880 --> 0:07:01.520 percentile of valuations. Now associating that with forward one year returns, 0:07:01.560 --> 0:07:05.120 you can still see valuations grind higher and companies perform. 0:07:05.240 --> 0:07:08.240 But ultimately that's where you also start to see some 0:07:08.320 --> 0:07:11.480 of these more meaningful selloffs. And as pedantic as it sounds, 0:07:11.560 --> 0:07:14.640 I think a lot of these developments over the past 0:07:14.640 --> 0:07:19.000 few days relate more to evaluation argument than anything that 0:07:19.200 --> 0:07:23.720 relates to kind of technology in the sense that this 0:07:23.800 --> 0:07:26.040 is big news and technology, but the bigger news is 0:07:26.040 --> 0:07:28.800 that there's no room in the narrative to account for 0:07:28.880 --> 0:07:31.360 anything but world domination when it comes to MAGS seven. 0:07:31.600 --> 0:07:33.640 And this is allowing us to think about the range 0:07:33.640 --> 0:07:35.920 of outcomes Martin. 0:07:35.960 --> 0:07:38.520 To play the movie forward a few years. Where do 0:07:38.560 --> 0:07:41.080 you see this taking us? Is this a parallel to 0:07:41.120 --> 0:07:45.040 the dot com era, which you know, eventually, after some 0:07:45.720 --> 0:07:48.920 troubles right around two thousand, really set us up for 0:07:49.120 --> 0:07:50.679 a long period of growth. 0:07:50.880 --> 0:07:52.880 What about AI? Are we in that moment? 0:07:53.880 --> 0:07:56.520 Well, that's that's the corollary that I reference. I know 0:07:56.600 --> 0:07:58.600 a lot of us are pointing back to the dot 0:07:58.640 --> 0:08:01.920 Com era and looking at evaluation argument. I think that's important, 0:08:02.120 --> 0:08:04.520 But I think the other important element is that this 0:08:04.640 --> 0:08:07.600 is an innovation supercycle. It feels like just the way 0:08:07.640 --> 0:08:09.880 the dot Com era was, and what we saw in 0:08:09.880 --> 0:08:14.840 that period was a massive expansion or growth for the 0:08:14.880 --> 0:08:17.560 broad market, not just for the tech area. So if 0:08:17.560 --> 0:08:20.440 we're looking at the trend growth from the nineteen nineties 0:08:20.480 --> 0:08:23.120 and comparing that to the realized trend growth of the 0:08:23.320 --> 0:08:27.440 subsequent years, we saw massive expansion, whether that's margins or 0:08:27.560 --> 0:08:31.640 enterprise value. That Internet world had a huge impact on 0:08:31.680 --> 0:08:33.679 the broad economy, and I think that's what we can 0:08:33.720 --> 0:08:36.280 expect to see with AI. And in many ways, this 0:08:36.440 --> 0:08:39.199 news around deepseek is part of that in the sense 0:08:39.240 --> 0:08:41.120 that it says we can economize, we can do this 0:08:41.200 --> 0:08:43.480 a lot better, and we can get deployment a lot 0:08:43.480 --> 0:08:45.920 more quickly than maybe we had otherwise thought. 0:08:46.559 --> 0:08:50.920 And on that idea of getting cheaper AI out to 0:08:51.040 --> 0:08:54.840 companies and businesses across the economy maybe a little bit faster. 0:08:55.200 --> 0:08:57.599 What sort of impact do you see and where do 0:08:57.640 --> 0:08:59.280 you see growth potential from that? 0:09:00.760 --> 0:09:03.160 Well, I do think there's this broad market where you 0:09:03.240 --> 0:09:10.040 have to understand spaces that we can't perfectly anticipate. I 0:09:10.080 --> 0:09:12.559 think we're seeing a lot hearing, a lot more conversations 0:09:12.600 --> 0:09:19.480 around retailers and the deployment of kind of acres more 0:09:19.600 --> 0:09:22.360 quickly with fewer steps along the way. I think that's 0:09:22.440 --> 0:09:26.520 one application we're also based square stocks hold up quite 0:09:26.520 --> 0:09:30.040 a bit better. I think that software can easily deploy 0:09:30.400 --> 0:09:33.720 or maybe more easily deploy AI within their world. I 0:09:33.720 --> 0:09:36.679 think that's a possibility. So I think that applications are 0:09:36.720 --> 0:09:39.480 somewhat in less the imagination. It has to be pretty 0:09:39.480 --> 0:09:42.200 broad here, and so there's some argument for kind of 0:09:42.200 --> 0:09:45.120 an equal weight exposure to try to capture that range 0:09:45.160 --> 0:09:47.800 of experiences, but with long term horizon. I don't think 0:09:47.840 --> 0:09:50.120 this is necessarily something that we're going to see in 0:09:50.240 --> 0:09:52.160 Q two Q three of twenty twenty five. 0:09:52.559 --> 0:09:54.680 We're taking a few tech hits to your line, but 0:09:54.679 --> 0:09:56.560 we're going to stick with it matter because I have 0:09:56.640 --> 0:09:59.760 this crucial point for the here and now with retail 0:10:00.080 --> 0:10:03.000 we buying back into in video institutionals not right now? 0:10:03.559 --> 0:10:06.520 What about a so called black swad event or ripple effect. 0:10:06.600 --> 0:10:09.160 We talk of course to nasin Talib yesterday. 0:10:09.200 --> 0:10:11.120 Just have a listen to what he said, right. 0:10:12.080 --> 0:10:17.240 The beginning of an adjustment of people for reality because 0:10:17.280 --> 0:10:21.320 now they realize now it's no longer flawless. You have 0:10:21.559 --> 0:10:25.200 a small little trip on the glass. 0:10:25.600 --> 0:10:28.200 A little chip in the glass of the valuation question, 0:10:28.400 --> 0:10:30.160 a little chip in the glass as to whether this 0:10:30.240 --> 0:10:34.200 is akin to a dot com crisis, but where could 0:10:34.400 --> 0:10:38.520 be sustained? Will Energy survive this? We saw such significant 0:10:38.520 --> 0:10:39.600 sell off to certain of those. 0:10:39.520 --> 0:10:44.280 Names, you know, I think when we're taking a look 0:10:44.280 --> 0:10:48.080 at Nvidia or Energy, I think there's more vulnerability there, 0:10:48.400 --> 0:10:51.480 simply because this is getting at the heart of the 0:10:51.480 --> 0:10:54.400 thesis there that they are the you know, especially in Vidia, 0:10:54.480 --> 0:10:58.199 the dominant provider of this capability that our roads lead 0:10:58.240 --> 0:11:00.679 back to in video when it comes to supply, And 0:11:00.720 --> 0:11:04.439 this story is really tackling that idea that potentially there's 0:11:04.480 --> 0:11:06.199 more ways to skin a cat. You can focus on 0:11:06.240 --> 0:11:09.720 efficiency rather than share mine. Of course, so I think 0:11:09.960 --> 0:11:11.840 the idea of a chip in the glass, I think 0:11:11.880 --> 0:11:14.559 that's important, and it's getting at the heart of Nvidia. 0:11:14.640 --> 0:11:17.960 But whether this is the beginning of the unraveling of 0:11:18.000 --> 0:11:20.959 all things AI, I think that's a little bit more 0:11:21.160 --> 0:11:24.360 extreme or a bridge too far today, simply because that 0:11:24.640 --> 0:11:27.880 unraveling would depend on the use case, the use case 0:11:27.960 --> 0:11:29.960 of AI being intact, and I don't think that's what 0:11:30.000 --> 0:11:32.800 we're seeing here. It's more of the supply chain question. 