1 00:00:02,750 --> 00:00:05,590 Speaker 1: Bloomberg Audio Studios. Podcasts. 2 00:00:05,890 --> 00:00:07,270 Speaker 2: Radio. News. 3 00:00:09,369 --> 00:00:22,100 Speaker 3: Hello and welcome to another episode of the Odd Lot Podcast. 4 00:00:22,239 --> 00:00:23,340 Speaker 4: I'm Joe Wiesenthal. 5 00:00:23,579 --> 00:00:24,480 Speaker 2: And I'm Traci Alloway. 6 00:00:25,350 --> 00:00:28,030 Speaker 3: Tracy, I got to say, this Jensen Wong story, I 7 00:00:28,850 --> 00:00:30,290 Speaker 3: think it's one of the biggest stories ever. 8 00:00:30,530 --> 00:00:30,910 Speaker 2: Yeah. 9 00:00:30,950 --> 00:00:34,440 Speaker 3: Just as a news story, it's about as big as 10 00:00:34,479 --> 00:00:35,360 Speaker 3: it could possibly get. 11 00:00:35,580 --> 00:00:37,200 Speaker 4: It has everything. It has everything. 12 00:00:37,360 --> 00:00:42,840 Speaker 5: Alleged Chinese spies, chip smuggling. It's exciting stuff. Also, it's 13 00:00:42,940 --> 00:00:46,960 Speaker 5: birthed like 5,000 memes, some of which are quite good. 14 00:00:47,280 --> 00:00:48,560 Speaker 4: I've definitely seen some good memes. 15 00:00:48,580 --> 00:00:50,959 Speaker 5: But actually, so on a serious note, if you think 16 00:00:51,020 --> 00:00:54,540 Speaker 5: about the AI trade, I mean, the sun around which 17 00:00:54,620 --> 00:00:58,300 Speaker 5: it all revolves is NVIDIA, both in terms of like 18 00:00:58,450 --> 00:01:00,790 Speaker 5: providing the actual chips that powers a lot of this 19 00:01:00,850 --> 00:01:04,209 Speaker 5: technology and also providing a lot of the financing and 20 00:01:04,270 --> 00:01:07,690 Speaker 5: investments for these companies. Yeah. And so any hint of 21 00:01:07,750 --> 00:01:11,590 Speaker 5: a scandal has big implications for AI as a whole. 22 00:01:12,110 --> 00:01:13,970 Speaker 3: No, I really like the way you put that, that 23 00:01:14,190 --> 00:01:17,419 Speaker 3: they are the son of the universe. Both from it, 24 00:01:17,510 --> 00:01:21,110 Speaker 3: they have the crown jewel technology, the chips. They have 25 00:01:21,160 --> 00:01:25,560 Speaker 3: the financing to back all kinds of partners and neoclouds 26 00:01:25,620 --> 00:01:26,920 Speaker 3: and everything all around the world. 27 00:01:27,120 --> 00:01:28,029 Speaker 4: Everyone wants it. 28 00:01:28,520 --> 00:01:33,580 Speaker 3: And it is a technology that is both commercially intensely 29 00:01:33,700 --> 00:01:36,690 Speaker 3: in demand and geopolitically in demand. And so in addition 30 00:01:36,730 --> 00:01:38,149 Speaker 3: to the fact that the tech is good and they 31 00:01:38,170 --> 00:01:41,570 Speaker 3: have the financing, there are all these fights about who 32 00:01:41,590 --> 00:01:44,009 Speaker 3: they could even sell it to starting, you know, there 33 00:01:44,030 --> 00:01:46,330 Speaker 3: are all these things of the Biden administration, lobbying of 34 00:01:46,350 --> 00:01:49,810 Speaker 3: the Trump administration, how liberal should the policies be about 35 00:01:50,140 --> 00:01:52,740 Speaker 3: who gets to buy them, et cetera. That is why 36 00:01:52,820 --> 00:01:57,060 Speaker 3: this story is just like, so it's huge. 37 00:01:57,400 --> 00:01:57,680 Speaker 2: Yeah. 38 00:01:57,800 --> 00:02:00,980 Speaker 5: And just on the geopolitical, atomic's a good word. On 39 00:02:01,000 --> 00:02:04,080 Speaker 5: the geopolitical angle itself, I mean, there's this tension that 40 00:02:04,140 --> 00:02:07,120 Speaker 5: we've seen more and more in recent years where some 41 00:02:07,180 --> 00:02:10,270 Speaker 5: people will argue that, well, you want the U.S., you 42 00:02:10,310 --> 00:02:14,070 Speaker 5: want a company like NVIDIA to be the backbone of 43 00:02:14,190 --> 00:02:17,260 Speaker 5: like the AI tech stack, right? And so maybe you 44 00:02:17,300 --> 00:02:19,040 Speaker 5: want lots of people using their chips. 45 00:02:19,060 --> 00:02:19,420 Speaker 4: Totally, totally. 46 00:02:19,639 --> 00:02:22,120 Speaker 5: But on the other hand, you know, you have people 47 00:02:22,180 --> 00:02:24,800 Speaker 5: in the Trump administration who argue that we are in 48 00:02:25,060 --> 00:02:28,940 Speaker 5: this existential AI race with China, and therefore it's really 49 00:02:29,000 --> 00:02:32,210 Speaker 5: important that China does not have access to the most 50 00:02:32,370 --> 00:02:35,570 Speaker 5: advanced chips. And so we're seeing that sort of play 51 00:02:35,710 --> 00:02:37,490 Speaker 5: out in a lot of the discussion on what's going 52 00:02:37,530 --> 00:02:38,470 Speaker 5: on with NVIDIA right now. 53 00:02:39,020 --> 00:02:41,860 Speaker 3: The other thing that's very important, and this will allow 54 00:02:41,919 --> 00:02:45,100 Speaker 3: us to get into the guest, is that in some sense, 55 00:02:46,200 --> 00:02:48,880 Speaker 3: there's actually, like, the story's kind of old news, right? 56 00:02:48,900 --> 00:02:53,400 Speaker 3: There's been a lot of reporting on this company, Megaspeed, 57 00:02:55,139 --> 00:02:58,239 Speaker 3: and its CEO, and the company's closeness to NVIDIA and 58 00:02:58,280 --> 00:03:02,139 Speaker 3: Jensen specifically, and where they know. So, like, in a way, like, 59 00:03:02,500 --> 00:03:04,820 Speaker 3: what is the new news here is part of the question, 60 00:03:04,860 --> 00:03:07,760 Speaker 3: but whatever, you know. I guess on some level, the 61 00:03:07,800 --> 00:03:13,359 Speaker 3: new news... is just specifically Jensen Wong admitting to NVIDIA 62 00:03:13,380 --> 00:03:18,359 Speaker 3: that he had this relationship, an affair, with this executive 63 00:03:18,440 --> 00:03:23,430 Speaker 3: or someone associated with Megaspeed and the investigation into it. Anyway, 64 00:03:23,450 --> 00:03:26,590 Speaker 3: without further ado, we really do have the perfect guest, 65 00:03:26,910 --> 00:03:28,829 Speaker 3: someone who I think is like the best beat in 66 00:03:28,870 --> 00:03:32,070 Speaker 3: the world. Whenever we get a chance to chat, I'm 67 00:03:32,190 --> 00:03:35,760 Speaker 3: always learning more about both the sort of like politics, 68 00:03:35,900 --> 00:03:40,290 Speaker 3: geopolitics of chip flows and, around the world, someone who's 69 00:03:40,350 --> 00:03:43,310 Speaker 3: reported on chips all around the world, from D.C. to 70 00:03:43,330 --> 00:03:45,770 Speaker 3: San Francisco, and now based in Hong Kong. We're going 71 00:03:45,790 --> 00:03:48,130 Speaker 3: to be speaking with tech reporter here at Bloomberg News, 72 00:03:48,550 --> 00:03:51,900 Speaker 3: Mackenzie Hawkins. So, Mackenzie, thank you so much for coming 73 00:03:51,960 --> 00:03:52,500 Speaker 3: on Odd Lots. 74 00:03:53,680 --> 00:03:55,320 Speaker 2: Thank you so much for having me. Quite a week. 75 00:03:55,640 --> 00:03:58,440 Speaker 5: Staying up late and coming on Odd Lots, because you're 76 00:03:58,460 --> 00:03:59,860 Speaker 5: based in Hong Kong, so thank you. 77 00:04:00,000 --> 00:04:01,440 Speaker 3: And you were just in the U.S., so you're probably 78 00:04:01,460 --> 00:04:03,480 Speaker 3: still kind of jet-lagged. 79 00:04:03,600 --> 00:04:07,030 Speaker 2: You know, I'm actually, like, This week I've gone back 80 00:04:07,050 --> 00:04:09,340 Speaker 2: to working U.S. hours, so I feel like I never 81 00:04:09,380 --> 00:04:11,280 Speaker 2: left the U.S. I finally got over the jet lag 82 00:04:11,340 --> 00:04:12,720 Speaker 2: and Jensen Wong is jet lagging me. 83 00:04:12,800 --> 00:04:15,460 Speaker 3: I can't believe how fried you must be. So again, 84 00:04:15,500 --> 00:04:17,660 Speaker 3: thank you for taking the time. Let me start with 85 00:04:17,700 --> 00:04:22,400 Speaker 3: this question. You know, reporters always think they know a 86 00:04:22,400 --> 00:04:24,220 Speaker 3: little bit more than what they can write at any 87 00:04:24,260 --> 00:04:26,039 Speaker 3: given moment, etc. 88 00:04:26,640 --> 00:04:28,440 Speaker 4: Because, you know, you're waiting for verification. 89 00:04:28,760 --> 00:04:31,480 Speaker 1: Let's say we were talking a week ago. 90 00:04:32,010 --> 00:04:34,900 Speaker 3: And so based on only what you're allowed to report, 91 00:04:34,920 --> 00:04:37,720 Speaker 3: let's say this conversation is happening sometime in October and 92 00:04:37,740 --> 00:04:40,919 Speaker 3: we say, Mackenzie, what's the deal with this company Megaspeed? 93 00:04:41,700 --> 00:04:43,799 Speaker 3: What was the answer you gave me a week ago? 94 00:04:44,260 --> 00:04:48,060 Speaker 3: Just that you're totally able, who is Megaspeed? And why 95 00:04:48,110 --> 00:04:50,110 Speaker 3: should we care about this company called Megaspeed? 96 00:04:51,490 --> 00:04:56,450 Speaker 2: So pre-affair admission, Megaspeed is the single largest Southeast Asian 97 00:04:56,470 --> 00:04:59,290 Speaker 2: buyer of NVIDIA chips. It's a major supplier of AI 98 00:04:59,310 --> 00:05:02,419 Speaker 2: compute to Alibaba. which is one of the most important 99 00:05:02,490 --> 00:05:06,610 Speaker 2: tech companies and frontier AI leaders in China. It purchased 100 00:05:06,610 --> 00:05:11,250 Speaker 2: $ 4. 6 billion worth of NVIDIA hardware after being spun off 101 00:05:11,350 --> 00:05:14,740 Speaker 2: from a Chinese gaming company, passing through a BVI shell, 102 00:05:15,339 --> 00:05:18,580 Speaker 2: and incorporating in Singapore, then becoming an official NVIDIA partner 103 00:05:18,640 --> 00:05:24,020 Speaker 2: four days later. Its executive had a close relationship with Jensen. 104 00:05:24,400 --> 00:05:27,330 Speaker 2: We didn't know quite how close. And it had been 105 00:05:27,430 --> 00:05:31,599 Speaker 2: under U.S. investigation and Singapore investigation for, if we're talking 106 00:05:31,620 --> 00:05:33,480 Speaker 2: a week ago, at least a year. 107 00:05:34,960 --> 00:05:38,640 Speaker 5: So why did this raise eyebrows in the first place? 108 00:05:38,720 --> 00:05:43,430 Speaker 5: So Megaspeed renting out GPU capacity to a company like Alibaba. 109 00:05:43,520 --> 00:05:47,210 Speaker 5: My understanding is like, OK, that's allowed. People know that 110 00:05:47,250 --> 00:05:50,010 Speaker 5: this is happening. But at the same time, if you're 111 00:05:50,050 --> 00:05:55,060 Speaker 5: worried about Chinese companies having access to technologically advanced NVIDIA chips, 112 00:05:55,140 --> 00:05:56,440 Speaker 5: it seems like a pretty big loophole. 113 00:05:57,900 --> 00:06:01,350 Speaker 2: Yeah. So, I mean, the arrangement that Megaspeed has with 114 00:06:01,390 --> 00:06:06,510 Speaker 2: Alibaba is extraordinarily common across the Southeast Asian data center ecosystem. 115 00:06:06,770 --> 00:06:09,190 Speaker 2: You know, in some ways, the backbone of Chinese compute 116 00:06:09,339 --> 00:06:14,000 Speaker 2: is like Johor, Malaysia. So it's not actually the remote 117 00:06:14,060 --> 00:06:17,660 Speaker 2: access cloud computing arrangement that got Megaspeed sort of looked 118 00:06:17,700 --> 00:06:19,859 Speaker 2: at by the U.S. government, later by the Singapore government. 119 00:06:20,400 --> 00:06:23,359 Speaker 2: It's the suspicion that the company has not just rented 120 00:06:23,420 --> 00:06:27,940 Speaker 2: compute to Chinese offtakers, but actually diverted those, like the 121 00:06:28,020 --> 00:06:32,290 Speaker 2: physical hardware into mainland China. Right. So that's the focus 122 00:06:32,310 --> 00:06:34,650 Speaker 2: of the U.S. investigation. It was looking at whether Megaspeed 123 00:06:34,670 --> 00:06:37,350 Speaker 2: had smuggled chips, which the company has categorically denied having 124 00:06:37,390 --> 00:06:40,989 Speaker 2: ever done. And also looking into its ownership structure, because 125 00:06:41,190 --> 00:06:44,150 Speaker 2: even if Megaspeed is a Singaporean entity, if it were 126 00:06:44,370 --> 00:06:46,950 Speaker 2: owned by a Chinese company, it would also be illegal 127 00:06:46,970 --> 00:06:50,010 Speaker 2: to sell advanced NVIDIA chips to Megaspeed without a U.S. license. 128 00:06:50,430 --> 00:06:52,650 Speaker 2: It is 100 percent Singapore owned, but by a shell. 129 00:06:53,470 --> 00:06:57,410 Speaker 3: Let's get into like there is this question of, you know, 130 00:06:57,510 --> 00:07:01,530 Speaker 3: what do chip restrictions even mean if you could still 131 00:07:01,570 --> 00:07:04,710 Speaker 3: just access a cloud in a different country? And like, 132 00:07:04,990 --> 00:07:08,150 Speaker 3: is this the mother of all loopholes? But let's actually 133 00:07:08,170 --> 00:07:12,900 Speaker 3: just start on the smuggling question. Why might investigators have 134 00:07:12,940 --> 00:07:18,460 Speaker 3: been concerned that Megaspeed was literally smuggling chips or not? Like, 135 00:07:18,860 --> 00:07:21,880 Speaker 3: what are the details that we know around this question? 136 00:07:21,940 --> 00:07:25,520 Speaker 3: The actual, the physical movement of chips from, say, a 137 00:07:25,560 --> 00:07:29,400 Speaker 3: delivery address in Johor, Malaysia to an address inside of 138 00:07:29,720 --> 00:07:30,200 Speaker 3: mainland China. 139 00:07:30,920 --> 00:07:31,200 Speaker 4: Yeah. 140 00:07:31,300 --> 00:07:34,120 Speaker 2: So, I mean, the U.S. government has a handful of 141 00:07:34,140 --> 00:07:38,440 Speaker 2: export control officers overseas, including in Southeast Asia. And when 142 00:07:38,460 --> 00:07:40,000 Speaker 2: I say a handful, it's really a handful. It's like 143 00:07:40,040 --> 00:07:42,160 Speaker 2: a dozen. They're going to more than double the number 144 00:07:42,230 --> 00:07:44,370 Speaker 2: this year to 25. For context, you know, when we 145 00:07:44,380 --> 00:07:47,970 Speaker 2: did our investigation into Megaspeed last year, I think I 146 00:07:48,010 --> 00:07:51,630 Speaker 2: was on singular email chains that had like 50 Bloomberg people. 147 00:07:51,650 --> 00:07:54,290 Speaker 2: So the capacity of the U.S. 148 00:07:54,330 --> 00:07:55,010 Speaker 1: Government. 149 00:07:55,050 --> 00:07:55,700 Speaker 4: Hold up, hold up. 150 00:07:56,590 --> 00:07:57,830 Speaker 5: Probably says more about Bloomberg. 151 00:07:57,890 --> 00:08:01,280 Speaker 3: All of us in this room have been on email 152 00:08:01,340 --> 00:08:05,060 Speaker 3: change with 50 Bloomberg people. A little inside baseball here, 153 00:08:05,100 --> 00:08:06,380 Speaker 3: but this is a real phenomenon. 154 00:08:06,400 --> 00:08:07,080 Speaker 4: All right, keep going. 155 00:08:08,020 --> 00:08:11,050 Speaker 2: But no, in all seriousness, the U.S. government has fairly 156 00:08:11,130 --> 00:08:16,510 Speaker 2: limited enforcement capacity overseas, but they do checks. And there's 157 00:08:16,630 --> 00:08:19,230 Speaker 2: unfortunately not a ton that I can say about the 158 00:08:19,560 --> 00:08:22,690 Speaker 2: initial trigger for the probe into mega speed. But through 159 00:08:22,710 --> 00:08:25,810 Speaker 2: the course of looking at shipments through Southeast Asia, they 160 00:08:26,240 --> 00:08:28,420 Speaker 2: thought that this company was worth digging into more. I 161 00:08:28,640 --> 00:08:30,520 Speaker 2: can talk a little about why I started looking into 162 00:08:30,540 --> 00:08:31,420 Speaker 2: Max Speed. Sure, go for it. 163 00:08:31,440 --> 00:08:32,080 Speaker 5: If that would be helpful. 164 00:08:32,140 --> 00:08:37,000 Speaker 2: Yeah. So in April or May of last year, I 165 00:08:37,000 --> 00:08:39,179 Speaker 2: had sort of recently arrived in Hong Kong and I 166 00:08:39,200 --> 00:08:41,970 Speaker 2: was working with a couple of colleagues here and in 167 00:08:42,010 --> 00:08:46,290 Speaker 2: the mainland on a story about this cluster of data 168 00:08:46,309 --> 00:08:49,230 Speaker 2: center operators in China's western deserts that had issued reports 169 00:08:49,620 --> 00:08:53,830 Speaker 2: procurement orders and local government tender notices that they had 170 00:08:53,870 --> 00:08:58,290 Speaker 2: fulfilled to build data centers seeking 115,000 restricted NVIDIA chips. 171 00:08:58,750 --> 00:09:02,210 Speaker 2: This was, you know, in order of magnitude larger than 172 00:09:02,290 --> 00:09:05,130 Speaker 2: what anybody thought might be in China. And so a 173 00:09:05,150 --> 00:09:08,350 Speaker 2: big question that we were trying to probe is, you know, 174 00:09:09,170 --> 00:09:11,870 Speaker 2: it's one thing for them to be constructing buildings. You know, 175 00:09:11,929 --> 00:09:14,590 Speaker 2: Chinese companies love to construct shells. I mean, I mean 176 00:09:14,610 --> 00:09:16,890 Speaker 2: that physically in this sense, but obviously that is the 177 00:09:16,910 --> 00:09:19,590 Speaker 2: case for shell companies as well. But it's one thing 178 00:09:19,610 --> 00:09:21,109 Speaker 2: to build the buildings. It's another thing to put the 179 00:09:21,130 --> 00:09:23,110 Speaker 2: chips inside. And so one of the big questions that 180 00:09:23,130 --> 00:09:24,770 Speaker 2: we were trying to tease out that we weren't able 181 00:09:24,790 --> 00:09:28,610 Speaker 2: to answer in that investigation was whether these particular operators 182 00:09:29,050 --> 00:09:32,480 Speaker 2: had the supply chain, the black market supply chain necessary 183 00:09:32,510 --> 00:09:35,719 Speaker 2: to get the GPUs. And when I was asking around 184 00:09:36,100 --> 00:09:39,319 Speaker 2: in Southeast Asia supply chains, everyone just kept saying, look 185 00:09:39,340 --> 00:09:42,640 Speaker 2: at this company called Megaspeed. And so we started looking 186 00:09:42,720 --> 00:09:45,459 Speaker 2: at them a year and a half ago, maybe a 187 00:09:45,480 --> 00:09:47,780 Speaker 2: bit over a year and a half ago. and found out, 188 00:09:48,020 --> 00:09:51,300 Speaker 2: you know, several months actually into our own investigation that 189 00:09:51,330 --> 00:09:53,270 Speaker 2: the U.S. government had been looking into them as well. 190 00:09:53,730 --> 00:09:57,530 Speaker 2: The U.S. investigation remains ongoing in a sort of complicated way, 191 00:09:57,590 --> 00:10:00,100 Speaker 2: which we can get into a bit later. In our reporting, 192 00:10:00,140 --> 00:10:03,059 Speaker 2: we didn't find concrete evidence that Megaspeed had smuggled any 193 00:10:03,080 --> 00:10:07,600 Speaker 2: GPUs to China. We found an extraordinary number of inconsistencies 194 00:10:07,740 --> 00:10:10,580 Speaker 2: in their data center footprint that neither the company nor 195 00:10:10,660 --> 00:10:11,860 Speaker 2: NVIDIA have been able to explain. 196 00:10:12,480 --> 00:10:15,059 Speaker 5: I want to ask about those inconsistencies, but before we do, 197 00:10:15,320 --> 00:10:19,290 Speaker 5: can you just describe how chip smuggling typically works? Like, 198 00:10:19,470 --> 00:10:22,010 Speaker 5: how do people in the black market actually do this? 199 00:10:23,460 --> 00:10:27,580 Speaker 2: Yeah, so, you know, we have enough examples of, you know, 200 00:10:27,620 --> 00:10:30,860 Speaker 2: public allegations, things that are being actively prosecuted to sort 201 00:10:30,900 --> 00:10:35,020 Speaker 2: of draw some thematic conclusions. There is a secondary market 202 00:10:35,200 --> 00:10:38,860 Speaker 2: for NVIDIA hardware, but, you know, most NVIDIA chips get 203 00:10:38,900 --> 00:10:43,340 Speaker 2: to customers via official allocation from NVIDIA. And NVIDIA doesn't 204 00:10:43,380 --> 00:10:45,929 Speaker 2: sell chips directly to the end user, like they wouldn't 205 00:10:45,970 --> 00:10:48,690 Speaker 2: be selling hardware directly to Megaspeed or directly to CoreWeave 206 00:10:48,710 --> 00:10:53,839 Speaker 2: or whoever. They sell to OEMs, original equipped manufacturers, server 207 00:10:53,860 --> 00:10:57,760 Speaker 2: makers like Supermicro or Dell or HP. Those server makers 208 00:10:57,800 --> 00:11:00,800 Speaker 2: typically sell to distributors. Those distributors might sell to another distributor. 209 00:11:01,160 --> 00:11:02,750 Speaker 2: then they go into a data center, and then they're 210 00:11:02,790 --> 00:11:05,750 Speaker 2: often rented to another company. So it's this extremely vast 211 00:11:05,809 --> 00:11:09,690 Speaker 2: distribution network for NVIDIA chips. But ultimately, in order to 212 00:11:09,790 --> 00:11:13,390 Speaker 2: place an order and get your order approved by NVIDIA, 213 00:11:13,530 --> 00:11:15,929 Speaker 2: you need to say, like, this is where they're going. Like, 214 00:11:16,010 --> 00:11:18,390 Speaker 2: this is the customer, this is the data center location. 215 00:11:19,010 --> 00:11:22,110 Speaker 2: And a consistent theme across, you know, what's now like 216 00:11:22,150 --> 00:11:24,670 Speaker 2: half a dozen cases being prosecuted, and this is certainly 217 00:11:24,720 --> 00:11:29,340 Speaker 2: the case in the indictment of Supermicro's co-founder earlier this year, 218 00:11:30,130 --> 00:11:32,800 Speaker 2: The stated location of where the chips are going to 219 00:11:32,820 --> 00:11:35,220 Speaker 2: wind up was not the intended location. You know, it's 220 00:11:35,260 --> 00:11:37,540 Speaker 2: not that there was some sort of rogue person at 221 00:11:38,260 --> 00:11:40,520 Speaker 2: one node in the supply chain that was like, yeah, 222 00:11:40,540 --> 00:11:43,400 Speaker 2: I'm just going to grab them. You have companies or 223 00:11:43,460 --> 00:11:47,490 Speaker 2: individual actors that place orders that are purportedly for a 224 00:11:47,530 --> 00:11:50,370 Speaker 2: destination outside of China with the intent of diverting that 225 00:11:50,390 --> 00:11:53,069 Speaker 2: hardware to China the entire time is at least the 226 00:11:53,110 --> 00:11:56,030 Speaker 2: allegation that's sort of laid out in several indictments that 227 00:11:56,670 --> 00:11:59,170 Speaker 2: And then it's a question of, well, how much work 228 00:11:59,290 --> 00:12:01,860 Speaker 2: does NVIDIA need to do and do the server manufacturers 229 00:12:01,880 --> 00:12:04,359 Speaker 2: need to do to make sure that the actual stated 230 00:12:04,380 --> 00:12:08,140 Speaker 2: destination of the chips exists? And there are some cases 231 00:12:08,260 --> 00:12:11,660 Speaker 2: where prosecutors have alleged the data center literally didn't exist. 232 00:12:11,720 --> 00:12:13,739 Speaker 2: There was never a place for those chips to go. 233 00:12:13,820 --> 00:12:15,600 Speaker 2: They were just always intended to go to China. 234 00:12:16,140 --> 00:12:18,520 Speaker 5: Wait, can I just, I'm a little bit disappointed. Are 235 00:12:18,559 --> 00:12:22,570 Speaker 5: there no chip mules out there? People who are just like, 236 00:12:23,170 --> 00:12:24,820 Speaker 5: I don't know. I have this vision of someone in 237 00:12:24,860 --> 00:12:26,880 Speaker 5: a trench coat and they like open it up and 238 00:12:26,920 --> 00:12:30,460 Speaker 5: it's lined with like, probably not H-100s, with black wells 239 00:12:30,500 --> 00:12:30,939 Speaker 5: or something. 240 00:12:32,020 --> 00:12:34,220 Speaker 2: Well, you know, it's funny that you say that because, 241 00:12:34,360 --> 00:12:37,219 Speaker 2: you know, in one of Jensen's many denials of chip 242 00:12:37,240 --> 00:12:40,000 Speaker 2: smuggling as an issue, he said, you know, the Grace 243 00:12:40,040 --> 00:12:41,860 Speaker 2: black well system wastes you tons. You're not putting that 244 00:12:41,960 --> 00:12:45,160 Speaker 2: in your backpack or pocket anytime soon. We reported a 245 00:12:45,179 --> 00:12:48,280 Speaker 2: couple months ago that the U.S. government has been investigating 246 00:12:48,340 --> 00:12:53,270 Speaker 2: a Singapore headquarter logistics company that literally specializes in tech 247 00:12:53,290 --> 00:12:58,060 Speaker 2: hardware transportation through complicated logistics. Those are fragile systems and 248 00:12:58,100 --> 00:13:01,660 Speaker 2: transported or is believed to have transported restricted hardware. And 249 00:13:02,360 --> 00:13:03,500 Speaker 2: then it's a question of like, OK, if you're a 250 00:13:03,520 --> 00:13:06,480 Speaker 2: logistics company, then how much is it your job to 251 00:13:06,520 --> 00:13:09,220 Speaker 2: know what's in the package? So a lot of these 252 00:13:09,260 --> 00:13:14,130 Speaker 2: cases also involve false declaration or allegedly involve false declaration, 253 00:13:14,470 --> 00:13:16,490 Speaker 2: you know, putting on customs forms that it's just a 254 00:13:16,530 --> 00:13:21,319 Speaker 2: computer server. Ambiguous what type of computer server. Sometimes random electronics, 255 00:13:21,360 --> 00:13:24,350 Speaker 2: there was one case where, and this is a case 256 00:13:24,370 --> 00:13:28,130 Speaker 2: where one person's actually pled guilty already, where they declared 257 00:13:28,790 --> 00:13:31,870 Speaker 2: restricted NVIDIA hardware to be, quote unquote, adapter modules. This 258 00:13:31,910 --> 00:13:34,750 Speaker 2: is a category of electronics that includes the thing you 259 00:13:34,850 --> 00:13:36,830 Speaker 2: use to plug your phone into a foreign power outlet. 