0:11:33.160 --> 0:11:36.520 And really at this point, I think there's still more 0:11:36.559 --> 0:11:39.000 that we would need to see to really suggest that 0:11:39.040 --> 0:11:41.640 companies are going to pull back on their capital expenditures 0:11:41.880 --> 0:11:43.080 in any meaningful way. 0:11:44.280 --> 0:11:47.840 Martin Norton Chief Investment Strategies that Empowered Thank You. 0:11:48.120 --> 0:11:49.480 Coming up, we're. 0:11:49.320 --> 0:11:52.400 Going to discuss deep seaks open source approach and what 0:11:52.440 --> 0:11:55.600 it means for competitors. The chief science officer at AI 0:11:55.720 --> 0:11:57.880 platform Hugging Face will join us on that. 0:11:58.280 --> 0:11:59.880 This is Bloomberg. 0:12:12.080 --> 0:12:16.200 China's deepseek AI model up ended markets largely though those 0:12:16.320 --> 0:12:17.680 AI infrastructure plays. 0:12:17.920 --> 0:12:19.319 What about model providers? 0:12:19.520 --> 0:12:22.520 Clearly open source is impacting a HI development. We're joined 0:12:22.559 --> 0:12:25.320 by Thomas Wolf Now here's the chief science officer at 0:12:25.360 --> 0:12:28.079 hugging Face and open source and collaborative platform for AI 0:12:28.160 --> 0:12:32.200 builders is our one the huge breakthrough that the market 0:12:32.240 --> 0:12:33.040 clearly thinks it is. 0:12:35.120 --> 0:12:37.280 I think there is a thin bida for overreaction. 0:12:37.440 --> 0:12:40.760 We've seen a steady increase in open source model performance, 0:12:41.280 --> 0:12:44.200 but we also have to be honest. It's the first model, 0:12:44.200 --> 0:12:46.800 I would say that really reached the performance of closed source. 0:12:46.800 --> 0:12:49.080 So the gap is closed now between the clos source 0:12:49.120 --> 0:12:51.040 models and the open source performance. 0:12:52.080 --> 0:12:54.960 Let's just talk about how the derivatives have expanded. Because 0:12:55.400 --> 0:12:58.160 you are all about open source, you're about community, you're 0:12:58.200 --> 0:13:01.920 about the innovation one model can provide to many others. 0:13:02.000 --> 0:13:04.240 From what we understand from Clem the CEO over with 0:13:04.360 --> 0:13:07.760 hugging face, he's saying that there's been five hundred derivatives 0:13:07.960 --> 0:13:11.679 already created from deep seek. How are you seeing an 0:13:11.679 --> 0:13:14.080 exponential growth of the use of this model. 0:13:15.559 --> 0:13:18.319 Yeah, it's even growing, right, So I was just checking. 0:13:18.360 --> 0:13:21.200 I can give you a fresh new stats just now. 0:13:21.520 --> 0:13:25.040 We're over six six hundred and seventeen models already created 0:13:25.040 --> 0:13:28.960 by the community, more than three point two million downloads 0:13:28.960 --> 0:13:32.240 of all these models, with seven hundred thousand dollars for 0:13:32.280 --> 0:13:33.079 the original model. 0:13:33.320 --> 0:13:36.360 So I think that's really the power of open source. Right. 0:13:36.520 --> 0:13:38.079 We see basically a growth of. 0:13:38.640 --> 0:13:41.839 More than thirty person of downloads day to day, which 0:13:41.920 --> 0:13:44.880 kind of is a testament to all this ecosystem that 0:13:45.080 --> 0:13:48.200 is already starting to build around deep seek with basically 0:13:48.240 --> 0:13:51.280 all these companies, all these teams, all these. 0:13:51.160 --> 0:13:53.000 Organizations that are already you. 0:13:52.880 --> 0:13:56.160 Know, taking this open source model and finding in it, 0:13:56.280 --> 0:13:59.199 adapting it, testing it on the unused cases. So that's 0:13:59.640 --> 0:14:02.600 that's maybe the most beautiful thing about these open source capabilities, 0:14:02.600 --> 0:14:06.720 which is still this ecosystem grow live day by day 0:14:06.880 --> 0:14:08.679 around the original Deepsick model. 0:14:09.960 --> 0:14:12.640 Thomas, we can't separate the story of deep seek from 0:14:12.679 --> 0:14:16.440 the geopolitics here. What concerns do you have that a 0:14:16.520 --> 0:14:21.080 reaction from Washington or other governments might be to restrict 0:14:21.160 --> 0:14:24.280 open source use in some fashion as a way of 0:14:24.360 --> 0:14:26.680 keeping technology out of China's hands. 0:14:28.400 --> 0:14:31.000 Yeah, I think there's a lot of I would say 0:14:31.600 --> 0:14:34.880 a lot of way of reframing these stories as a 0:14:35.000 --> 0:14:38.720 US China. But really the more general story here is 0:14:38.760 --> 0:14:41.120 the open source effect of this model. It could have 0:14:41.200 --> 0:14:44.280 come out of almost any countries, and I would expect 0:14:44.320 --> 0:14:47.080 actually open source model to keep coming out of China. 0:14:47.160 --> 0:14:50.080 But we also have European companies like Mixed Trial starting 0:14:50.120 --> 0:14:52.600 to open source model and very small teams. 0:14:52.640 --> 0:14:55.920 So mordinally, I think hopefully. 0:14:55.560 --> 0:14:59.640 We'll see move from a geopolitical inpretation of this story 0:14:59.720 --> 0:15:03.360 to really interpretation at open source versus closed source, and 0:15:03.520 --> 0:15:06.840 our belief attacking face is open source is really the 0:15:06.920 --> 0:15:10.920 way to foster development, to foster a breakthrough technology, to 0:15:11.000 --> 0:15:14.360 foster growing communities, and a lot more business use cases. 0:15:14.480 --> 0:15:19.080 Basically, Thomas, as a software creator yourself, you must have 0:15:19.160 --> 0:15:22.200 some admiration for what Deep Seek has pulled off, apparently 0:15:22.520 --> 0:15:24.720 at a much lower cost and rivals. 0:15:24.360 --> 0:15:25.160 Here in the US. 0:15:25.360 --> 0:15:28.320 But I bet you have some questions too, what would 0:15:28.360 --> 0:15:30.760 you like to learn more about what the company did 0:15:30.800 --> 0:15:31.840 and how it pulled it off? 0:15:33.480 --> 0:15:36.520 So what is interesting is we already started a tragging 0:15:36.560 --> 0:15:40.080 pace to actually explore, you know, can we reproduce this model? 0:15:40.120 --> 0:15:42.960 Can we reproduce it in poticlar on the open right? 0:15:43.040 --> 0:15:45.400 So we have right the diepsick model. I was trying 0:15:45.400 --> 0:15:49.920 its last year, last day yesterday, and it does really, 0:15:50.240 --> 0:15:52.720 you know, offer what we see in the benchmark, which 0:15:52.800 --> 0:15:55.640 is it's a very powerful model. But what we would 0:15:55.720 --> 0:15:58.040 like to know is exactly can you retrain it? Can 0:15:58.040 --> 0:16:00.920 you actually apply the same recess? Is two more models, 0:16:00.960 --> 0:16:03.800 So I've study this project called open Air one, which 0:16:03.840 --> 0:16:07.760 is basically a reproduction and open reprodiction of the deep 0:16:07.800 --> 0:16:10.920 Sick pipeline. Thankfully, they shared a lot of details on 0:16:10.960 --> 0:16:11.440 how they. 0:16:11.320 --> 0:16:14.280 Train the model, much more than than we see recently. 