260 00:13:37,550 --> 00:13:37,809 Speaker 6: Really? 261 00:13:37,870 --> 00:13:42,270 Speaker 2: Adapter modules that weigh 2,000 pounds? So there are lots 262 00:13:42,309 --> 00:13:45,380 Speaker 2: of cases where it's like, really? That's an eyebrow-raising declaration 263 00:13:45,390 --> 00:13:50,640 Speaker 2: of But there are lots of colorful characters in broker markets. 264 00:13:51,559 --> 00:13:56,260 Speaker 2: It's actually a very sophisticated black market. There are contracts. 265 00:13:56,320 --> 00:14:00,590 Speaker 2: Brokers go and solicit hardware on behalf of specific suppliers. 266 00:14:00,650 --> 00:14:05,150 Speaker 2: There was one broker who complained to us that the 267 00:14:05,190 --> 00:14:06,750 Speaker 2: trade is so good and so many people are trying 268 00:14:06,770 --> 00:14:08,439 Speaker 2: to get in on it. that now you have these 269 00:14:08,500 --> 00:14:11,000 Speaker 2: scammers in the black market who are like purporting to 270 00:14:11,040 --> 00:14:11,540 Speaker 2: have hardware. 271 00:14:11,679 --> 00:14:12,490 Speaker 4: Oh, yeah. 272 00:14:12,540 --> 00:14:14,830 Speaker 5: Good, honest black markets. 273 00:14:14,510 --> 00:14:18,449 Speaker 2: For like legitimate, honest black market brokers. No, it's a 274 00:14:18,470 --> 00:14:20,860 Speaker 2: pretty extraordinary and sophisticated and sprawling ecosystem. 275 00:14:37,050 --> 00:14:41,210 Speaker 3: Just one slight detail. You said that technically an entity 276 00:14:41,290 --> 00:14:44,670 Speaker 3: like a Megaspeed or even a CoreWeave does not actually 277 00:14:44,750 --> 00:14:48,780 Speaker 3: buy the chips from NVIDIA directly? And that actually the 278 00:14:48,960 --> 00:14:51,540 Speaker 3: first buyer of the chips is the server maker? 279 00:14:52,600 --> 00:14:55,460 Speaker 2: Yeah, typically speaking, NVIDIA is not selling directly to the 280 00:14:55,480 --> 00:14:55,860 Speaker 2: end user. 281 00:14:56,000 --> 00:15:00,900 Speaker 3: But the end user gets the allocation. So the end user, 282 00:15:00,950 --> 00:15:04,550 Speaker 3: a CoreWeave, a Megaspeed, whoever, will get some allocation directly. 283 00:15:04,800 --> 00:15:07,380 Speaker 3: that NVIDIA decides, okay, they're gonna get this many, but 284 00:15:07,400 --> 00:15:11,090 Speaker 3: the actual chain first goes through one of the server 285 00:15:11,130 --> 00:15:13,490 Speaker 3: makers like a Dell or a Supermicro or something like that. 286 00:15:13,530 --> 00:15:13,770 Speaker 2: Correct. 287 00:15:14,010 --> 00:15:14,190 Speaker 6: Okay. 288 00:15:14,810 --> 00:15:16,450 Speaker 2: Yeah, that's what makes NVIDIA the kingmaker. 289 00:15:16,470 --> 00:15:17,070 Speaker 4: All right. 290 00:15:17,770 --> 00:15:22,580 Speaker 3: Explain, as you said, this question of smuggling is not 291 00:15:22,640 --> 00:15:26,800 Speaker 3: locked down, but you said inconsistencies discovered in your reporting. 292 00:15:27,040 --> 00:15:28,660 Speaker 3: Let's talk about these inconsistencies. 293 00:15:28,870 --> 00:15:29,500 Speaker 2: On Megaspeed? 294 00:15:29,570 --> 00:15:29,690 Speaker 5: Yeah. 295 00:15:29,770 --> 00:15:33,440 Speaker 2: All right, let's get into it. So, you know, one 296 00:15:33,480 --> 00:15:36,860 Speaker 2: of the sort of first ports of call that I 297 00:15:37,020 --> 00:15:39,160 Speaker 2: look at when I'm investing in a company for smuggling, 298 00:15:39,180 --> 00:15:41,120 Speaker 2: which has now become a big part of my job, 299 00:15:41,800 --> 00:15:45,850 Speaker 2: is do they have the requisite data center space? Like 300 00:15:46,030 --> 00:15:48,270 Speaker 2: literally just the physical space in the building, the power 301 00:15:48,290 --> 00:15:53,850 Speaker 2: supply to support the hardware that they have purchased. And we. 302 00:15:56,200 --> 00:15:59,060 Speaker 2: Learned at a point in our investigation the number of 303 00:15:59,120 --> 00:16:02,990 Speaker 2: GPUs that NVIDIA found when asked by the U.S. government 304 00:16:03,040 --> 00:16:05,950 Speaker 2: to go and check Megaspeed's facilities, which they did several times. 305 00:16:05,970 --> 00:16:08,850 Speaker 2: You know, they tabulated all of the chips across over 306 00:16:08,850 --> 00:16:12,130 Speaker 2: 12 data centers owned by eight companies in three countries. 307 00:16:12,950 --> 00:16:15,230 Speaker 2: They added up all the chips. They reported to the U.S. government. 308 00:16:15,570 --> 00:16:18,290 Speaker 2: And around this time, you know, we got a statement 309 00:16:18,330 --> 00:16:22,410 Speaker 2: from them in mid-November last year that said everything that 310 00:16:22,450 --> 00:16:23,810 Speaker 2: was shipped to our partners is where it should be. 311 00:16:25,790 --> 00:16:31,790 Speaker 2: We compared that number to Malaysian and Indonesian trade records 312 00:16:31,970 --> 00:16:34,990 Speaker 2: that we had gotten and found that there were about 313 00:16:34,990 --> 00:16:39,140 Speaker 2: 50,000 chips unaccounted for, which is like not a small 314 00:16:39,200 --> 00:16:42,200 Speaker 2: amount of compute. And most of them were Blackwell chips, 315 00:16:42,420 --> 00:16:46,140 Speaker 2: which at the time were NVIDIA's best hardware. And so 316 00:16:46,200 --> 00:16:48,660 Speaker 2: we went back to NVIDIA and we said, you know, 317 00:16:48,680 --> 00:16:50,830 Speaker 2: we're not talking about one or two chips that are missing. 318 00:16:50,850 --> 00:16:54,570 Speaker 2: We're talking about a significant proportion of the overall hardware 319 00:16:54,590 --> 00:16:58,000 Speaker 2: that Megaspeed has purchased. And after we asked about that, 320 00:16:58,120 --> 00:17:01,540 Speaker 2: they went back to Malaysia and they said, oh, we 321 00:17:01,620 --> 00:17:06,700 Speaker 2: found these warehouses that contain what's missing. And I said, OK, great. 322 00:17:07,080 --> 00:17:11,369 Speaker 2: So we're going to subtract from the amount of hardware 323 00:17:11,390 --> 00:17:14,540 Speaker 2: that we found in trade records and the number of 324 00:17:14,600 --> 00:17:17,400 Speaker 2: chips that we know you guys told the U.S. government 325 00:17:17,440 --> 00:17:19,320 Speaker 2: you found, and that should be the amount of hardware 326 00:17:19,359 --> 00:17:22,600 Speaker 2: that's in these warehouses that we didn't know existed until 327 00:17:22,619 --> 00:17:24,020 Speaker 2: a week ago. We have been talking to NVIDIA about 328 00:17:24,040 --> 00:17:27,060 Speaker 2: this for a really long time by this point. This 329 00:17:27,119 --> 00:17:29,250 Speaker 2: is the number, right? And they declined to comment on 330 00:17:29,350 --> 00:17:31,570 Speaker 2: whether that's the number. I was like, okay. And then 331 00:17:31,590 --> 00:17:36,250 Speaker 2: they eventually submitted an updated statement that said we've identified, quote-unquote, 332 00:17:36,290 --> 00:17:40,030 Speaker 2: substantially all of the chips that Megaspeed had purchased via 333 00:17:40,070 --> 00:17:43,010 Speaker 2: its hardware partners. And we'll do additional checks in the 334 00:17:43,030 --> 00:17:45,950 Speaker 2: near future. So that was December of last year. I've 335 00:17:46,090 --> 00:17:49,470 Speaker 2: probably asked NVIDIA 10 to 12 times over the past 336 00:17:49,510 --> 00:17:52,930 Speaker 2: year whether those additional checks have occurred. I'm interested in 337 00:17:52,990 --> 00:17:55,940 Speaker 2: particular in a handful of data centers. One, which is 338 00:17:55,960 --> 00:17:59,600 Speaker 2: Megaspeed's largest data center, dropped the company as a tenant 339 00:17:59,840 --> 00:18:03,540 Speaker 2: a couple of months after our investigation and a story 340 00:18:03,600 --> 00:18:05,740 Speaker 2: by the New York Times that first broke the news 341 00:18:05,810 --> 00:18:08,240 Speaker 2: of the U.S. investigation into Megaspeed. I don't know the 342 00:18:08,359 --> 00:18:10,340 Speaker 2: reason that they were dropped, but if they're no longer 343 00:18:10,380 --> 00:18:12,740 Speaker 2: a tenant and that's where a significant portion of the 344 00:18:12,830 --> 00:18:16,129 Speaker 2: NVIDIA GPUs that NVIDIA told the U.S. government it found 345 00:18:16,170 --> 00:18:19,070 Speaker 2: were located, that raises some questions about where that hardware 346 00:18:19,090 --> 00:18:22,590 Speaker 2: might be now. We've not gotten answers to those questions. 347 00:18:22,770 --> 00:18:25,270 Speaker 2: So that's one of the eight data center operators. Two 348 00:18:25,310 --> 00:18:28,710 Speaker 2: of the others directly contradicted to us on the record 349 00:18:28,880 --> 00:18:31,780 Speaker 2: NVIDIA's account of the hardware as described to the U.S. government. 350 00:18:32,250 --> 00:18:34,710 Speaker 2: And so that's of the hardware that NVIDIA said it 351 00:18:34,750 --> 00:18:36,380 Speaker 2: found in the data centers. And then you have this 352 00:18:36,390 --> 00:18:39,820 Speaker 2: stuff that's, you know, in these warehouses. And we asked, okay, 353 00:18:39,840 --> 00:18:44,139 Speaker 2: so NVIDIA's policy is, you know, we prioritize customers that 354 00:18:44,160 --> 00:18:46,840 Speaker 2: are ready to put their products to use. They don't 355 00:18:46,880 --> 00:18:49,590 Speaker 2: want GPUs sitting in warehouses for months and months on end. 356 00:18:49,650 --> 00:18:52,869 Speaker 2: We asked, you know, what is Megaspeed's plan with this hardware? 357 00:18:52,910 --> 00:18:54,210 Speaker 2: When are we going to see it installed in the 358 00:18:54,250 --> 00:18:57,869 Speaker 2: data center? We didn't get answers to those questions. So 359 00:18:57,910 --> 00:19:03,980 Speaker 2: that's just on the hardware side. allocation piece. Then we 360 00:19:04,020 --> 00:19:09,139 Speaker 2: look at what Megaspeed has represented to investors and its 361 00:19:09,200 --> 00:19:12,760 Speaker 2: data center footprint as the company describes it. And on 362 00:19:12,800 --> 00:19:15,899 Speaker 2: its website, it had three data centers listed. We know 363 00:19:15,920 --> 00:19:17,919 Speaker 2: that's an undercount of their total number of data centers, 364 00:19:18,140 --> 00:19:23,000 Speaker 2: all in Southeast Asia. But in a presentation to investors 365 00:19:23,200 --> 00:19:28,530 Speaker 2: in 2024, Megaspeed described five data centers across three countries 366 00:19:28,550 --> 00:19:32,919 Speaker 2: and Two in Indonesia, two in Malaysia, and one in 367 00:19:33,040 --> 00:19:36,399 Speaker 2: a specific area of an unnamed third country that on 368 00:19:36,460 --> 00:19:39,940 Speaker 2: another slide of that same investor deck was the only 369 00:19:40,020 --> 00:19:44,470 Speaker 2: one described as not in Southeast Asia. And the rendering 370 00:19:44,510 --> 00:19:48,210 Speaker 2: that Megaspeed used to represent that data center is of 371 00:19:48,310 --> 00:19:53,520 Speaker 2: a sprawling project in Shanghai. where Megaspeed's sort of Chinese 372 00:19:53,580 --> 00:19:57,420 Speaker 2: corporate twin has been hiring engineers to work on restricted 373 00:19:57,440 --> 00:20:02,720 Speaker 2: NVIDIA hardware. Companies that Alice, who has become sort of 374 00:20:02,830 --> 00:20:06,270 Speaker 2: infamous in this scandal over the past couple of weeks, 375 00:20:06,720 --> 00:20:11,020 Speaker 2: Companies that she has long-time business ties to have been 376 00:20:11,119 --> 00:20:15,760 Speaker 2: investors in that data center. And so we asked, what's 377 00:20:15,820 --> 00:20:17,960 Speaker 2: up with this product in Shanghai? Why is that the 378 00:20:18,020 --> 00:20:20,389 Speaker 2: rendering that Megaspeed used to represent this facility on its 379 00:20:20,430 --> 00:20:23,190 Speaker 2: investor deck? We never got answers to those questions either. 