0:16:14.600 --> 0:16:17.440 So I'm quite confident in the coming month we'll be 0:16:17.480 --> 0:16:20.760 able to basically, you know, understand all the breakthrough that 0:16:20.800 --> 0:16:22.040 we need to make in this model. 0:16:22.800 --> 0:16:26.040 What does it mean for a closed source focused open 0:16:26.080 --> 0:16:29.440 AI or for anthropic in this. 0:16:29.520 --> 0:16:33.480 Moment, Well, I think it's quite positive. And you saw 0:16:33.520 --> 0:16:35.760 probably the reaction of some outman, right. 0:16:35.840 --> 0:16:38.160 I think for a good competition in the field is 0:16:38.200 --> 0:16:41.320 something that is just a net positive for developing a technology. 0:16:41.760 --> 0:16:43.680 And in particular here because a lot of. 0:16:43.600 --> 0:16:45.480 This model and as I was saying, a lot of 0:16:45.560 --> 0:16:48.280 recipes are open, it means I would I would be 0:16:48.400 --> 0:16:51.440 very surprised they don't, you know, directly take the recipes 0:16:51.480 --> 0:16:54.760 that can be useful for improving both open AI, anthropic 0:16:55.000 --> 0:16:59.160 Google or the coming LAMMA models, and that this model basically. 0:16:58.680 --> 0:17:01.800 You know, quickly catch So I think it's a good thing. 0:17:01.920 --> 0:17:04.399 That's also the good thing about open sources because you 0:17:04.760 --> 0:17:07.679 share a lot of information about your models. You know, 0:17:07.800 --> 0:17:11.959 you actually lift the whole fields up with you by 0:17:12.040 --> 0:17:14.160 you know, explaining how how you make your breakthrough. 0:17:15.040 --> 0:17:18.320 But some people don't want to see everyone rise up. 0:17:18.560 --> 0:17:22.280 They don't want to see China rise up. Ultimately, is 0:17:22.320 --> 0:17:25.480 that a false narrative a strong man to be, even 0:17:25.560 --> 0:17:27.800 saying that China is behind the US when there's open 0:17:27.840 --> 0:17:29.160 source communities such as yours. 0:17:30.880 --> 0:17:34.400 Yeah, the Open Source Committee generally don't really know any border, right, 0:17:34.880 --> 0:17:37.440 as I was saying today, it's a Chinese team, but generally, 0:17:37.520 --> 0:17:39.960 you know, it's just it's just a very smart, edge 0:17:40.320 --> 0:17:40.920 edge round team. 0:17:41.000 --> 0:17:41.200 Right. 0:17:41.400 --> 0:17:43.800 There are teams like that in many countries, and that's 0:17:43.800 --> 0:17:46.439 why we keep saying it would keep seeing new actors 0:17:46.480 --> 0:17:50.919 in AI, right, So I think ultimately, I believe in 0:17:51.000 --> 0:17:54.800 fair market. I think having competition, sharing more actively, you know, 0:17:54.920 --> 0:17:58.960 is a good way to make progress together. So you know, 0:17:59.040 --> 0:18:01.200 in your future, we will like to see much more 0:18:01.240 --> 0:18:03.840 actors active in AI. We would like to see basically 0:18:04.000 --> 0:18:07.280 a larger company a bit everywhere in the world. And 0:18:07.320 --> 0:18:09.320 I think that's maybe the first step, you know, in 0:18:09.359 --> 0:18:09.919 this direction. 0:18:11.280 --> 0:18:15.280 Are we expecting to see similar innovations coming from Europe? 0:18:15.359 --> 0:18:18.760 There have been questions and concerns raised in the European 0:18:18.880 --> 0:18:21.680 Union about the level of regulation and then it might 0:18:21.720 --> 0:18:23.920 be restricting development in AI. 0:18:26.280 --> 0:18:28.440 Yeah, it's obviously a big discussion, right. 0:18:28.720 --> 0:18:30.919 I was a divorce a bit last week, and if 0:18:30.960 --> 0:18:32.960 you see a lot of the discussion was, Okay, what 0:18:33.480 --> 0:18:34.680 is about your regulation? 0:18:35.080 --> 0:18:36.760 I just think the teams are really good. 0:18:36.840 --> 0:18:39.600 I would be surprised, for instance, in the UK, in Germany, 0:18:39.720 --> 0:18:41.760 you know in France that we that we don't see 0:18:41.760 --> 0:18:44.879 all these teams that basically helped train you know, the 0:18:44.960 --> 0:18:48.879 models of Mistual, at Meta, at Opena, at Entropic. 0:18:48.920 --> 0:18:51.040 We've seen a lot of people also living op Ai 0:18:51.200 --> 0:18:52.440 to start the und startup. 0:18:52.520 --> 0:18:55.840 So I think one takeaway from Deep Sick is that 0:18:55.920 --> 0:19:00.280 basically the recipe to build a very good quality large 0:19:00.320 --> 0:19:04.400 language model nowadays is something that's almost accessible to everyone, right, 0:19:04.480 --> 0:19:07.520 So I think all these people leaving the big tech 0:19:07.520 --> 0:19:10.720 companies will start the out team, and because you kind 0:19:10.720 --> 0:19:12.920 of just need a few millions, I would be surprised 0:19:12.920 --> 0:19:14.639 we don't see a lot more teams coming out of 0:19:14.640 --> 0:19:17.400 Europe and Botska, but also as a region this year. 0:19:18.760 --> 0:19:21.840 Thomas Wolf, chief science officer at Hugging Face, thanks for 0:19:21.920 --> 0:19:33.160 joining us. Microsoft is in talks that acquire the US 0:19:33.280 --> 0:19:36.080 arm of TikTok. At least that's according to President Trump 0:19:36.160 --> 0:19:38.440 last night, who did not elaborate further. 0:19:38.800 --> 0:19:40.240 Microsoft declient to comment. 0:19:40.520 --> 0:19:43.479 Let's bring in Bloomberg's Balance and Power co host Kaylee 0:19:43.520 --> 0:19:47.040 lines for more. Kaylee, is this another case of Trump 0:19:47.320 --> 0:19:50.560 trying to make a deal happen? He casts himself as 0:19:50.600 --> 0:19:53.960 the consummate deal maker. We are so short on details, though, 0:19:54.240 --> 0:19:56.920 walk us through what more we know, if anything beyond 0:19:57.000 --> 0:19:57.480 last night. 0:19:58.880 --> 0:20:00.960 Well, the details are pretty limited, Mike. 0:20:01.000 --> 0:20:03.720 He was asked directly by a reporter if Microsoft was 0:20:03.760 --> 0:20:06.359 in talks to acquire TikTok, and the President said, quote, 0:20:06.359 --> 0:20:08.520 I would say yes, And that's really all we got 0:20:08.520 --> 0:20:11.120 on that specific matter. As he said, Microsoft isn't commenting 0:20:11.440 --> 0:20:13.440 on this either. Keep in mind, though, that during the 0:20:13.480 --> 0:20:16.040 first Trump administration, back in the summer of twenty twenty, 0:20:16.040 --> 0:20:19.600 Bloomberg did report Microsoft was looking at acquiring TikTok's US 0:20:19.640 --> 0:20:22.760 operations back then, when the first kind of pressure around 0:20:22.760 --> 0:20:26.399 divesting or banning it over national security concerns was a rising. 