380 00:20:24,170 --> 00:20:26,770 Speaker 5: It's such a weird decision to talk about, oh, and 381 00:20:26,790 --> 00:20:28,830 Speaker 5: we also have this project in a country that we're 382 00:20:28,869 --> 00:20:34,410 Speaker 5: not going to name, but here's an illustration of Shanghai. Yeah. OK, 383 00:20:34,490 --> 00:20:37,260 Speaker 5: and just so we're clear, can you outline what exactly 384 00:20:37,300 --> 00:20:41,080 Speaker 5: are the responsibilities of NVIDIA here? So if it sells 385 00:20:41,220 --> 00:20:45,070 Speaker 5: chips to a server company and the server company declares 386 00:20:45,130 --> 00:20:49,050 Speaker 5: that like the end user is definitely not a Chinese firm, 387 00:20:49,650 --> 00:20:53,130 Speaker 5: is it incumbent on NVIDIA to like double check those 388 00:20:53,230 --> 00:20:56,659 Speaker 5: claims or investigate? Like what do they actually what are 389 00:20:56,680 --> 00:20:57,419 Speaker 5: they expected to do? 390 00:20:58,330 --> 00:21:02,070 Speaker 2: You know, this is sort of the central tension that 391 00:21:02,109 --> 00:21:05,100 Speaker 2: has been playing out over the past year is whose 392 00:21:05,160 --> 00:21:08,139 Speaker 2: job is it to know? So, you know, in the 393 00:21:08,160 --> 00:21:11,120 Speaker 2: case of Megaspeed, NVIDIA told us last year we asked 394 00:21:11,160 --> 00:21:13,530 Speaker 2: the U.S. government if we need to cut off Megaspeed 395 00:21:13,619 --> 00:21:18,250 Speaker 2: and we were told no. In the case where, you know, 396 00:21:18,290 --> 00:21:21,229 Speaker 2: Supermicro's co-founders alleged to have smuggled two and a half 397 00:21:21,250 --> 00:21:24,879 Speaker 2: billion or sold $ 2. 5 billion worth of hardware to an 398 00:21:25,010 --> 00:21:27,629 Speaker 2: unnamed Southeast Asian intermediary that then smuggled the majority of 399 00:21:27,670 --> 00:21:30,210 Speaker 2: it to China. We've reported that that Southeast Asian intermediary 400 00:21:30,410 --> 00:21:33,090 Speaker 2: is one of the most important AI companies in Thailand. 401 00:21:34,250 --> 00:21:37,660 Speaker 2: That company is alleged to never have had the data 402 00:21:37,680 --> 00:21:41,100 Speaker 2: centers to install that hardware. So a lot of people 403 00:21:41,140 --> 00:21:44,899 Speaker 2: have asked, did NVIDIA check did Supermicro check. And actually, 404 00:21:44,980 --> 00:21:49,290 Speaker 2: Supermicro did check, according to the U.S. indictment. They audited 405 00:21:49,350 --> 00:21:52,910 Speaker 2: because the volumes that this Thai company was purchasing were 406 00:21:52,950 --> 00:21:55,970 Speaker 2: so significant. They audited, but then they allowed the shipments 407 00:21:55,990 --> 00:22:01,470 Speaker 2: to continue multiple times. I think companies are always going 408 00:22:01,490 --> 00:22:03,840 Speaker 2: to say, tell us what the law is and we 409 00:22:03,859 --> 00:22:05,580 Speaker 2: will follow the law. If you don't want us to 410 00:22:05,600 --> 00:22:07,780 Speaker 2: sell to this particular company, then tell us we can't 411 00:22:07,800 --> 00:22:11,270 Speaker 2: sell to that company. But I think the question is, 412 00:22:11,550 --> 00:22:15,409 Speaker 2: are there red flags that are so obvious that if 413 00:22:15,450 --> 00:22:18,409 Speaker 2: you have sort of any competent due diligence program, you 414 00:22:18,430 --> 00:22:21,650 Speaker 2: should have seen that? This is it's it's a conversation 415 00:22:21,670 --> 00:22:24,900 Speaker 2: that's sort of being actively litigated in Washington, shall we say. 416 00:22:25,160 --> 00:22:31,060 Speaker 2: And NVIDIA has started actually in some ways to self-regulate. Um, 417 00:22:31,200 --> 00:22:35,060 Speaker 2: you know, they have this like top tier of NVIDIA 418 00:22:35,119 --> 00:22:38,690 Speaker 2: cloud partners, these sort of like stamp of approval companies 419 00:22:38,710 --> 00:22:41,430 Speaker 2: that are totally, you tell the industry, like these are, 420 00:22:41,710 --> 00:22:43,689 Speaker 2: these are our top tier, uh, these are our top 421 00:22:43,710 --> 00:22:46,190 Speaker 2: tier partners. Megaspeed is, is one of them, is still 422 00:22:46,270 --> 00:22:47,830 Speaker 2: one of them. It became one of them four days 423 00:22:47,930 --> 00:22:53,580 Speaker 2: after it was incorporated in Singapore. Another NCP, its CEO 424 00:22:53,619 --> 00:22:58,740 Speaker 2: is on trial in Singapore for, uh, document for defrauding 425 00:22:58,780 --> 00:23:01,760 Speaker 2: server suppliers about the ultimate destination of hardware. And there 426 00:23:01,780 --> 00:23:04,100 Speaker 2: was like a mansion seize that was the alleged approach 427 00:23:04,119 --> 00:23:08,180 Speaker 2: seize of chip smuggling. The Thai company was like sort 428 00:23:08,200 --> 00:23:14,330 Speaker 2: of the initial hardware supplier for the like national AI company, 429 00:23:14,369 --> 00:23:17,450 Speaker 2: Sovereign Cloud of Thailand, which is the first NCP in Thailand. 430 00:23:17,770 --> 00:23:19,990 Speaker 2: So you have like definitely this interesting set of connections 431 00:23:20,010 --> 00:23:21,109 Speaker 2: happening at the NCP layer. 432 00:23:21,390 --> 00:23:23,870 Speaker 4: The NCP stands for? 433 00:23:23,890 --> 00:23:25,330 Speaker 2: NVIDIA Cloud Partner. Oh, yeah, yeah. 434 00:23:25,369 --> 00:23:25,770 Speaker 4: Sorry, sorry. 435 00:23:26,250 --> 00:23:29,230 Speaker 2: Sorry. Yeah, yeah. So there's also like an internal whitelist 436 00:23:29,790 --> 00:23:33,090 Speaker 2: of companies that NVIDIA has sort of pre-approved for purchase. 437 00:23:33,869 --> 00:23:38,650 Speaker 2: And after the Supermicro co-founder's indictment earlier this year, NVIDIA 438 00:23:39,050 --> 00:23:41,430 Speaker 2: went and cleaved a lot of companies from that whitelist. 439 00:23:42,130 --> 00:23:45,290 Speaker 2: And it's this interesting thing because on the one hand, 440 00:23:45,350 --> 00:23:48,590 Speaker 2: this is the type of due diligence maybe that some 441 00:23:48,630 --> 00:23:51,429 Speaker 2: Washington regulators would like to see the chipmaker take. On 442 00:23:51,450 --> 00:23:53,709 Speaker 2: the other, it raises questions of, you know, On what 443 00:23:53,760 --> 00:23:55,899 Speaker 2: basis did you remove those companies? Why were they on 444 00:23:55,920 --> 00:23:58,439 Speaker 2: the whitelist in the first place? And so, you know, 445 00:23:58,460 --> 00:24:01,840 Speaker 2: I think also for NVIDIA, which is trying in some 446 00:24:01,859 --> 00:24:05,160 Speaker 2: ways to demonstrate to the U.S. that it does take 447 00:24:05,200 --> 00:24:09,470 Speaker 2: this seriously, but also has so deeply lost credibility on 448 00:24:09,490 --> 00:24:11,909 Speaker 2: this issue that it's like a little bit of a catch-22. 449 00:24:11,930 --> 00:24:13,929 Speaker 2: And that's not trying to absolve NVIDIA of responsibility, just 450 00:24:13,950 --> 00:24:16,490 Speaker 2: sort of explaining a political reality that the company now 451 00:24:16,530 --> 00:24:17,060 Speaker 2: finds itself in. 452 00:24:17,600 --> 00:24:19,770 Speaker 5: This was going to be my next question because NVIDIA, 453 00:24:19,850 --> 00:24:22,209 Speaker 5: I mean, the thing we hear constantly is chips are 454 00:24:22,270 --> 00:24:25,230 Speaker 5: in short supply. Everyone is scrambling to get chips. There's 455 00:24:25,270 --> 00:24:28,190 Speaker 5: so much demand out there. It's hard to fill it all. 456 00:24:28,609 --> 00:24:33,149 Speaker 5: They can afford to be selective with their customers, right? 457 00:24:33,540 --> 00:24:37,760 Speaker 5: Do you have a working theory for why would they 458 00:24:37,880 --> 00:24:40,959 Speaker 5: even take the chance if demand is so strong that 459 00:24:40,980 --> 00:24:42,340 Speaker 5: they could sell to pretty much anyone? 460 00:24:42,740 --> 00:24:49,040 Speaker 2: That's a good question. NVIDIA's primary arguments on export controls 461 00:24:49,340 --> 00:24:53,740 Speaker 2: are about the market share, the global market share of 462 00:24:53,859 --> 00:24:59,230 Speaker 2: AI infrastructure that is American. They sort of view Huawei, 463 00:24:59,290 --> 00:25:02,970 Speaker 2: their primary competitor in China, as this sort of existential 464 00:25:03,010 --> 00:25:06,070 Speaker 2: threat to U.S. dominance. And their arguments are basically that 465 00:25:06,130 --> 00:25:10,040 Speaker 2: NVIDIA's interests are U.S. interests. And so, you know, when 466 00:25:10,080 --> 00:25:14,680 Speaker 2: you talk about the question of remote access, you renting 467 00:25:14,800 --> 00:25:17,340 Speaker 2: the use of NVIDIA hardware that Chinese companies can't purchase 468 00:25:17,359 --> 00:25:20,900 Speaker 2: themselves to those Chinese companies. You know, that's a policy 469 00:25:20,940 --> 00:25:22,760 Speaker 2: choice by the U.S. government. You know, a lot of 470 00:25:22,780 --> 00:25:24,970 Speaker 2: people describe that as like, oh, it's a workaround of 471 00:25:25,010 --> 00:25:27,150 Speaker 2: the export controls. It's the loophole of the export controls. 472 00:25:27,830 --> 00:25:30,710 Speaker 2: It's legal. The U.S. has the ability to make it 473 00:25:30,830 --> 00:25:33,330 Speaker 2: illegal and it's legal in the vast majority of cases. 474 00:25:33,790 --> 00:25:36,790 Speaker 2: You know, for NVIDIA, that's a good thing because NVIDIA 475 00:25:36,810 --> 00:25:39,040 Speaker 2: chips are filling all these data centers in Southeast Asia 476 00:25:39,080 --> 00:25:41,820 Speaker 2: that in many cases wouldn't exist if not for the 477 00:25:41,880 --> 00:25:45,100 Speaker 2: offtake demand from Chinese companies. It's the same reason that 478 00:25:45,180 --> 00:25:47,420 Speaker 2: NVIDIA wants to be able to freely sell its chips 479 00:25:47,460 --> 00:25:49,219 Speaker 2: in China. We need to be able to compete with 480 00:25:49,270 --> 00:25:52,890 Speaker 2: Huawei on its own turf. A question that a lot 481 00:25:52,950 --> 00:25:56,050 Speaker 2: of regulators have pushed back on NVIDIA on this topic 482 00:25:56,130 --> 00:25:59,850 Speaker 2: is Huawei can't supply the world's data centers. We're so 483 00:25:59,910 --> 00:26:01,649 Speaker 2: far off from a world of Huawei being able to 484 00:26:01,670 --> 00:26:04,330 Speaker 2: supply the world's data centers. And so how existential is 485 00:26:04,369 --> 00:26:07,290 Speaker 2: the threat of Chinese compute on the global stage or 486 00:26:07,330 --> 00:26:10,909 Speaker 2: even within China that the NVIDIA sales are aiming to counteract? 487 00:26:11,570 --> 00:26:14,370 Speaker 2: But I think NVIDIA is very focused on these sort 488 00:26:14,390 --> 00:26:18,520 Speaker 2: of battleground or what could become intense battleground regions if 489 00:26:18,560 --> 00:26:22,000 Speaker 2: China builds its own AI ship production capacity that they 490 00:26:22,020 --> 00:26:24,460 Speaker 2: want to make sure that they're selling in these regions. 491 00:26:24,780 --> 00:26:26,200 Speaker 4: I just want to point out one thing. 492 00:26:26,700 --> 00:26:30,500 Speaker 3: We've talked to Ian Dunning twice from Hudson River Trading, 493 00:26:30,560 --> 00:26:33,790 Speaker 3: who has said he is not having a hard time 494 00:26:33,830 --> 00:26:36,790 Speaker 3: getting GPUs. He's having a hard time getting GPUs. GPU 495 00:26:36,869 --> 00:26:37,810 Speaker 3: places to plug in. 496 00:26:38,109 --> 00:26:38,350 Speaker 6: Yeah. 497 00:26:38,690 --> 00:26:42,580 Speaker 1: So one could, it's possible, at least it seems possible 498 00:26:43,119 --> 00:26:47,060 Speaker 1: that there is not as much like unlimited demand for 499 00:26:47,119 --> 00:26:51,320 Speaker 1: GPUs per se as maybe NVIDIA would. 500 00:26:51,220 --> 00:26:52,080 Speaker 4: Like to claim. 