0:20:26.440 --> 0:20:28.600 Oracle was reportedly looking at it too, And we know 0:20:28.720 --> 0:20:31.560 just last week when the Stargate project was being announced 0:20:31.560 --> 0:20:33.439 at the White House and Larry Ellison of Oracle was 0:20:33.440 --> 0:20:35.399 in the room, Donald Trump suggested he'd be open to 0:20:35.440 --> 0:20:38.720 Ellison buying TikTok or Elon Musk when he was asked 0:20:38.760 --> 0:20:41.640 about that. We know there are other players involved as well, 0:20:41.680 --> 0:20:44.439 including Frank McCourt, who you speak with frequently on this program, 0:20:44.440 --> 0:20:46.159 who have made bus and it does seem that is 0:20:46.160 --> 0:20:49.040 what Donald Trump's preference really is here, a bidding war. 0:20:49.119 --> 0:20:51.280 He said as much last night that he thinks bidding 0:20:51.280 --> 0:20:54.080 war's result in the best deals. As for what that 0:20:54.119 --> 0:20:56.919 deal ultimately looks like, that remains unclear. What he was 0:20:56.960 --> 0:20:59.720 clear on yesterday when speaking to the House Republican Conference 0:20:59.760 --> 0:21:02.479 in his to Raw Club in Florida is that he 0:21:02.560 --> 0:21:05.640 does not want China involved in whatever happens here. Of course, 0:21:05.720 --> 0:21:08.040 China is going to have some say in that ultimately, 0:21:08.280 --> 0:21:11.040 would the US government have involvement still, Katie. 0:21:11.760 --> 0:21:12.639 Well, that's the question. 0:21:12.680 --> 0:21:15.560 He's talked about this fifty percent ownership structure, whoever buys 0:21:15.600 --> 0:21:18.440 it sharing equity with the United States essentially, so it 0:21:18.440 --> 0:21:20.600 would be some kind of half and half deal, But 0:21:20.680 --> 0:21:22.560 it remains unclear whether or not that can be done 0:21:22.560 --> 0:21:25.120 by the now April fourth deadline, which has been extended 0:21:25.160 --> 0:21:28.040 by Trump's executive order. There's also a massive question around 0:21:28.080 --> 0:21:30.399 whether anti trust concerns would be raised by some of 0:21:30.400 --> 0:21:35.040 these individuals or companies like Microsoft acquiring operations of this size, 0:21:35.040 --> 0:21:37.080 which could be tens of billions of dollars and obviously, 0:21:37.080 --> 0:21:38.920 as you guys will know, one hundred and seventy million 0:21:39.000 --> 0:21:42.280 users in the US. While Donald Trump and the people 0:21:42.280 --> 0:21:44.720 he's instilled at places like the FTC or at the 0:21:44.760 --> 0:21:47.720 Department of Justice may have some alignment on this, it 0:21:47.800 --> 0:21:51.600 still could face scrutiny, especially from big tech skeptical lawmakers 0:21:51.760 --> 0:21:53.200 on Capitol Hill and. 0:21:53.320 --> 0:21:56.679 Whether bite don'ts whatever said it, Katie Hines, Yeah, thanks 0:21:56.680 --> 0:21:57.080 so much. 0:22:05.480 --> 0:22:07.840 Welcome back to med Technology. I'm Caroline Hyde in New 0:22:07.920 --> 0:22:09.200 York and I'm. 0:22:09.040 --> 0:22:10.520 Mike Shepard in San Francisco. 0:22:10.960 --> 0:22:13.600 Let's discuss the impact that Deep seak is having on 0:22:13.720 --> 0:22:17.640 US tech companies now with Bloomberg Sharen Gafari. Sharen, thanks 0:22:17.680 --> 0:22:20.840 so much and thanks for all your reporting today. There's 0:22:20.920 --> 0:22:23.600 a lot of soul searching going on in Silicon Valley 0:22:23.760 --> 0:22:26.200 over the past forty eight hours. Tell us a little 0:22:26.240 --> 0:22:28.680 bit more about what's happening. They set up war rooms. 0:22:28.920 --> 0:22:30.399 What is the conversation like? 0:22:31.640 --> 0:22:34.439 So the top a I labs right now in the 0:22:34.600 --> 0:22:37.920 US are trying to figure out how the Chinese startup 0:22:38.000 --> 0:22:41.160 Deep Seek was able to catch up so quickly. Right, 0:22:41.200 --> 0:22:43.720 you have their latest R one model. Really exploding on 0:22:43.760 --> 0:22:46.600 the scene in the past week, and everyone is astounded 0:22:46.680 --> 0:22:49.960 by how competitive it is. It is actually leading by 0:22:49.960 --> 0:22:53.720 some metrics on these AI models that have taken Western companies, 0:22:53.960 --> 0:22:55.679 you know, years end, a lot more money. 0:22:56.280 --> 0:22:57.520 The Chinese company. 0:22:57.240 --> 0:23:00.320 Deep Sea claims to build well, we can old bait 0:23:00.359 --> 0:23:04.080 as to how much it was copying building off Western technology. 0:23:04.119 --> 0:23:06.919 But the key question is will capex remain the same 0:23:07.160 --> 0:23:09.800 and ultimately what it means for open AI and Anthropic 0:23:09.880 --> 0:23:11.719 in terms of reducing the cost of their models. 0:23:11.800 --> 0:23:13.680 What do you think the ripple effects Ascherene. 0:23:15.119 --> 0:23:17.359 Yeah, I think we're starting to see people really question 0:23:18.440 --> 0:23:23.080 whether we need these astronomical budgets for building the most 0:23:23.080 --> 0:23:26.119 advanced AI models and whether, you know, because. 0:23:25.760 --> 0:23:27.960 Deep Seek was able to make some gains in. 0:23:27.960 --> 0:23:32.480 How efficiently they use the computing power, if those you know, 0:23:32.520 --> 0:23:34.680 gains can be applied now to US companies and sort 0:23:34.680 --> 0:23:37.040 of questioning why US companies didn't come up with that first. 0:23:38.600 --> 0:23:39.920 It's a fascinating big take. 0:23:40.000 --> 0:23:43.000 Go and read it from Sharene Gafari and the meta 0:23:43.040 --> 0:23:44.679 war rooms that are currently being set up. 0:23:44.720 --> 0:23:45.479 We appreciate it. 0:23:45.520 --> 0:23:48.480 Now we've got to assess what President Trump had to 0:23:48.480 --> 0:23:50.280 say about the competition from deep Seek. 0:23:51.560 --> 0:23:54.080 I think if it's if it's fact and if it's true, 0:23:54.720 --> 0:23:57.080 and nobody really knows what it is, but I view 0:23:57.119 --> 0:23:59.000 that as a positive because you'll be doing that too, 0:23:59.080 --> 0:24:00.960 so you won't be spending much and you'll get the 0:24:00.960 --> 0:24:06.720 same result. Hopefully, the release of Deepseek AI from a 0:24:06.880 --> 0:24:09.600 Chinese company should be a wake up call for our 0:24:09.680 --> 0:24:13.280 industries that we need to be laser focused on competing 0:24:13.320 --> 0:24:13.600 to win. 0:24:14.080 --> 0:24:15.200 Let's talk about that competition. 0:24:15.280 --> 0:24:17.639 Jacqueline Rice Nelson is with us CO Founracy of Tribe 0:24:17.640 --> 0:24:20.800 Ai provides AI services for leading enterprises, and I have 0:24:20.840 --> 0:24:22.720 a feeling a lot of them are rushing to understand 0:24:22.720 --> 0:24:25.640 how they can use this cheap and innovative AI model. 0:24:25.960 --> 0:24:30.119 Absolutely first, thrilled to be here. I think this twenty 0:24:30.160 --> 0:24:33.840 twenty five is the year for enterprises to get ROI 0:24:34.040 --> 0:24:38.159 from their AI efforts, and the ROI calculation just changed dramatically, 0:24:38.560 --> 0:24:41.679 So I think that kind of is the headline for businesses, 0:24:41.800 --> 0:24:44.440 which is use cases that weren't possible just a few 0:24:44.440 --> 0:24:48.840 weeks ago or even last week are now possible. And 0:24:49.200 --> 0:24:51.440 I think we're about to see an explosion of AI 0:24:51.560 --> 0:24:55.920 experimentation and also value delivery to organizations and enterprises away. 0:24:56.040 --> 0:24:58.199 Have any of those organizations been reticent to use it 0:24:58.200 --> 0:24:59.640 simply because of where it was born? 0:24:59.720 --> 0:25:01.479 And it's a great question. 