501 00:26:52,520 --> 00:26:55,699 Speaker 3: And that actually, like, let's say the bottleneck is in 502 00:26:55,900 --> 00:26:58,050 Speaker 3: places to plug them in, that if you want to 503 00:26:58,170 --> 00:27:00,630 Speaker 3: actually sell the most number of chips for the most 504 00:27:00,670 --> 00:27:03,590 Speaker 3: number of dollars, setting aside these sort of broad market 505 00:27:03,609 --> 00:27:07,650 Speaker 3: share questions, that actually the Southeast Asian customers are important 506 00:27:08,090 --> 00:27:11,770 Speaker 3: if we believe what we've heard from other purchasers of 507 00:27:11,869 --> 00:27:15,770 Speaker 3: NVIDIA GPUs, that they're available if you can plug them in. 508 00:27:15,810 --> 00:27:19,920 Speaker 3: But before we get into the sort of NVIDIA versus 509 00:27:19,940 --> 00:27:22,159 Speaker 3: Huawei part, which I do want to talk about, I 510 00:27:22,480 --> 00:27:26,780 Speaker 3: want to just clarify this remote access question. Are you 511 00:27:26,820 --> 00:27:29,540 Speaker 3: saying like, let's say I start a data center company 512 00:27:30,380 --> 00:27:33,400 Speaker 3: and I'm explicitly like, I'm going to build data centers 513 00:27:33,660 --> 00:27:38,480 Speaker 3: in Malaysia. and I'm going to rent them out to 514 00:27:39,590 --> 00:27:43,350 Speaker 3: deep seek and a bunch of other like Chinese frontier 515 00:27:43,530 --> 00:27:46,150 Speaker 3: AI companies, technically. And I'm going to buy a video, 516 00:27:46,170 --> 00:27:48,270 Speaker 3: like technically there is nothing as far as we know, 517 00:27:48,310 --> 00:27:51,730 Speaker 3: like per se illegal about that arrangement. And then beyond that, 518 00:27:52,190 --> 00:27:54,889 Speaker 3: how do the Hawks or the people who argue in 519 00:27:54,990 --> 00:28:01,760 Speaker 3: favor of export controls into mainland China, that seems really inconsistent. 520 00:28:02,020 --> 00:28:06,240 Speaker 1: Like, Can you walk me through, like, is that actually legal? 521 00:28:06,320 --> 00:28:08,780 Speaker 3: And how is that, you say it's not a loophole 522 00:28:08,840 --> 00:28:12,550 Speaker 3: that it's the law, but how is that law justified? 523 00:28:15,119 --> 00:28:17,129 Speaker 2: Yeah, in the vast majority of cases, this is legal. 524 00:28:17,310 --> 00:28:19,890 Speaker 2: I mean, this is like literally what Megaspeed's business is. 525 00:28:19,910 --> 00:28:22,109 Speaker 2: This is what a lot of Southeast Asian neoclouds do. 526 00:28:22,830 --> 00:28:25,270 Speaker 2: The exceptions to that would be if you know that 527 00:28:25,290 --> 00:28:29,010 Speaker 2: the compute is going to be used for intelligence activities 528 00:28:29,030 --> 00:28:32,070 Speaker 2: by a foreign adversary military. It's a very narrow band 529 00:28:32,109 --> 00:28:32,570 Speaker 2: of restrictions. 530 00:28:32,900 --> 00:28:37,560 Speaker 3: But there's no restrictions on like selling chips into Southeast 531 00:28:37,660 --> 00:28:41,520 Speaker 3: Asia to be used to train the most advanced open 532 00:28:41,580 --> 00:28:43,060 Speaker 3: source models from China. 533 00:28:43,120 --> 00:28:45,740 Speaker 4: There's actually nothing illegal about it. 534 00:28:45,740 --> 00:28:50,460 Speaker 5: Doesn't DeepSea rents from like a Southeast Asia data center, right? 535 00:28:51,800 --> 00:28:56,570 Speaker 2: So the U.S. government has said that Moonshot rents NVIDIA 536 00:28:56,610 --> 00:28:57,690 Speaker 2: compute in Thailand. 537 00:28:57,850 --> 00:28:58,190 Speaker 6: Okay. 538 00:28:58,930 --> 00:29:01,680 Speaker 2: DeepSeq, by my understanding, actually does more of it, has 539 00:29:01,700 --> 00:29:06,180 Speaker 2: more of its compute onshore. But, you know, generally speaking 540 00:29:06,520 --> 00:29:10,770 Speaker 2: across Chinese companies, you know, this is an existentially important 541 00:29:10,810 --> 00:29:14,510 Speaker 2: resource is NVIDIA compute in Southeast Asia. You ask, you know, 542 00:29:14,650 --> 00:29:16,970 Speaker 2: how is it that the hawks have not restricted this? 543 00:29:17,070 --> 00:29:20,860 Speaker 2: They actually did. They restricted it at the end of 544 00:29:20,920 --> 00:29:24,900 Speaker 2: the Biden administration. And it was part of this broad 545 00:29:24,980 --> 00:29:28,980 Speaker 2: set of global restrictions on NVIDIA hardware called the AI 546 00:29:29,020 --> 00:29:33,630 Speaker 2: diffusion rule that was set to take effect in May 2025. 547 00:29:33,630 --> 00:29:36,310 Speaker 2: And the Trump administration is not enforcing that regulation. 548 00:29:36,890 --> 00:29:41,400 Speaker 3: Sorry, just to be clear, NVIDIA still can't... sell their 549 00:29:41,440 --> 00:29:46,760 Speaker 3: most advanced chips into mainland China. But the Trump administration 550 00:29:47,240 --> 00:29:50,940 Speaker 3: so far has been cool with sell them into Southeast 551 00:29:51,000 --> 00:29:52,770 Speaker 3: Asia to then remote access. 552 00:29:53,050 --> 00:29:53,850 Speaker 4: I mean, it's the internet. 553 00:29:53,910 --> 00:29:55,610 Speaker 1: I don't even know what borders mean. Why are like, 554 00:29:56,950 --> 00:29:57,470 Speaker 1: It's the internet. 555 00:29:57,490 --> 00:29:59,190 Speaker 4: What are we talking about? What does it matter? 556 00:29:59,270 --> 00:30:05,150 Speaker 2: Export controls on hardware don't respect the cloud. Yeah, yeah. No, 557 00:30:05,250 --> 00:30:08,380 Speaker 2: it's sometimes a bit baffling. 558 00:30:08,440 --> 00:30:11,560 Speaker 1: As long as I'm not the only one who's baffled. 559 00:30:12,480 --> 00:30:15,520 Speaker 2: This is like an active, raging debate in Washington, to 560 00:30:15,560 --> 00:30:18,800 Speaker 2: be fair. And, you know, one interesting thing that happened 561 00:30:18,860 --> 00:30:23,160 Speaker 2: this summer. So, you know, after Kimmy comes out with K3, 562 00:30:23,540 --> 00:30:25,520 Speaker 2: it's sort of, it's like, seen as like a second 563 00:30:25,560 --> 00:30:30,740 Speaker 2: deep-seek moment almost. The science and tech advisor in the 564 00:30:30,760 --> 00:30:35,760 Speaker 2: White House, Michael Kratzios, posts on X that the Moonshot 565 00:30:35,780 --> 00:30:39,170 Speaker 2: team acquired Blackwell chips, which is illegal. It raises questions 566 00:30:39,210 --> 00:30:41,050 Speaker 2: of why has the U.S. government not, you know, why 567 00:30:41,070 --> 00:30:43,370 Speaker 2: does the Commerce Department not take any action if that's 568 00:30:43,410 --> 00:30:46,130 Speaker 2: the U.S. government's determination? And also that they had remotely 569 00:30:46,190 --> 00:30:48,250 Speaker 2: accessed these chips in Thailand. And a couple of days 570 00:30:48,330 --> 00:30:51,800 Speaker 2: after that, the enforcement unit of the part of the 571 00:30:51,820 --> 00:30:55,240 Speaker 2: Commerce Department that handles chip export control. So the people 572 00:30:55,260 --> 00:30:57,760 Speaker 2: who are like, oh, you violated, like you smuggled, we're 573 00:30:57,780 --> 00:31:02,060 Speaker 2: going to go after you. That unit started mapping where 574 00:31:02,100 --> 00:31:05,540 Speaker 2: in the world remote access happens, which is very interesting 575 00:31:05,640 --> 00:31:08,680 Speaker 2: because remote access is not illegal. And so if you're 576 00:31:08,700 --> 00:31:12,040 Speaker 2: an enforcer of export controls, you know, why is that 577 00:31:12,060 --> 00:31:16,140 Speaker 2: your purview? But they started sort of crafting this list 578 00:31:16,380 --> 00:31:19,660 Speaker 2: of countries where that's known to happen and And, you know, 579 00:31:19,700 --> 00:31:21,540 Speaker 2: there have been at a very senior level of the 580 00:31:21,560 --> 00:31:27,370 Speaker 2: Trump administration, pretty intense debate over whether to impose such restrictions. 581 00:31:28,130 --> 00:31:30,590 Speaker 2: There's also interest on Capitol Hill. There's interest in Capitol 582 00:31:30,610 --> 00:31:33,440 Speaker 2: Hill in being more hawkish on a lot of these policies, 583 00:31:33,820 --> 00:31:37,920 Speaker 2: but including remote access. It's, you know, this is sort 584 00:31:37,960 --> 00:31:41,460 Speaker 2: of nightmare scenario for NVIDIA if remote access were to 585 00:31:41,500 --> 00:31:44,920 Speaker 2: be restricted. It's also a nightmare scenario for some hyperscalers 586 00:31:44,940 --> 00:31:50,100 Speaker 2: like Oracle that counts Chinese offtakers as a significant portion 587 00:31:50,140 --> 00:31:52,959 Speaker 2: of its business. There was great reporting in the Financial 588 00:31:52,980 --> 00:31:56,640 Speaker 2: Times recently about Oracle having a cloud computing arrangement with 589 00:31:56,680 --> 00:32:00,440 Speaker 2: Tencent that then the Trump administration stepped in to kill, basically, 590 00:32:00,850 --> 00:32:02,560 Speaker 2: even though it was legal. And so it's like, OK, 591 00:32:02,590 --> 00:32:04,790 Speaker 2: are we just going to do this ad hoc? You know, 592 00:32:04,810 --> 00:32:06,310 Speaker 2: they do a lot of things ad hoc, but we'll see. 593 00:32:06,330 --> 00:32:09,080 Speaker 2: You know, there's it's an active regulatory debate. Like, you know, 594 00:32:09,120 --> 00:32:11,100 Speaker 2: the answer to this question could change by the time 595 00:32:11,120 --> 00:32:12,620 Speaker 2: this podcast airs, although I doubt it. 596 00:32:12,780 --> 00:32:15,360 Speaker 5: Well, OK, speaking of ad hoc, I mean, part of 597 00:32:15,400 --> 00:32:20,600 Speaker 5: the scandal here is that NVIDIA reportedly intervened in the U.S. 598 00:32:20,620 --> 00:32:24,850 Speaker 5: government's Megaspeed investigation and said, yeah, it's all good. 599 00:32:25,030 --> 00:32:25,810 Speaker 4: Yeah, they're cool. 600 00:32:26,140 --> 00:32:28,840 Speaker 5: They're cool. Trust us. What do you expect to happen? 601 00:32:29,120 --> 00:32:31,590 Speaker 5: I don't even know the status of that investigation now, 602 00:32:31,640 --> 00:32:34,170 Speaker 5: but like, is it picking up again or like what's 603 00:32:34,210 --> 00:32:36,550 Speaker 5: going on there? Would you expect the DOJ to actually 604 00:32:36,610 --> 00:32:37,150 Speaker 5: pursue those? 605 00:32:38,470 --> 00:32:41,410 Speaker 2: Yeah, yeah. So I just want to clarify quickly on 606 00:32:41,650 --> 00:32:45,090 Speaker 2: the sort of scope of NVIDIA's intervention. So this is 607 00:32:45,370 --> 00:32:47,910 Speaker 2: reporting from like two days ago. And this is such 608 00:32:47,930 --> 00:32:50,600 Speaker 2: an active story that I'm saying what I know as 609 00:32:50,640 --> 00:32:54,820 Speaker 2: of 10.38 p.m. Hong Kong time on Wednesday. So NVIDIA's 610 00:32:54,860 --> 00:32:58,410 Speaker 2: general counsel, a guy called Tim Teeter, called NVIDIA. the 611 00:32:58,650 --> 00:33:00,540 Speaker 2: head of the Bureau of Industry and Security, like the 612 00:33:00,640 --> 00:33:04,600 Speaker 2: export controls enforcement and the agency that writes and enforces 613 00:33:04,640 --> 00:33:07,400 Speaker 2: export controls at the Commerce Department in late summer last 614 00:33:07,460 --> 00:33:12,880 Speaker 2: year to inquire about whether BIS was involved in the 615 00:33:12,920 --> 00:33:16,390 Speaker 2: detention of Alice, who we now know was, you know, 616 00:33:16,430 --> 00:33:18,730 Speaker 2: having an extramarital affair with Jensen. But NVIDIA says that 617 00:33:18,770 --> 00:33:20,610 Speaker 2: Tim Teeter didn't know that at the time of this call, 618 00:33:21,210 --> 00:33:23,450 Speaker 2: called to ask if BIS was involved in her being 619 00:33:23,490 --> 00:33:26,550 Speaker 2: detained in Singapore. she had been brought in for questioning 620 00:33:26,690 --> 00:33:29,690 Speaker 2: and restricted from leaving the country as part of Singapore's 621 00:33:29,800 --> 00:33:34,980 Speaker 2: own investigation into Megaspeed, which Megaspeed said recently had concluded 622 00:33:35,080 --> 00:33:38,120 Speaker 2: this year with no adverse findings. So Tim Teeter called 623 00:33:38,140 --> 00:33:42,920 Speaker 2: to ask about that. He later emailed BIS staff who 624 00:33:42,940 --> 00:33:44,520 Speaker 2: were working on the case and asked to meet with 625 00:33:44,600 --> 00:33:47,740 Speaker 2: one of them. And they forwarded those emails to Jeffrey 626 00:33:47,760 --> 00:33:51,040 Speaker 2: Kessler because Tim Teeter invoked Kessler's name. He's like, oh, 627 00:33:51,060 --> 00:33:52,570 Speaker 2: like Kessler said it was OK to talk to you. 628 00:33:52,580 --> 00:33:55,010 Speaker 2: And this was like this is not normal practice for 629 00:33:55,030 --> 00:33:59,310 Speaker 2: an export enforcement investigation. And, you know, Kessler said, continue 630 00:33:59,330 --> 00:34:02,229 Speaker 2: doing things by the book. You know, the Commerce Department said, 631 00:34:02,650 --> 00:34:05,190 Speaker 2: or a spokesperson for this agency within the Commerce Department said, 632 00:34:05,210 --> 00:34:08,770 Speaker 2: you know, everything has been above board. But the sort 633 00:34:08,790 --> 00:34:11,439 Speaker 2: of nature of the investigation and the ability of export 634 00:34:11,460 --> 00:34:15,580 Speaker 2: enforcement agents to act in the weeks and months after 635 00:34:15,800 --> 00:34:19,660 Speaker 2: this sort of saga with Tim Teeter was diminished. You know, 636 00:34:19,920 --> 00:34:23,160 Speaker 2: and agents or people working on export enforcement investigations you know, 637 00:34:23,520 --> 00:34:28,489 Speaker 2: we're restricted from requesting information directly from NVIDIA without getting 638 00:34:28,510 --> 00:34:31,450 Speaker 2: Kessler's approval beforehand or routing it through him. That's a 639 00:34:31,469 --> 00:34:34,630 Speaker 2: policy that's unique to NVIDIA. It's not how things operate 640 00:34:34,670 --> 00:34:37,340 Speaker 2: with any other major tech company. You were told not 641 00:34:37,380 --> 00:34:39,760 Speaker 2: to mention Megaspeed or Speed Matrix, the name of Megaspeed's 642 00:34:39,780 --> 00:34:44,420 Speaker 2: Malaysian subsidiary and external communications. And eventually the Department of Justice, 643 00:34:44,560 --> 00:34:48,910 Speaker 2: which typically pursues these investigations alongside BIS. So, you know, 644 00:34:48,950 --> 00:34:50,610 Speaker 2: if anything is going to be turned into like a 645 00:34:50,750 --> 00:34:53,810 Speaker 2: criminal indictment, it's going to involve the DOJ. The Department 646 00:34:53,830 --> 00:34:56,450 Speaker 2: of Justice had been working alongside BIS for a bit 647 00:34:56,530 --> 00:35:01,590 Speaker 2: on this case, ended up asking BIS to stop its efforts. 648 00:35:01,950 --> 00:35:04,340 Speaker 2: So now that case is with DOJ. You know, I 649 00:35:04,380 --> 00:35:06,719 Speaker 2: think my my understanding of the status of the case 650 00:35:06,760 --> 00:35:09,000 Speaker 2: a few weeks ago is would be probably very different 651 00:35:09,040 --> 00:35:11,900 Speaker 2: from where things are right now. There's obviously intense interest 652 00:35:12,140 --> 00:35:15,549 Speaker 2: in this within the Trump administration on Capitol Hill. So, 653 00:35:15,570 --> 00:35:17,210 Speaker 2: you know, it's funny, like a lot of a lot 654 00:35:17,230 --> 00:35:20,150 Speaker 2: of what we're talking about now was known such a 655 00:35:20,190 --> 00:35:22,170 Speaker 2: long time ago. You know, New Times reported on it. 656 00:35:22,210 --> 00:35:24,830 Speaker 2: We reported on it. But it's got a new life, 657 00:35:24,969 --> 00:35:26,330 Speaker 2: shall we say, in the past couple of weeks. 658 00:35:26,730 --> 00:35:29,359 Speaker 5: I feel like we should mention BIS is the Bureau 659 00:35:29,380 --> 00:35:31,920 Speaker 5: of Industry and Security, because whenever I hear BIS, I 660 00:35:31,960 --> 00:35:34,260 Speaker 5: always think Bank for International Settlements. 661 00:35:34,420 --> 00:35:37,680 Speaker 2: But yes. Different type of nerd, you guys. 662 00:35:37,700 --> 00:35:41,469 Speaker 4: You can only have so many three-letter acronyms, right? That's true. 663 00:35:41,510 --> 00:35:42,730 Speaker 4: And so you're bound to. 664 00:35:42,750 --> 00:35:45,850 Speaker 2: No, DC says every day we will invent a new one. Oh, yeah. 665 00:36:02,670 --> 00:36:04,880 Speaker 3: One more question on Megaspeed before I want to get 666 00:36:04,920 --> 00:36:06,880 Speaker 3: a little bit, talk a little bit about Huawei. One 667 00:36:06,920 --> 00:36:11,060 Speaker 3: more question on Megaspeed. What do we know right now? 668 00:36:11,080 --> 00:36:13,950 Speaker 3: You know, you described at the very beginning, spun off 669 00:36:13,989 --> 00:36:17,850 Speaker 3: from this gaming company, BVI shell company, etc. 670 00:36:18,290 --> 00:36:19,629 Speaker 4: What do we know. 671 00:36:19,910 --> 00:36:24,540 Speaker 3: About its owners and its backers and how they have 672 00:36:24,560 --> 00:36:26,300 Speaker 3: the money to buy a bunch of NVIDIA chips. 673 00:36:26,120 --> 00:36:26,600 Speaker 4: And all that stuff? 674 00:36:27,460 --> 00:36:32,080 Speaker 2: Yeah, so Megaspeed on paper is 100% owned by a 675 00:36:32,239 --> 00:36:34,740 Speaker 2: Singapore shell company that is 100% owned by a CEO 676 00:36:34,760 --> 00:36:37,529 Speaker 2: who is a Singapore national. We tried many times to 677 00:36:37,550 --> 00:36:39,669 Speaker 2: get in touch with him last year. We were unsuccessful. 678 00:36:40,350 --> 00:36:44,469 Speaker 2: In terms of where it got the money, this is 679 00:36:44,530 --> 00:36:48,009 Speaker 2: a big open question. $ 4. 6 billion worth of hardware is 680 00:36:48,030 --> 00:36:53,160 Speaker 2: not a small amount of hardware. It has a compute 681 00:36:53,219 --> 00:36:56,410 Speaker 2: rental agreement with Alibaba. Alibaba, you know, the way that 682 00:36:56,450 --> 00:36:58,370 Speaker 2: these agreements often work is a company will put sort 683 00:36:58,410 --> 00:37:01,430 Speaker 2: of a prepayment, sort of a guarantee, like, hey, like, 684 00:37:01,489 --> 00:37:02,850 Speaker 2: you know, it's like you put a deposit down on 685 00:37:02,870 --> 00:37:05,350 Speaker 2: an expensive purchase. Like, we are going to offtake this 686 00:37:05,370 --> 00:37:08,820 Speaker 2: in the future. So there are, you know, financial transactions 687 00:37:08,890 --> 00:37:13,840 Speaker 2: between Alibaba via a Singaporean subsidiary of Alibaba's two mega 688 00:37:13,860 --> 00:37:17,920 Speaker 2: speeds where, you know, the entirety of its Malaysia business 689 00:37:17,940 --> 00:37:20,660 Speaker 2: was offered as collateral actually. These are transactions that happened 690 00:37:20,680 --> 00:37:22,810 Speaker 2: a couple of years ago, but don't sort of fully 691 00:37:22,890 --> 00:37:24,870 Speaker 2: answer the question of where it got all of the 692 00:37:24,910 --> 00:37:28,010 Speaker 2: money that it did to purchase the hardware. But on paper, 693 00:37:28,070 --> 00:37:30,430 Speaker 2: it is a Singaporean company, and I don't have evidence 694 00:37:30,469 --> 00:37:31,110 Speaker 2: to the contrary. 695 00:37:31,250 --> 00:37:34,590 Speaker 3: Let's talk a little bit about, you mentioned Huawei, and 696 00:37:34,630 --> 00:37:38,330 Speaker 3: you mentioned that Huawei just doesn't have the capacity, the 697 00:37:38,370 --> 00:37:41,770 Speaker 3: skill yet to replace NVIDIA on a global level, probably 698 00:37:41,810 --> 00:37:44,530 Speaker 3: not even at a China level, given the demand. 699 00:37:45,360 --> 00:37:47,020 Speaker 4: Among Chinese AI firms. 700 00:37:47,120 --> 00:37:50,600 Speaker 2: Yeah, you know, Huawei's senior executives themselves say it'll take 701 00:37:50,700 --> 00:37:53,020 Speaker 2: until 2030 to satiate domestic demand. 702 00:37:53,040 --> 00:37:55,580 Speaker 3: All right, so talk a little bit about, like, where 703 00:37:55,760 --> 00:37:59,900 Speaker 3: are they on this process? And, you know, in AI, 704 00:37:59,920 --> 00:38:01,700 Speaker 3: we're always talking about, like, are they six months behind, 705 00:38:01,739 --> 00:38:02,800 Speaker 3: two years behind, whatever. 706 00:38:02,860 --> 00:38:04,180 Speaker 1: Like, tell us a little. 707 00:38:03,980 --> 00:38:08,140 Speaker 3: Bit about, like, where Huawei is in its fight, I guess, 708 00:38:08,160 --> 00:38:09,819 Speaker 3: to be, like, a real competitor with Nvidia. 709 00:38:10,469 --> 00:38:14,160 Speaker 2: Yeah, so, you know, Huawei is... is such a black box. 710 00:38:15,400 --> 00:38:18,779 Speaker 2: It's the sort of preeminent competitor to NVIDIA in China. 711 00:38:18,800 --> 00:38:21,299 Speaker 2: There are a handful of other companies, of course, that are, 712 00:38:21,440 --> 00:38:23,259 Speaker 2: you know, a lot of in the same way that 713 00:38:23,280 --> 00:38:25,200 Speaker 2: a lot of American companies are like designing their own chips. 714 00:38:25,219 --> 00:38:26,700 Speaker 2: You have that same pattern among a lot of the 715 00:38:26,719 --> 00:38:28,839 Speaker 2: Chinese companies. It's just a matter of like whether they 716 00:38:28,860 --> 00:38:31,930 Speaker 2: can get the production capacity. You know, that's the constraint 717 00:38:32,250 --> 00:38:34,969 Speaker 2: in large part because the U.S. and allied nations have 718 00:38:35,010 --> 00:38:38,540 Speaker 2: restricted the flow of semiconductor manufacturing equipment into China. But 719 00:38:38,580 --> 00:38:42,390 Speaker 2: Huawei has made material progress on its chips over the 720 00:38:42,410 --> 00:38:46,810 Speaker 2: past couple of years. And you're starting to see Chinese 721 00:38:46,950 --> 00:38:50,430 Speaker 2: companies become more interested in adopting Huawei hardware. You know, 722 00:38:50,450 --> 00:38:52,170 Speaker 2: I remember actually coming on OddLots a year ago and 723 00:38:52,190 --> 00:38:53,830 Speaker 2: I was like, you know, one of the things that 724 00:38:53,850 --> 00:38:55,730 Speaker 2: I'm waiting for is like, is Huawei ever going to 725 00:38:55,830 --> 00:38:58,780 Speaker 2: export Ascend chips, which are the name of its AI chips? 726 00:38:59,140 --> 00:39:01,000 Speaker 2: And also, are we going to see, you know, a 727 00:39:01,040 --> 00:39:04,899 Speaker 2: frontier Chinese lab complete a successful training run on Huawei chips? 728 00:39:05,250 --> 00:39:07,330 Speaker 2: You know, right now you're starting to see model releases 729 00:39:07,370 --> 00:39:09,350 Speaker 2: out of China that have what are called like zero-day 730 00:39:09,390 --> 00:39:11,990 Speaker 2: inferencing capabilities on Huawei, which is like, this is sort 731 00:39:12,030 --> 00:39:14,580 Speaker 2: of optimized to run best on Huawei hardware from the 732 00:39:14,620 --> 00:39:16,759 Speaker 2: moment that we release it. But that's a very different 733 00:39:16,800 --> 00:39:21,440 Speaker 2: thing from doing, you know, pre-training on Huawei. And, you know, 734 00:39:21,520 --> 00:39:25,750 Speaker 2: one data point for this is, gosh, it's been a 735 00:39:25,770 --> 00:39:27,310 Speaker 2: busy news cycle, maybe like a month and a half ago, 736 00:39:27,370 --> 00:39:30,969 Speaker 2: two months ago, we reported that DeepSeq, had placed a 737 00:39:31,010 --> 00:39:36,450 Speaker 2: very substantial order for Huawei hardware, 160,000 of Huawei's best 738 00:39:36,530 --> 00:39:40,660 Speaker 2: AI chips called 950DTs for use at a data center 739 00:39:40,680 --> 00:39:43,140 Speaker 2: in Inner Mongolia that DeepSeq is building as its sort 740 00:39:43,160 --> 00:39:47,180 Speaker 2: of like next really, really large domestic computing cluster. Incidentally, 741 00:39:47,200 --> 00:39:50,279 Speaker 2: this is the same location within China that the U.S. 742 00:39:50,300 --> 00:39:54,080 Speaker 2: government believes that DeepSeq has installed NVIDIA's Blackwell chips. We 743 00:39:54,120 --> 00:39:56,319 Speaker 2: don't know what the overall composition of compute in that 744 00:39:56,360 --> 00:39:58,580 Speaker 2: data center will be. Huawei is not going to be 745 00:39:58,620 --> 00:40:01,359 Speaker 2: able to fulfill that order for more than a year. 746 00:40:01,700 --> 00:40:05,609 Speaker 2: DeepSeq is actually lobbying the Chinese government to pressure Huawei 747 00:40:05,640 --> 00:40:08,850 Speaker 2: to speed up its allocation. It would be great for 748 00:40:08,910 --> 00:40:12,730 Speaker 2: Chinese labs if they had this alternative source of supply, right? 749 00:40:12,770 --> 00:40:15,700 Speaker 2: It's obviously the U.S. government. you know, obviously we talked 750 00:40:15,719 --> 00:40:17,680 Speaker 2: about their Monaxis piece, but, you know, everybody would like 751 00:40:17,700 --> 00:40:20,759 Speaker 2: to own their chips, or at least have them in-country. 752 00:40:21,080 --> 00:40:23,500 Speaker 2: The U.S. government has made it very difficult to use NVIDIA. 753 00:40:23,760 --> 00:40:25,839 Speaker 2: They would love people to use Huawei. It's a matter 754 00:40:25,860 --> 00:40:28,319 Speaker 2: of whether they can do a training run. And, you know, 755 00:40:28,400 --> 00:40:31,160 Speaker 2: at the time of our reporting, Huawei was planning on 756 00:40:31,280 --> 00:40:35,020 Speaker 2: using these 950DTs for inference rather than training, even though, 757 00:40:35,040 --> 00:40:37,900 Speaker 2: or sorry, DeepSea was planning on using Huawei 950DTs for 758 00:40:37,940 --> 00:40:40,239 Speaker 2: inference rather than training, even though they were designed as 759 00:40:40,280 --> 00:40:43,620 Speaker 2: training chips. And Huawei's overall production capacity of that hardware 760 00:40:43,860 --> 00:40:46,920 Speaker 2: is in the low hundreds of thousands. You know, it's 761 00:40:47,020 --> 00:40:51,280 Speaker 2: just orders of magnitude less than what NVIDIA is shipping globally. 