0:25:02.160 --> 0:25:05.760 I actually had the same question and dove into this myself, 0:25:06.359 --> 0:25:08.560 and I think it all comes down to how you 0:25:08.680 --> 0:25:12.240 use these models. So there are lots of ways to 0:25:12.280 --> 0:25:16.240 be using LAMA or open source models today in ways 0:25:16.240 --> 0:25:18.320 that have lots of guardrails that are set up to 0:25:18.359 --> 0:25:20.879 be more secure and can be run locally on your 0:25:20.920 --> 0:25:23.639 own environment. In many ways, actually it's almost easier with 0:25:23.680 --> 0:25:26.520 open source than it is with closed source models, And 0:25:26.600 --> 0:25:31.000 so everything comes down to the implementations of these models. 0:25:31.040 --> 0:25:34.399 I think today a lot of businesses have these same questions, 0:25:35.119 --> 0:25:37.600 but that's in many ways why they're sort of coming 0:25:37.680 --> 0:25:40.880 to Tribe and trying to help navigate the landscape across 0:25:40.920 --> 0:25:43.280 all of these different models. What are best to use 0:25:43.400 --> 0:25:46.280 and for what types of use cases and how to 0:25:46.280 --> 0:25:49.480 set it up in the optimal ways Jacqueline. 0:25:49.480 --> 0:25:51.840 One of the key questions that's come up in relation 0:25:51.960 --> 0:25:54.760 to deep seek is the one on spending. How is 0:25:54.800 --> 0:25:57.960 that factoring into the way you are guiding people through 0:25:58.000 --> 0:25:59.119 the AI landscape? 0:25:59.119 --> 0:25:59.320 Now? 0:25:59.640 --> 0:26:01.919 Is it a different message that you're having to deliver 0:26:02.480 --> 0:26:07.600 about efficiency and about bringing a product to market for 0:26:07.640 --> 0:26:08.320 a lot less. 0:26:09.240 --> 0:26:12.880 So the short answer is yes, I think that we are. 0:26:13.440 --> 0:26:16.199 I compare it to a highway. We have just added 0:26:16.280 --> 0:26:18.600 a lane or multiple lanes to the highway. When you 0:26:18.640 --> 0:26:21.719 do that, the traffic does not go down. Actually it 0:26:21.760 --> 0:26:25.159 only increases, and that's because driving and going on road 0:26:25.200 --> 0:26:28.160 trips gets more attractive. The same is true right now 0:26:28.280 --> 0:26:31.320 for AI. So I think we're about to see really 0:26:31.359 --> 0:26:36.000 an explosion of activity and that sort of across companies, 0:26:36.040 --> 0:26:39.760 across enterprises within the large and to your point on 0:26:39.840 --> 0:26:44.320 efficiency and for product innovation, I think we're going to 0:26:44.359 --> 0:26:48.680 see proliferation across these use cases. And then that doesn't 0:26:48.680 --> 0:26:51.679 even include the large hyper scalers who are still on 0:26:51.760 --> 0:26:55.440 this quest for AGI and nothing changes there. I don't 0:26:55.440 --> 0:26:58.280 think anyone is taking their foot off the gas. And 0:26:58.320 --> 0:27:00.760 if anything, competition is just in dramatically. 0:27:02.040 --> 0:27:04.800 And the open source model, is this something that is 0:27:05.320 --> 0:27:10.119 feasible to be scaled at large enterprises, Big corporations that 0:27:10.240 --> 0:27:14.680 also may have security considerations and other factors to bring 0:27:14.760 --> 0:27:17.440 in them might make them question whether open source is 0:27:17.480 --> 0:27:18.600 the right choice for them. 0:27:19.160 --> 0:27:21.480 Yeah, I think there are a lot of considerations that 0:27:21.560 --> 0:27:24.800 go into what is the right model selection, and so 0:27:25.240 --> 0:27:27.760 I think it's really clear that we have entered an era, 0:27:27.840 --> 0:27:29.920 and actually I've always thought we were in this era, 0:27:30.280 --> 0:27:36.159 which is that we are in a multimodel world. Businesses 0:27:36.240 --> 0:27:39.159 need to be able to build, to swap between models 0:27:39.160 --> 0:27:45.640 seamlessly to optimize for performance, for latency, for accuracy, and 0:27:45.680 --> 0:27:48.800 for cost. And what that means is they also need 0:27:49.080 --> 0:27:52.600 the guardrails, the security, the infrastructure to set those things 0:27:52.680 --> 0:27:55.919 up in proper ways. They also need the evals, the 0:27:55.960 --> 0:27:59.120 evaluation framework to actually be able to compare and contrast 0:27:59.119 --> 0:28:02.919 across models so they know what models to use. And 0:28:02.960 --> 0:28:06.520 so I think there, while there might be some hesitance, 0:28:07.560 --> 0:28:10.080 we have done lots of development on closed source models 0:28:10.080 --> 0:28:12.760 and lots of development on open source models, and the 0:28:12.840 --> 0:28:16.480 considerations still come down to what is that company trying 0:28:16.480 --> 0:28:18.760 to achieve and what are the things we're optimizing for. 0:28:19.040 --> 0:28:22.159 That's what determines the model. So I, oh, please go ahead. 0:28:22.520 --> 0:28:24.159 Well I'm interested in that because I want to go 0:28:24.200 --> 0:28:27.560 back to sort of the Jevons paradox piece that you know, 0:28:27.760 --> 0:28:30.280 the more AI innovation there is, ultimately the more people 0:28:30.280 --> 0:28:33.080 on those highways the more innovation, but also the same 0:28:33.119 --> 0:28:34.919 amount of compute. Can you speak a bit to like 0:28:34.960 --> 0:28:37.240 the inferencing part of this, because that seems to be 0:28:37.280 --> 0:28:39.840 the silver lining that indeeded videos trying to see from 0:28:39.880 --> 0:28:41.800 this look, you're still going to need our GPUs, just 0:28:41.840 --> 0:28:42.479 not for training. 0:28:42.480 --> 0:28:43.200 It's the inference. 0:28:43.880 --> 0:28:46.120 I think that's fair. Look, I think we have not 0:28:46.240 --> 0:28:49.160 cracked the code on the hardware that's needed. Like it's 0:28:49.200 --> 0:28:51.920 clear that compute is still needed at really high levels. 0:28:52.320 --> 0:28:55.120 There is no way that we have even with these 0:28:55.160 --> 0:28:58.239 new deep seek models, we haven't found the answers yet. Right, 0:28:58.280 --> 0:29:00.800 We're not stopping. And I think that that's the goal 0:29:00.920 --> 0:29:04.760 case for how and why you're going to see continued 0:29:04.840 --> 0:29:08.440 investment here is that we are in a continuous innovation 0:29:08.640 --> 0:29:13.520 cycle and ultimately, I think that drives utilization and I 0:29:13.560 --> 0:29:16.760 think the question is for what. And then the other 0:29:16.800 --> 0:29:20.360 piece is that ultimately consumers are the biggest winners here, 0:29:21.160 --> 0:29:23.920 and I think that is the most exciting story. There's 0:29:23.920 --> 0:29:26.280 a lot of fear, there's a lot of you know, 0:29:26.360 --> 0:29:30.200 China versus US dynamics. I think the reality is that 0:29:30.600 --> 0:29:34.240 competition is good. And what I'm hearing from the US 0:29:34.360 --> 0:29:38.240 companies is you know, nothing has kind of the US 0:29:38.400 --> 0:29:41.000 large AI companies, nothing has kind of blown their socks 0:29:41.080 --> 0:29:44.400 deef right. This is innovation they at least, you know, 0:29:44.520 --> 0:29:48.040 feel that they already have. And the difference is that 0:29:48.080 --> 0:29:50.280 there has been an optimization for open source and an 0:29:50.320 --> 0:29:53.280 optimization for cost in ways that they have not done. 0:29:53.600 --> 0:29:55.360 And so I think we're about to see a lot 0:29:55.400 --> 0:29:59.120 of fast follow and Sam Alman signaled as much last night. 