762 00:40:51,320 --> 00:40:54,799 Speaker 2: But you still have started to see Huawei make some 763 00:40:54,900 --> 00:40:59,310 Speaker 2: export attempts. You know, I wrote about a couple of 764 00:40:59,370 --> 00:41:03,350 Speaker 2: months ago as well that they had submitted in response 765 00:41:03,410 --> 00:41:06,549 Speaker 2: to a tender from the Egyptian government to build AI 766 00:41:06,570 --> 00:41:10,520 Speaker 2: compute for use by Egyptian defense companies. And it was 767 00:41:10,560 --> 00:41:13,160 Speaker 2: a very small amount of compute. It was like 2000 ships, 768 00:41:13,219 --> 00:41:15,350 Speaker 2: but they were prepared to offer their best ships. And 769 00:41:15,390 --> 00:41:17,549 Speaker 2: it was like, when you're that supply constraint, it's like 770 00:41:17,630 --> 00:41:20,149 Speaker 2: everything that you're offering overseas is something that you're not 771 00:41:20,250 --> 00:41:23,190 Speaker 2: offering domestically. And so as a company, you know, Huawei 772 00:41:23,230 --> 00:41:25,850 Speaker 2: has to decide like, you know, as important as it 773 00:41:25,969 --> 00:41:28,390 Speaker 2: is to get more Chinese developers, you know, working on 774 00:41:28,430 --> 00:41:30,569 Speaker 2: its ecosystem, because you also get the sort of like 775 00:41:30,760 --> 00:41:33,500 Speaker 2: iterative feedback loop when you have AI companies training on 776 00:41:33,520 --> 00:41:35,779 Speaker 2: your hardware that helps you make your hardware better. They 777 00:41:35,840 --> 00:41:38,040 Speaker 2: also want to make sure that they have these footholds 778 00:41:38,080 --> 00:41:40,339 Speaker 2: in global markets. It's the same strategy that made them 779 00:41:40,380 --> 00:41:43,960 Speaker 2: sort of the preeminent telecommunications provider. And that, you know, 780 00:41:43,980 --> 00:41:46,540 Speaker 2: the U.S. government watched that happen years ago when Huawei 781 00:41:46,580 --> 00:41:49,960 Speaker 2: became extremely successful in 3G, 4G. And they want to 782 00:41:50,020 --> 00:41:52,120 Speaker 2: avoid that happening with AI infrastructure. So actually, when they 783 00:41:52,180 --> 00:41:54,980 Speaker 2: learned of this bid that Huawei had made in Egypt, 784 00:41:55,580 --> 00:41:58,040 Speaker 2: they like gathered this consortium of American companies and they 785 00:41:58,060 --> 00:42:00,819 Speaker 2: were like, counterbid. And, you know, a lot of these 786 00:42:00,840 --> 00:42:04,180 Speaker 2: companies had never really thought of Egypt as an important market. But, 787 00:42:04,620 --> 00:42:07,850 Speaker 2: you know, they had these conversations with the Egyptian government. 788 00:42:07,989 --> 00:42:10,529 Speaker 2: You know, we don't know the outcome of that tender 789 00:42:10,570 --> 00:42:13,710 Speaker 2: process yet. But it's just sort of an interesting, you know, 790 00:42:13,930 --> 00:42:16,250 Speaker 2: Egypt is an interesting battleground for Huawei because they feel 791 00:42:16,290 --> 00:42:18,810 Speaker 2: a little bit blocked or boxed out of the Persian 792 00:42:18,850 --> 00:42:21,370 Speaker 2: Gulf because there were these, you know, big American AI 793 00:42:21,390 --> 00:42:24,350 Speaker 2: infrastructure deals last year, which I talked with you guys 794 00:42:24,390 --> 00:42:26,730 Speaker 2: about about a year ago, how controversial those were. And 795 00:42:27,650 --> 00:42:28,969 Speaker 2: Egypt is sort of an anchor in the Middle East, 796 00:42:29,010 --> 00:42:31,509 Speaker 2: also an anchor in Africa. So you're starting to see 797 00:42:31,530 --> 00:42:33,680 Speaker 2: these battles play out. You know, the Malaysian government is 798 00:42:34,530 --> 00:42:38,180 Speaker 2: seriously evaluating use of an earlier generation of Huawei Ascend 799 00:42:38,200 --> 00:42:41,040 Speaker 2: chips for its sovereign AI capabilities. But, you know, it's 800 00:42:41,060 --> 00:42:43,210 Speaker 2: important to remember that the volumes that we're talking about 801 00:42:43,250 --> 00:42:45,609 Speaker 2: here are just infinitesimally small at this time. 802 00:42:46,460 --> 00:42:49,759 Speaker 5: So this is very much a live story, as you've 803 00:42:49,780 --> 00:42:52,719 Speaker 5: been emphasizing throughout this conversation, and one reason why you're 804 00:42:52,760 --> 00:42:54,930 Speaker 5: very sleep deprived at the moment. But what are you 805 00:42:54,969 --> 00:42:58,149 Speaker 5: watching out for next? So we still we don't really 806 00:42:58,170 --> 00:43:00,089 Speaker 5: know what the DOJ is going to do. And now 807 00:43:00,130 --> 00:43:05,290 Speaker 5: we have this independent investigation going into Jensen and the 808 00:43:05,350 --> 00:43:09,200 Speaker 5: spy affair at NVIDIA. What do you think is going 809 00:43:09,219 --> 00:43:09,700 Speaker 5: to be important? 810 00:43:11,410 --> 00:43:14,330 Speaker 2: Yeah, I mean, look, obviously Jensen's fate is sort of 811 00:43:14,570 --> 00:43:18,590 Speaker 2: important for global markets, for the industry, you know, what 812 00:43:18,640 --> 00:43:23,460 Speaker 2: the board ends up deciding to do. I think from 813 00:43:23,800 --> 00:43:27,140 Speaker 2: a policy and regulatory perspective, you know, this is a 814 00:43:27,560 --> 00:43:33,550 Speaker 2: very unique moment in an administration that, generally speaking, has, 815 00:43:33,830 --> 00:43:36,730 Speaker 2: you know, adopted a lot of the policies that NVIDIA 816 00:43:36,750 --> 00:43:41,740 Speaker 2: would like it to. There is... rage, honestly, on Capitol 817 00:43:41,780 --> 00:43:44,180 Speaker 2: Hill and in many parts of the Trump administration about 818 00:43:44,219 --> 00:43:48,080 Speaker 2: the overall state of export controls. Obviously, this scandal with 819 00:43:48,140 --> 00:43:50,640 Speaker 2: NVIDIA gives people a lot of things to latch on to. 820 00:43:50,960 --> 00:43:53,589 Speaker 2: So you're starting to see these voices that have not 821 00:43:53,640 --> 00:43:58,290 Speaker 2: historically been particularly active on this issue take great interest 822 00:43:58,350 --> 00:44:00,430 Speaker 2: in it. And then you have this sort of baseline 823 00:44:00,469 --> 00:44:04,250 Speaker 2: contingent of China hawks who have been upset by what 824 00:44:04,290 --> 00:44:06,850 Speaker 2: they see as a lack of action by the Bureau 825 00:44:07,010 --> 00:44:10,690 Speaker 2: of Industry and Security, that BIS, within the Trump administration 826 00:44:10,870 --> 00:44:13,150 Speaker 2: on this topic and are trying to legislate some of 827 00:44:13,210 --> 00:44:15,109 Speaker 2: the rules to at least like, OK, let's not roll 828 00:44:15,150 --> 00:44:17,629 Speaker 2: back export controls more than they already have been. Let's 829 00:44:17,650 --> 00:44:22,529 Speaker 2: put additional restrictions on semiconductor equipment shipments into China. Let's 830 00:44:22,570 --> 00:44:25,330 Speaker 2: make sure that it's unambiguous that the executive branch has 831 00:44:25,370 --> 00:44:29,420 Speaker 2: the authority to restrict cloud compute access by Chinese companies. 832 00:44:29,940 --> 00:44:33,219 Speaker 2: And a lot of these measures are sort of going 833 00:44:33,239 --> 00:44:34,880 Speaker 2: to be litigated over the next couple of months as 834 00:44:34,920 --> 00:44:37,739 Speaker 2: part of an annual must pass defense package. So it'll 835 00:44:37,760 --> 00:44:40,529 Speaker 2: be a really interesting couple of months on Capitol Hill. Obviously, 836 00:44:40,550 --> 00:44:44,469 Speaker 2: you know, Democrats are taking Congress next year. I would 837 00:44:44,510 --> 00:44:48,089 Speaker 2: expect a significant amount of oversight into not just NVIDIA, but, 838 00:44:48,130 --> 00:44:51,270 Speaker 2: you know, the major players in the ecosystem. And then 839 00:44:51,310 --> 00:44:54,529 Speaker 2: I think, you know, for Southeast Asia, which is a 840 00:44:54,530 --> 00:44:57,049 Speaker 2: bit closer to where I sit, you know, there's been 841 00:44:57,090 --> 00:45:01,260 Speaker 2: this sort of whiplash over the past couple of years where, 842 00:45:01,360 --> 00:45:03,419 Speaker 2: you know, remember like the AI diffusion rule I mentioned 843 00:45:03,440 --> 00:45:06,380 Speaker 2: that the Trump administration is not enforcing? So this was 844 00:45:06,400 --> 00:45:09,480 Speaker 2: going to impose, you know, licensing requirements, meaning you need 845 00:45:09,500 --> 00:45:12,440 Speaker 2: to get the U.S. government's permission to ship NVIDIA ships 846 00:45:12,620 --> 00:45:16,739 Speaker 2: to Singapore, Malaysia, Thailand, Indonesia, all these places where, you know, 847 00:45:16,760 --> 00:45:19,560 Speaker 2: we have these smuggling cases happening. And so actually some 848 00:45:19,580 --> 00:45:23,020 Speaker 2: of the companies that are suspected of being involved in 849 00:45:23,040 --> 00:45:26,140 Speaker 2: smuggling like massively ramped up their shipments right before those 850 00:45:26,200 --> 00:45:27,700 Speaker 2: restrictions were set to take effect. And then they didn't 851 00:45:27,719 --> 00:45:32,040 Speaker 2: take effect. And then the Trump administration worked on a regulation. 852 00:45:32,060 --> 00:45:33,759 Speaker 2: They were like, hey, we're not doing that, but we 853 00:45:33,820 --> 00:45:36,279 Speaker 2: are going to restrict to Malaysia and Thailand specifically. This 854 00:45:36,320 --> 00:45:38,800 Speaker 2: was a draft rule that I read about in June 855 00:45:38,960 --> 00:45:41,879 Speaker 2: or July of last year. And then they scrapped that 856 00:45:41,940 --> 00:45:45,410 Speaker 2: amid diplomatic pressure from the Malaysian government. And then they 857 00:45:45,430 --> 00:45:48,110 Speaker 2: worked on another replacement regulation that would have done global 858 00:45:48,230 --> 00:45:52,050 Speaker 2: export controls. We heard about that in March, I would say. 859 00:45:52,070 --> 00:45:54,770 Speaker 2: And that was withdrawn shortly after it was sent to 860 00:45:54,850 --> 00:45:58,660 Speaker 2: other U.S. government agencies. And so, you know, for companies 861 00:45:58,680 --> 00:46:01,700 Speaker 2: that are trying to build infrastructure in this region, you know, 862 00:46:02,400 --> 00:46:04,339 Speaker 2: even the ones that are doing it completely legally, it's like, 863 00:46:04,719 --> 00:46:06,219 Speaker 2: are we going to be able to get the chips? 864 00:46:06,300 --> 00:46:08,359 Speaker 2: Like it's this constant sort of, you know, the U.S. 865 00:46:08,390 --> 00:46:09,750 Speaker 2: government's a little bit the boy who cried wolf on 866 00:46:09,770 --> 00:46:13,489 Speaker 2: this issue. You know, we have this intense scrutiny into 867 00:46:13,530 --> 00:46:17,969 Speaker 2: mega speed. Asian banks sort of go through like ebbs 868 00:46:18,030 --> 00:46:20,779 Speaker 2: and flows of how comfortable they are with, you you know, 869 00:46:20,860 --> 00:46:24,500 Speaker 2: financing deals that have a significant amount of Chinese offtakes? 870 00:46:24,540 --> 00:46:26,400 Speaker 2: Or like, what if the U.S. government does restrict it 871 00:46:26,440 --> 00:46:28,299 Speaker 2: and then are suddenly, you know, left holding the bag? 872 00:46:28,860 --> 00:46:31,070 Speaker 2: And this is, you know, one of the fastest growing, 873 00:46:31,370 --> 00:46:34,029 Speaker 2: you know, AI infrastructure regions in the world. You know, 874 00:46:34,050 --> 00:46:36,270 Speaker 2: the amount of compute online right now is like, it 875 00:46:37,230 --> 00:46:39,490 Speaker 2: doesn't hold a candle to what exists in the U.S., 876 00:46:39,510 --> 00:46:41,549 Speaker 2: but it is fast growing. A lot of companies are 877 00:46:41,570 --> 00:46:44,359 Speaker 2: investing here and it's just under, yeah, an intense period 878 00:46:44,380 --> 00:46:47,239 Speaker 2: of scrutiny. So it's, yes, that's the developments that I'll 879 00:46:47,280 --> 00:46:50,160 Speaker 2: be watching. And then on the Huawei front, you know, 880 00:46:50,750 --> 00:46:53,410 Speaker 2: Huawei is making export attempts. And I think, you know, 881 00:46:53,430 --> 00:46:55,989 Speaker 2: the U.S. government, it's funny, like over a year ago, 882 00:46:56,030 --> 00:46:58,640 Speaker 2: there was an executive order to promote the export of 883 00:46:58,680 --> 00:47:01,840 Speaker 2: the American AI stack to basically like design these packages 884 00:47:01,900 --> 00:47:05,149 Speaker 2: that you could sort of offer to you know, other 885 00:47:05,210 --> 00:47:08,069 Speaker 2: countries of like adopt, you know, up and down, you know, 886 00:47:08,110 --> 00:47:11,640 Speaker 2: from applications to models to infrastructure, the American AI stack. 887 00:47:12,180 --> 00:47:14,460 Speaker 2: And we haven't seen like anything come out of that 888 00:47:14,500 --> 00:47:17,460 Speaker 2: program yet. And now we actually do have Huawei like 889 00:47:17,500 --> 00:47:19,980 Speaker 2: starting to make these attempts. And the U.S. is sort 890 00:47:20,000 --> 00:47:22,320 Speaker 2: of responding to them on like an ad hoc basis. 891 00:47:23,030 --> 00:47:24,549 Speaker 2: And I'm sure that we'll see more of that over 892 00:47:24,570 --> 00:47:26,649 Speaker 2: the next year. So you'll start to see like these 893 00:47:26,730 --> 00:47:29,250 Speaker 2: battlegrounds in the global south that will be really interesting 894 00:47:29,310 --> 00:47:32,319 Speaker 2: to watch, you know, how foreign governments are make decisions 895 00:47:32,390 --> 00:47:34,870 Speaker 2: about their own sovereign capabilities and whether they think, you know, 896 00:47:34,890 --> 00:47:37,390 Speaker 2: the U.S. is a reliable partner or they're willing to take, 897 00:47:37,430 --> 00:47:39,890 Speaker 2: you know, maybe worse trips from China that, you know, 898 00:47:39,910 --> 00:47:41,950 Speaker 2: Huawei is happy to like send their engineers there and 899 00:47:41,969 --> 00:47:44,029 Speaker 2: do the whole data center for you. And yeah, I 900 00:47:44,050 --> 00:47:45,250 Speaker 2: know it's exciting times. 901 00:47:47,050 --> 00:47:48,190 Speaker 1: Mackenzie, that was fantastic. 902 00:47:48,210 --> 00:47:51,589 Speaker 4: Get some sleep. Well-deserved. You're going to need it. But no, 903 00:47:51,610 --> 00:47:52,110 Speaker 4: that was great. 904 00:47:52,160 --> 00:47:54,279 Speaker 3: Thank you for staying up late, catching us up to 905 00:47:54,320 --> 00:47:57,120 Speaker 3: speed on all this. And I don't think this will 906 00:47:57,140 --> 00:47:58,140 Speaker 3: be the last time we chat. 907 00:47:58,940 --> 00:48:00,339 Speaker 4: And I appreciate your coming on the hotline. 908 00:48:00,630 --> 00:48:14,990 Speaker 6: Yeah, thanks so much for having me. Always a pleasure, guys. 909 00:48:15,430 --> 00:48:21,160 Speaker 4: She's so great. These are such complicated stories. 910 00:48:21,219 --> 00:48:21,920 Speaker 5: They really are. 911 00:48:21,980 --> 00:48:26,299 Speaker 3: They're really complicated. There's so many different... dimensions to them. 912 00:48:26,580 --> 00:48:28,779 Speaker 5: Also, there's a lot of legal technicality in what can 913 00:48:28,820 --> 00:48:31,879 Speaker 5: and can't be said, which I think sometimes people don't 914 00:48:31,920 --> 00:48:35,620 Speaker 5: necessarily appreciate from a journalistic perspective. And she's doing that 915 00:48:35,840 --> 00:48:39,020 Speaker 5: all on like, I think she said she hadn't slept 916 00:48:39,040 --> 00:48:42,299 Speaker 5: for 70 hours or something. So that's pretty intense. No, 917 00:48:42,340 --> 00:48:45,740 Speaker 5: it's a fascinating story. I keep thinking it feels like 918 00:48:46,239 --> 00:48:48,620 Speaker 5: a big tail risk right now for the chip makers. 919 00:48:49,120 --> 00:48:53,050 Speaker 5: I mean, obviously what happens with Jensen and the investigation, 920 00:48:53,170 --> 00:48:56,649 Speaker 5: but just the idea that as a lot of this 921 00:48:56,930 --> 00:49:00,600 Speaker 5: becomes more public knowledge, as politicians become aware of it, 922 00:49:00,719 --> 00:49:05,460 Speaker 5: that that legal loophole of Chinese companies being able to 923 00:49:05,620 --> 00:49:11,260 Speaker 5: rent GPUs via Southeast Asia, even if they can't actually 924 00:49:11,340 --> 00:49:14,930 Speaker 5: buy certain chips, that comes under scrutiny. And you can 925 00:49:14,969 --> 00:49:18,620 Speaker 5: imagine if that... like it's such a weird one it's. 926 00:49:18,410 --> 00:49:22,500 Speaker 3: So weird this term remote access what does that even 927 00:49:22,580 --> 00:49:25,020 Speaker 3: mean it's a cloud of course it's remote access it's 928 00:49:25,060 --> 00:49:28,279 Speaker 3: like oh as opposed to like the tenant of the 929 00:49:28,380 --> 00:49:31,020 Speaker 3: user of the chips walking into the data instead of 930 00:49:31,040 --> 00:49:33,840 Speaker 3: themselves are plugging in a laptop it's all remote access 931 00:49:34,060 --> 00:49:37,589 Speaker 3: this so it's a it's always been very baffling to 932 00:49:37,620 --> 00:49:40,489 Speaker 3: me that we have these intense conversations about, oh, you 933 00:49:40,550 --> 00:49:41,710 Speaker 3: can't sell into China. 934 00:49:42,010 --> 00:49:43,870 Speaker 4: But actually, it's really not. 935 00:49:43,670 --> 00:49:47,629 Speaker 3: That big of a deal to sell into, and it's 936 00:49:47,690 --> 00:49:53,210 Speaker 3: not illegal currently, to sell into anywhere else who then 937 00:49:53,530 --> 00:49:55,830 Speaker 3: rents that capacity to Chinese firms. 938 00:49:55,910 --> 00:50:01,540 Speaker 4: Like, what? I don't know. I do, like, Who can 939 00:50:01,600 --> 00:50:02,280 Speaker 4: explain this. 940 00:50:02,200 --> 00:50:03,590 Speaker 1: In just a moment? 941 00:50:03,730 --> 00:50:05,770 Speaker 5: Yeah. The other thing that I thought was really interesting 942 00:50:05,870 --> 00:50:08,089 Speaker 5: and I wasn't aware of was McKenzie saying that there's 943 00:50:08,130 --> 00:50:12,790 Speaker 5: basically like a handful of U.S. Customs officers in Southeast Asia. 944 00:50:12,870 --> 00:50:15,160 Speaker 5: And you can imagine like if you're going to ramp 945 00:50:15,260 --> 00:50:19,120 Speaker 5: up export controls and expect them to be effective, you're 946 00:50:19,140 --> 00:50:22,399 Speaker 5: going to have to staff up, right? It reminds me, 947 00:50:22,420 --> 00:50:26,160 Speaker 5: I did an article on the corporate bond market many, 948 00:50:26,200 --> 00:50:28,730 Speaker 5: many years ago. And at the time, I think this 949 00:50:28,750 --> 00:50:31,780 Speaker 5: was like... well, I can't remember, maybe an eight trillion market. 950 00:50:32,260 --> 00:50:35,960 Speaker 5: And there was one guy at the SEC who devoted 951 00:50:36,180 --> 00:50:41,140 Speaker 5: half his time to actually monitoring the corporate bond market, 952 00:50:41,180 --> 00:50:42,220 Speaker 5: which blew my mind. 953 00:50:42,739 --> 00:50:46,419 Speaker 3: You know, I really think actually, and I hadn't appreciated 954 00:50:46,480 --> 00:50:51,430 Speaker 3: this until McKenzie brought up Huawei's sale to Egypt. That is, 955 00:50:51,550 --> 00:50:54,130 Speaker 3: I forget, you know, what is it? We did the 956 00:50:54,170 --> 00:50:56,790 Speaker 3: story with Eva Doe, who wrote the book, House of Huawei. 957 00:50:56,830 --> 00:51:00,010 Speaker 3: And what was the story there? The story there was 958 00:51:00,510 --> 00:51:04,560 Speaker 3: this is a company that is very aggressive about gaining 959 00:51:04,620 --> 00:51:08,260 Speaker 3: footholds in parts of the world that, like, other companies 960 00:51:08,320 --> 00:51:09,420 Speaker 3: aren't as aggressive about. 961 00:51:09,460 --> 00:51:12,259 Speaker 4: So Egypt is, like, as she explained, a pretty good 962 00:51:12,420 --> 00:51:13,020 Speaker 4: example of this. 963 00:51:13,460 --> 00:51:17,400 Speaker 3: Even with a scarce number of Ascend chips, that is 964 00:51:17,420 --> 00:51:20,540 Speaker 3: sort of, like, in their corporate DNA, like, go out and, like, 965 00:51:20,600 --> 00:51:24,469 Speaker 3: win these global markets, which then, of course, implicitly, as 966 00:51:24,510 --> 00:51:28,040 Speaker 3: you gain this foothold, then of course implicitly help sort 967 00:51:28,060 --> 00:51:31,020 Speaker 3: of like broader Chinese, like soft and hard power. 968 00:51:31,060 --> 00:51:31,379 Speaker 6: Et cetera. 969 00:51:31,880 --> 00:51:33,980 Speaker 3: But then I was thinking about, okay, let's say after 970 00:51:34,020 --> 00:51:36,880 Speaker 3: this mega speed scandal, you're like, you know what? We 971 00:51:36,920 --> 00:51:39,620 Speaker 3: really are no more of this remote stuff. We really 972 00:51:39,660 --> 00:51:41,680 Speaker 3: got to like crack down on who we're selling to 973 00:51:41,980 --> 00:51:47,160 Speaker 3: in Southeast Asia. It's clear that today, you know, it 974 00:51:47,200 --> 00:51:51,190 Speaker 3: doesn't sound like Huawei could replace Nvidia today, but if, 975 00:51:51,480 --> 00:51:54,860 Speaker 3: you know, maybe a few years from now, like a 976 00:51:55,010 --> 00:51:58,150 Speaker 3: number of these countries do become like Huawei countries. 977 00:51:58,230 --> 00:51:58,590 Speaker 2: Yeah. 978 00:51:58,870 --> 00:52:00,170 Speaker 4: It's going to be very tricky. 979 00:52:00,489 --> 00:52:04,830 Speaker 5: Well, the other thing I just want to reiterate is like, again, 980 00:52:05,310 --> 00:52:10,050 Speaker 5: with NVIDIA itself, there's a weird like contradiction here where 981 00:52:10,310 --> 00:52:14,669 Speaker 5: if chip demand is as strong as they describe, they 982 00:52:14,710 --> 00:52:17,450 Speaker 5: should be very selective with their customers, right? 983 00:52:17,469 --> 00:52:18,030 Speaker 2: Yes, yes. 984 00:52:18,050 --> 00:52:21,900 Speaker 5: Like there shouldn't be any question marks over anyone, basically. 985 00:52:21,940 --> 00:52:22,580 Speaker 2: Yeah, yeah. 986 00:52:23,250 --> 00:52:26,400 Speaker 5: And so I find that really interesting as well. And 987 00:52:26,480 --> 00:52:29,680 Speaker 5: probably like an under-discussed angle in all of this. Like 988 00:52:29,700 --> 00:52:32,799 Speaker 5: you can afford to like- Be discriminating. Be discriminating and 989 00:52:32,840 --> 00:52:36,400 Speaker 5: have like the most intense compliance process in the world 990 00:52:36,460 --> 00:52:39,170 Speaker 5: if you want. Anyway, shall we leave it there? 991 00:52:39,190 --> 00:52:40,390 Speaker 4: Let's leave it there. 992 00:52:40,630 --> 00:52:43,009 Speaker 5: This has been another episode of the All Thoughts Podcast. 993 00:52:43,150 --> 00:52:45,890 Speaker 5: I'm Tracy Allaway. You can follow me at Tracy Allaway. 994 00:52:46,090 --> 00:52:49,430 Speaker 3: And I'm Joe Wiesenthal. You can follow me at The Stalwart. 995 00:52:49,690 --> 00:52:51,419 Speaker 3: Follow our guest Mackenzie Hawkins. 996 00:52:51,560 --> 00:52:53,000 Speaker 4: She's at Mack Hawk. 997 00:52:53,290 --> 00:52:57,770 Speaker 3: Follow our producers, Carmen Rodriguez at CarmenArmond, Dashiell Bennett at Dashbot, 998 00:52:58,130 --> 00:53:01,010 Speaker 3: Kale Brooks at Kale Brooks, and Kevin Lozano at Kevin 999 00:53:01,030 --> 00:53:03,230 Speaker 3: Lloyd Lozano. And you can chat about all of these 1000 00:53:03,270 --> 00:53:08,380 Speaker 3: topics 24-7 in our Discord, discord.gg slash oddlots. And for 1001 00:53:08,390 --> 00:53:11,040 Speaker 3: more Oddlots content, check out our daily newsletter, which you 1002 00:53:11,060 --> 00:53:13,200 Speaker 3: can find at bloomberg.com slash oddlots. 1003 00:53:13,739 --> 00:53:16,060 Speaker 5: And if you enjoyed this conversation, if you like it 1004 00:53:16,219 --> 00:53:18,660 Speaker 5: when we talk about chip smuggling, then please leave us 1005 00:53:18,680 --> 00:53:21,730 Speaker 5: a positive review on your favorite podcast platform. And if 1006 00:53:21,750 --> 00:53:24,509 Speaker 5: you're watching on video, then please like or leave a 1007 00:53:24,530 --> 00:53:26,270 Speaker 5: comment or better yet, subscribe. 1008 00:53:26,469 --> 00:53:28,910 Speaker 3: And remember, if you're a Bloomberg subscriber, you can listen 1009 00:53:28,930 --> 00:53:31,830 Speaker 3: to all of our episodes absolutely ad-free. All you need 1010 00:53:31,850 --> 00:53:35,410 Speaker 3: to do is find the Bloomberg channel on Apple Podcasts 1011 00:53:35,489 --> 00:53:36,040 Speaker 3: and follow. 1012 00:53:35,880 --> 00:53:36,660 Speaker 4: The instructions there. 1013 00:53:37,200 --> 00:53:37,930 Speaker 5: Thanks for joining. 1014 00:53:54,460 --> 00:53:54,890 Speaker 3: Thank you.