0:29:59.240 --> 0:30:01.880 Yeah, he said, we're going to drop models faster or is. 0:30:01.880 --> 0:30:03.920 It going to drop the cost because that it must 0:30:03.960 --> 0:30:05.080 be something that puts off your. 0:30:04.960 --> 0:30:07.680 Client's definitely going to drop the cost. And I actually 0:30:07.680 --> 0:30:12.880 think we are likely to see a cost curve decrease 0:30:13.000 --> 0:30:16.280 like that's dramatic from them, but also a latency decrease. 0:30:16.600 --> 0:30:18.200 And I think that that's how they're going to play 0:30:18.240 --> 0:30:20.680 the game is to try to now one up completely 0:30:21.040 --> 0:30:22.160 on both dimensions. 0:30:23.560 --> 0:30:28.480 Jacqueline Rice Nelson, CEO of Tribe AI, Thank you, Caraw's just. 0:30:28.560 --> 0:30:31.080 Check on these markets because these are all the questions 0:30:31.280 --> 0:30:33.480 that investors are trained on at the moment. What does 0:30:33.520 --> 0:30:35.560 it mean for compute going forwards, what does it mean 0:30:35.600 --> 0:30:38.520 for the competitors in the generative AI model space? And 0:30:38.560 --> 0:30:40.680 what does it ultimately mean for software too? Now we're 0:30:40.680 --> 0:30:42.600 clawing back some of our losses of yesterday, but not 0:30:42.640 --> 0:30:44.440 March eight tens percent higher on than as that one 0:30:44.520 --> 0:30:45.560 hundred Meta. 0:30:45.320 --> 0:30:46.120 In a new record high. 0:30:46.160 --> 0:30:49.200 We'll talk more about Lama there, but Bitcoin up nine 0:30:49.240 --> 0:30:50.840 tens percent. We're going to delve into the world of 0:30:50.880 --> 0:30:52.400 crypto and the ripple effects in a moment where one 0:30:52.480 --> 0:30:53.600 hundred and two thousand move on. 0:30:53.800 --> 0:30:55.080 The individual movers. 0:30:54.800 --> 0:30:56.800 That you've got to keep your eyes trained on have been, 0:30:56.840 --> 0:30:59.640 of course some of the points contributors we've seen Meta 0:30:59.800 --> 0:31:02.320 new record. As I mentioned, Apple has been doing significantly 0:31:02.360 --> 0:31:04.600 well in video bounces back, but hardly at all. Up 0:31:04.640 --> 0:31:07.360 two point eight percent after seventeen percent sell off, We're 0:31:07.400 --> 0:31:09.600 only up about thirty billion. It will six hundred billion 0:31:09.640 --> 0:31:12.320 yesterday Tesla of by two percent. Remember, earnings are almost 0:31:12.400 --> 0:31:14.920 upon us Tesla tomorrow out of the gate. We're gonna 0:31:14.920 --> 0:31:16.760 have the likes of Microsoft to chew on and Meta 0:31:16.840 --> 0:31:19.080 as well. But coming up more on musks World and 0:31:19.160 --> 0:31:23.479 questions around his approach to efficiency, because apparently it's rather 0:31:23.520 --> 0:31:50.560 good at it, don't It would seem this is gluebog technology. 0:31:52.440 --> 0:31:55.360 Elil musks X has announced a new partnership with Visa. 0:31:55.440 --> 0:31:57.600 The social media platform is tapping the payment company to 0:31:57.720 --> 0:32:00.680 enable digital wallets on its app and website, in what 0:32:00.720 --> 0:32:02.920 marks the first step for the company to create the 0:32:02.920 --> 0:32:04.040 so called Everything app. 0:32:04.320 --> 0:32:04.480 Now. 0:32:04.480 --> 0:32:07.080 The move comes as bankers are now looking to offload 0:32:07.120 --> 0:32:10.360 some free billion dollars of X's debt. According to sources, 0:32:10.360 --> 0:32:12.800 Pimco Apollo are now said to be among the asset 0:32:12.800 --> 0:32:15.200 managers looking to purchase a portion of the debt being 0:32:15.240 --> 0:32:17.600 sold by group of banks at by Morgan Stanley. Some 0:32:17.680 --> 0:32:21.840 interesting sweetness in that offering too, and staying with Elil Musk, 0:32:22.080 --> 0:32:25.600 Many are now scrutinizing the tech billionaires approach to efficiency 0:32:26.040 --> 0:32:27.160 with his own companies. 0:32:27.200 --> 0:32:29.120 This as he gears up to cut spending and waste 0:32:29.120 --> 0:32:29.360 in the. 0:32:29.400 --> 0:32:33.400 US government with of course, the Dose initiative for mores Cretrudel, 0:32:33.520 --> 0:32:36.000 it's a perfect person to discuss as to how good 0:32:36.080 --> 0:32:38.280 is mascut efficiency in the private sector. 0:32:39.800 --> 0:32:42.200 Well, I think it's it's worth kind of, you know, 0:32:42.240 --> 0:32:45.560 taking a look at his track record and maybe sort 0:32:45.600 --> 0:32:47.920 of you know, being open to the idea that there's 0:32:48.160 --> 0:32:52.240 maybe some more similarities in how he has run his 0:32:52.320 --> 0:32:56.800 companies and you know, how his criticism of the government 0:32:57.480 --> 0:33:02.200 is you know, characterized, then then he is possibly let on. 0:33:02.320 --> 0:33:05.680 You know, this is it's been the case that Tesla, 0:33:05.760 --> 0:33:09.160 for example, you know, took a good decade before it 0:33:09.200 --> 0:33:13.040 was making its investors any money. Uh And and you 0:33:13.080 --> 0:33:16.080 know there was plenty of waste, you know, even according 0:33:16.120 --> 0:33:19.960 to him, or inefficiency according to him within the company 0:33:20.600 --> 0:33:23.120 that was sort of allowed to. 0:33:22.640 --> 0:33:24.880 Fester for some time before he turned things around. 0:33:24.920 --> 0:33:27.560 And so you know, no one's disputing that Tesla and 0:33:27.600 --> 0:33:31.680 SpaceX aren't you know, quite efficient and now quite you know, 0:33:31.800 --> 0:33:35.800 lucrative companies for their shareholders. But you can find, you know, 0:33:35.880 --> 0:33:40.360 your fair share of inefficiency or waste within his own companies. 0:33:40.400 --> 0:33:43.200 And you know, even on his own earnings calls, you know, 0:33:43.280 --> 0:33:47.200 him sort of talking about ways in which his companies 0:33:47.320 --> 0:33:50.960 haven't necessarily been as efficient as you would expect from 0:33:51.000 --> 0:33:54.280 somebody who's about to run run Doge. 0:33:55.080 --> 0:34:00.360 Craig, your reporting pointed to some pretty visible example of 0:34:00.480 --> 0:34:04.240 maybe a pro fligod approach flying tires from the Czech 0:34:04.280 --> 0:34:07.400 Republic to the US to ensure their delivery that had 0:34:07.400 --> 0:34:10.480 to have been expensive. How does that kind of approach 0:34:10.640 --> 0:34:15.080 translate to to government where the consequences and stakeholders are 0:34:15.320 --> 0:34:16.000 very different. 0:34:17.520 --> 0:34:21.440 I think what what Denni's Danna Hole in San Francisco, uh, 0:34:21.600 --> 0:34:24.319 you know, pointed out in this story is essentially that 0:34:24.400 --> 0:34:28.920 you know, there's there's a willingness to sort of allow 0:34:29.040 --> 0:34:33.400 for you know, some inefficiency if it means sort of uh, 0:34:33.480 --> 0:34:36.920 you know, other deliverables and uh, you know within the government. 0:34:37.000 --> 0:34:39.919 I think you know, the answer that we're clearly going 0:34:39.960 --> 0:34:42.799 to get is you know, something different in terms of 0:34:43.160 --> 0:34:46.400 you know, not so much forgiveness of Okay, yes we 0:34:46.440 --> 0:34:49.640 were inefficient, but we delivered X or y uh. 0:34:49.840 --> 0:34:51.760 You know what we're what we're seeing from. 0:34:51.640 --> 0:34:54.640 Doze early on is is this you know, relatively new 0:34:54.800 --> 0:34:57.360 X account that is just you know, kind of highlighting 0:34:57.400 --> 0:34:59.560 all the ways that the government is you know, so 0:34:59.760 --> 0:35:03.000 as of blowing taxpayer dollars and all the ways in 0:35:03.040 --> 0:35:05.480 which you know, Musk and his team are coming in 0:35:05.920 --> 0:35:09.479 and sort of you know, writing those wrongs in his view. 0:35:09.560 --> 0:35:14.000 And so I think, you know, he's not showing a 0:35:14.000 --> 0:35:17.239 whole lot of willingness to you know, cut slack in 0:35:17.320 --> 0:35:20.359 terms of trying different things or you know, trying things 0:35:20.360 --> 0:35:22.839 and then not going well as he maybe has been 0:35:22.920 --> 0:35:26.640 in his sort of you know times as CEO of 0:35:26.680 --> 0:35:27.560 many of these companies. 0:35:28.800 --> 0:35:30.520 Bloomberg's Craig Trudell, thank you. 0:35:30.920 --> 0:35:33.600 Turning to crypto, deep seek sell off pose a threat 0:35:33.640 --> 0:35:38.000 across all markets, including crypto, with bitcoin seeing its biggest 0:35:38.040 --> 0:35:40.520 intra day drop in more than a month. All this 0:35:40.640 --> 0:35:44.040 comes just as President Trump signed an executive order last 0:35:44.080 --> 0:35:46.799 week calling for the creation of a White House Advisory 0:35:46.840 --> 0:35:50.640 Group on digital currencies. Joining us now to discuss all 0:35:50.680 --> 0:35:54.360 of this is Melton Demr's Crucible Capital Group General partner 0:35:54.440 --> 0:35:56.440 and founder, Melton we. 0:35:56.440 --> 0:35:57.920 Really have to get right to it. 0:35:58.000 --> 0:36:01.480 We were talking about Elon Musk Department of Government efficiency. 0:36:01.640 --> 0:36:05.120 One of the biggest questions facing your industry is regulation. 0:36:05.680 --> 0:36:08.640 David Sachs, the new AI and cryptos are what are 0:36:08.680 --> 0:36:11.880 you looking for him and the Trump team to clear away? 0:36:13.200 --> 0:36:15.600 Well, I think I'm going to say something controversial. 0:36:15.880 --> 0:36:18.759 The last four years by an administration, the lack of 0:36:18.840 --> 0:36:22.800 clarity on regulatory policy in many ways I think was 0:36:22.840 --> 0:36:25.719 actually constructive for the crypto industry in the sense that 0:36:26.320 --> 0:36:30.080 there was not a lot of external pressure. Clarity I 0:36:30.080 --> 0:36:32.600 think can be really challenging for markets. 0:36:32.800 --> 0:36:34.560 Yes, there are obviously bright spots. 0:36:34.600 --> 0:36:37.960 The rollback of SAB one twenty one will now allow 0:36:38.280 --> 0:36:40.680 banks to hold crypto on their balance sheet, make it 0:36:40.760 --> 0:36:45.719 easier for institutions to hold crypto. Sure, that's great, but clarity, 0:36:45.760 --> 0:36:50.239 I think also creates more perspective on what will work 0:36:50.280 --> 0:36:52.120 and what won't work in crypto. 0:36:52.239 --> 0:36:54.719 So far, the view has been number go up. 0:36:54.800 --> 0:36:54.960 Right. 0:36:55.000 --> 0:36:57.759 We sometimes joke that coin as number go up technology, 0:36:58.160 --> 0:37:00.160 and there's a meme in the industry. 0:37:00.160 --> 0:37:01.279 Me we're all going to make it. 0:37:01.760 --> 0:37:03.959 I think what's happening in week one of the Trump 0:37:04.000 --> 0:37:08.640 administration is this clarity is helping people and particularly markets 0:37:08.719 --> 0:37:11.600 understand that we are not all going to make it. 0:37:11.680 --> 0:37:14.000 They're going to be winners and they're going to be losers. 0:37:14.280 --> 0:37:17.359 And this administration has made it very clear that they're 0:37:17.400 --> 0:37:19.840 perfectly content picking winners and losers. 0:37:20.760 --> 0:37:23.040 AI coin is going to be winners in the future. 0:37:23.080 --> 0:37:26.560 There's been this Steve tailing of the AI trade within crypto, 0:37:26.840 --> 0:37:29.000 and they've taken a brutal hit of course, in the 0:37:29.080 --> 0:37:29.720 CAR narrative. 0:37:30.600 --> 0:37:33.120 Now, look, I think the AI crypto narrative is a 0:37:33.160 --> 0:37:36.480 confusing one. Two things I look at One, how does 0:37:36.560 --> 0:37:40.040 crypto make AI better or safer? And the jury is 0:37:40.080 --> 0:37:43.000 out there and then the reverse is how does. 0:37:42.920 --> 0:37:45.880 AI make crypto better or safer? 0:37:46.400 --> 0:37:50.440 There I think we've used AI to create new casinos 0:37:50.520 --> 0:37:54.279 in crypto in the form of agents or online accounts 0:37:54.600 --> 0:37:57.920 tied to lms that are creating their own coins and 0:37:58.000 --> 0:37:59.200 launching their own coins. 0:37:59.560 --> 0:38:01.520 To me, that's not what's interesting. 0:38:01.800 --> 0:38:04.759 What I'm looking at is how can crypto and some 0:38:04.920 --> 0:38:10.600 of the opportunities around aggregation, optimization, and financialization help crypto 0:38:10.800 --> 0:38:13.799 make AI better. And then I'm really looking at what 0:38:13.880 --> 0:38:17.680 Apple is doing smaller models run locally on devices I 0:38:17.680 --> 0:38:21.600 think is very interesting opportunity for crypto. Obviously, the big 0:38:21.719 --> 0:38:24.319 story around deep seek is what's going to happen to 0:38:24.360 --> 0:38:27.399 all of this energy and compute capex and I think 0:38:27.480 --> 0:38:29.680 there are some of the early efforts we're seeing in 0:38:29.719 --> 0:38:34.080 crypto to aggregate the long tail of compute into these 0:38:34.160 --> 0:38:37.320 marketplaces where you can pay in stable coins or dollars. 0:38:37.520 --> 0:38:38.520 Is really interesting. 0:38:38.719 --> 0:38:41.320 Oh, guys, all the way back to file coin of old. 0:38:41.440 --> 0:38:42.879 But I'm interested. 0:38:42.560 --> 0:38:47.920 In Melton in what the regulatory spirits have meant underlying 0:38:48.120 --> 0:38:51.680 all of what is sent meant and felt like to 0:38:51.800 --> 0:38:54.600 go up on bitcoin, What has the trading told you. 0:38:55.719 --> 0:38:57.520 Yeah, if we look at markets right. 0:38:57.840 --> 0:39:00.439 What I think is always so interesting is you see 0:39:00.440 --> 0:39:03.680 a lot of sentiment online and it's easy to say something, 0:39:03.760 --> 0:39:05.760 but if you want to know what people are truly thinking, 0:39:05.800 --> 0:39:07.640 you have to look at markets, and you have to. 0:39:07.560 --> 0:39:08.320 Look at flows. 0:39:08.680 --> 0:39:11.440 The biggest week of inflows we had in the last 0:39:11.520 --> 0:39:14.280 year into crypto ETFs, right, which were a great proxy 0:39:14.719 --> 0:39:17.800 four million dollars the week that Trump won the election. Okay, 0:39:18.120 --> 0:39:20.920 last week with all of the regulatory quote unquote clarity 0:39:21.040 --> 0:39:24.200 or at least setting of directions with Trump administration and 0:39:24.200 --> 0:39:26.160 the executive orders being signed. 0:39:26.239 --> 0:39:27.440 Two billion in inflows. 0:39:27.640 --> 0:39:30.840 So the expectation is always better than the reality. 0:39:31.040 --> 0:39:33.560 Right. Markets not reacting positively. 0:39:33.960 --> 0:39:38.120 CM Bitcoin futures contract had the biggest drop in open 0:39:38.160 --> 0:39:41.839 interest in its entire trading history yesterday, So it is 0:39:42.000 --> 0:39:46.640 very clear that I'm crypto traders. Markets are feeling overextended, 0:39:46.880 --> 0:39:49.000 just like they did on the big AI names. 0:39:48.680 --> 0:39:49.840 And there's been a pullback. 0:39:50.080 --> 0:39:52.200 The question is how much of that is going to 0:39:52.239 --> 0:39:54.719 come back with clarity, and how much of that is 0:39:54.760 --> 0:39:58.000 going to come back when we actually start to see reality. 0:39:57.600 --> 0:40:01.640 Catch up with hype trade the room of sell the news, 0:40:01.640 --> 0:40:04.080 Malton damares, it's so good to have you on Creaseable 0:40:04.080 --> 0:40:05.480 Capital Group general partner. 0:40:06.000 --> 0:40:06.600 We thank you. 0:40:14.920 --> 0:40:17.680 Meta is set to report fourth quarter earnings after the 0:40:17.719 --> 0:40:20.359 bell tomorrow. For more of what we can expect from 0:40:20.360 --> 0:40:23.880 the social media giant, Bloomberg's Kurt Wagner joins us Now, 0:40:24.239 --> 0:40:28.320 Kurt Metta somehow escaped the big sell off from deep 0:40:28.360 --> 0:40:31.439 Seek yesterday. Why is that and what does that tell 0:40:31.520 --> 0:40:33.160 us going into the results tomorrow? 0:40:34.560 --> 0:40:37.239 Yeah, it might be two things. I think one is, 0:40:37.920 --> 0:40:39.759 you may recall, Mike, at the end of last week, 0:40:39.800 --> 0:40:42.160 they announced their big plan for the year, all the 0:40:42.200 --> 0:40:45.360 spending they were going to be doing on AI. You know, 0:40:45.640 --> 0:40:48.800 they announced that Threads was going to start running ads. 0:40:48.840 --> 0:40:50.920 Like they kind of bront run to their own earnings 0:40:50.920 --> 0:40:53.799 a little bit with some of the biggest news that 0:40:53.880 --> 0:40:56.480 they have planned for twenty twenty five. And so I 0:40:56.520 --> 0:40:58.439 think maybe some people just saw a little bit less 0:40:58.480 --> 0:41:00.560 uncertainty with Meta because some of that was out there. 0:41:00.640 --> 0:41:03.800 I think the second is that this deep Seek stuff 0:41:04.040 --> 0:41:07.000 in a way sort of validates the AI strategy that 0:41:07.040 --> 0:41:09.000 Meta has been pushing towards this whole time. You know, 0:41:09.080 --> 0:41:12.439 Mark Zuckerberg has been arguing for more than a year 0:41:12.880 --> 0:41:15.440 that where this was going was open source, and that 0:41:15.880 --> 0:41:17.719 is where Meta was going to go as well, was 0:41:17.760 --> 0:41:20.799 to create an open source model that other people could 0:41:20.800 --> 0:41:23.040 build on. Now we saw that is what deep Seek 0:41:23.520 --> 0:41:25.960 is building as well. So there's obviously more competition for 0:41:26.040 --> 0:41:28.680 Meta here, but I think the strategy that they were 0:41:28.719 --> 0:41:31.920 employing might be seen as the best way forward here. 0:41:31.920 --> 0:41:34.759 So perhaps some investors were less spooked because they thought 0:41:34.760 --> 0:41:36.759 the road ahead for Meta made a little bit more 0:41:36.800 --> 0:41:37.800 sense than some of the others. 0:41:37.880 --> 0:41:41.320 Gene Munster deep Water Asset Management reflected that exact issue, 0:41:41.360 --> 0:41:43.800 so too did City, and they're also thinking that maybe 0:41:43.920 --> 0:41:47.080 the AI models becoming cheaper and cheaper are also going 0:41:47.120 --> 0:41:50.799 to just really push on the overll profitability the advertising improvements. 0:41:50.840 --> 0:41:52.040 Is that what we're going to have to hear the 0:41:52.160 --> 0:41:55.600 our return on AI investment from Mark as well as 0:41:55.680 --> 0:41:57.840 his investment going forward in Capex. 0:41:59.120 --> 0:42:01.239 Yeah, this has been one of the biggest questions, not 0:42:01.320 --> 0:42:03.680 just for Meta but for all these tech companies is 0:42:04.000 --> 0:42:05.319 when is all this investment going. 0:42:05.239 --> 0:42:05.799 To pay off? 0:42:06.120 --> 0:42:08.600 I think in twenty twenty four, actually Medaid probably a 0:42:08.600 --> 0:42:10.479 better job than a lot of its peers that sort 0:42:10.480 --> 0:42:13.799 of conveying where AI was impacting the product right, not 0:42:13.840 --> 0:42:18.440 only in making the ads more efficient, but in dispersing 0:42:18.480 --> 0:42:21.040 their AI assistant across all of their different apps. Having 0:42:21.040 --> 0:42:23.440 the ray Band smart glasses right, they had sort of 0:42:23.480 --> 0:42:26.960 tangible products that they could roll out and show people. Obviously, 0:42:26.960 --> 0:42:28.640 with this massive investment they're going to be making in 0:42:28.680 --> 0:42:30.520 twenty twenty five, it will be even more important for 0:42:30.560 --> 0:42:31.920 them to continue to show that. 0:42:32.320 --> 0:42:33.080 But given what they did. 0:42:33.040 --> 0:42:35.120 In twenty twenty four, at least it seems like the 0:42:35.160 --> 0:42:37.160 street sort of knows what to expect from them a 0:42:37.200 --> 0:42:39.120 little bit. So we'll see if they're able to continue 0:42:39.120 --> 0:42:40.040 that tomorrow. 0:42:40.120 --> 0:42:43.680 For earnings, shares at a record high, currently trading six 0:42:43.800 --> 0:42:46.000 hundred and seventy seven kut Wagner, We thank you. That 0:42:46.080 --> 0:42:48.440 does it for this edition of Bloomberg Technology. You do 0:42:48.520 --> 0:42:50.840 not want to forget a podcast. You can find it 0:42:50.880 --> 0:42:52.960 on the terminal as well as online on Apple, Spotify, 0:42:52.960 --> 0:42:53.560 and iHeart. 0:42:54.080 --> 0:42:55.360 This is Bloomberg Technology.