1 00:00:02,160 --> 00:00:06,000 Speaker 1: Also media, Hello and welcome to Better Offline. 2 00:00:06,000 --> 00:00:20,600 Speaker 2: I'm your host ed ZiT Tron. And of course you 3 00:00:20,600 --> 00:00:23,400 Speaker 2: can go into the notes for the episode. You can 4 00:00:23,440 --> 00:00:25,280 Speaker 2: take a look at links from a newsletter. You can 5 00:00:25,320 --> 00:00:28,160 Speaker 2: go and buy some merch as always, but much more importantly, 6 00:00:28,160 --> 00:00:31,200 Speaker 2: today I'm joined by the esteem Steve Burke of Gamers Nexus. Steve, 7 00:00:31,200 --> 00:00:32,080 Speaker 2: thank you for joining. 8 00:00:31,920 --> 00:00:33,199 Speaker 3: Us, Thanks for having me. 9 00:00:33,960 --> 00:00:36,960 Speaker 2: So you did an incredible video on the black market 10 00:00:37,080 --> 00:00:40,559 Speaker 2: smuggling of in videos AI GPUs into China, and then 11 00:00:40,600 --> 00:00:43,160 Speaker 2: there has been a thing that's happened where the video 12 00:00:43,200 --> 00:00:46,840 Speaker 2: has been pulled down due to a DMCA complaint from Bloomberg. 13 00:00:46,920 --> 00:00:47,760 Speaker 2: What is happening? 14 00:00:50,320 --> 00:00:54,960 Speaker 3: Yeah, it's so the video, it's really cool video where 15 00:00:55,000 --> 00:00:58,200 Speaker 3: we went all over China finding these GPUs. As part 16 00:00:58,240 --> 00:01:02,480 Speaker 3: of that, we included a recap of all the news 17 00:01:02,480 --> 00:01:05,880 Speaker 3: of export controls or the US controls GPUs that are 18 00:01:06,840 --> 00:01:12,440 Speaker 3: above a certain performance threshold and the allowance to export 19 00:01:12,480 --> 00:01:15,160 Speaker 3: those to China by American companies, and so that was 20 00:01:15,200 --> 00:01:17,240 Speaker 3: the story we wanted to look into. Okay, but they're 21 00:01:17,240 --> 00:01:20,320 Speaker 3: still getting to China and so how is that? And 22 00:01:20,959 --> 00:01:23,479 Speaker 3: as part of that recap, go through all the news 23 00:01:23,480 --> 00:01:26,600 Speaker 3: we've got clips of you know, various politicians. There was 24 00:01:26,600 --> 00:01:29,039 Speaker 3: a clip of Obama in there, and there was a 25 00:01:29,040 --> 00:01:31,800 Speaker 3: clip of Trump in there, a couple of them talking 26 00:01:31,840 --> 00:01:35,400 Speaker 3: about AI or GPU export control bands. Of course AI, 27 00:01:35,520 --> 00:01:41,400 Speaker 3: I know ed is one of your favorite topics. Yes, yes, 28 00:01:41,920 --> 00:01:48,080 Speaker 3: me as well. And yeah, so we received a copyright 29 00:01:48,080 --> 00:01:56,400 Speaker 3: strike from small, small publication called the Bloomberg. Yeah. So 30 00:01:56,480 --> 00:02:01,080 Speaker 3: that strike took the video offline and now we are 31 00:02:01,120 --> 00:02:04,600 Speaker 3: currently challenging it. And so the strike was for specifically, 32 00:02:04,600 --> 00:02:07,680 Speaker 3: I believe it was minutes twenty two until minute twenty 33 00:02:07,720 --> 00:02:09,400 Speaker 3: three point fifteen in the video. 34 00:02:10,000 --> 00:02:14,200 Speaker 2: Of Donald Trump speaking the President of America. Correct, Yes, 35 00:02:14,720 --> 00:02:18,120 Speaker 2: that is ludicrous. And you went to New York specifically 36 00:02:18,160 --> 00:02:20,440 Speaker 2: to have it out with them, did they Where do 37 00:02:20,480 --> 00:02:22,200 Speaker 2: we stand as we speak, because this is going out 38 00:02:22,320 --> 00:02:24,480 Speaker 2: in about a week from now. It's August twenty eighth, 39 00:02:24,520 --> 00:02:27,080 Speaker 2: so who knows what twenty sixth even, I don't know 40 00:02:27,120 --> 00:02:28,239 Speaker 2: what happened next. 41 00:02:28,480 --> 00:02:30,640 Speaker 3: Yeah, if it goes out in about a week, then 42 00:02:30,880 --> 00:02:32,720 Speaker 3: that'll be pretty close to when we should have an 43 00:02:32,800 --> 00:02:34,600 Speaker 3: update from YouTube. 44 00:02:35,360 --> 00:02:36,960 Speaker 2: Because the video is off right now. 45 00:02:37,400 --> 00:02:41,880 Speaker 3: Correct, the video's gone. And it's very interesting how copyright 46 00:02:41,960 --> 00:02:45,720 Speaker 3: strikes work because if basically the way it works is 47 00:02:45,760 --> 00:02:50,760 Speaker 3: a company files a strike. They decide if there's infringement 48 00:02:50,800 --> 00:02:55,720 Speaker 3: of some kind, and they file it. YouTube then basically, 49 00:02:55,840 --> 00:02:59,919 Speaker 3: especially for comedy, Bloomberg size will automatically approve it. It's then 50 00:03:00,120 --> 00:03:03,080 Speaker 3: up to the creator to dispute that the videos offline. 51 00:03:03,120 --> 00:03:07,040 Speaker 3: During this entire process, any money from the video is 52 00:03:07,080 --> 00:03:11,560 Speaker 3: held at least in escrow, and so then you can 53 00:03:11,600 --> 00:03:14,480 Speaker 3: dispute it if you want. At that stage, YouTube then 54 00:03:14,520 --> 00:03:16,960 Speaker 3: decides if it's going to accept your dispute. It accepted. 55 00:03:16,960 --> 00:03:19,040 Speaker 3: Ours is written by a lawyer, so that makes sense, 56 00:03:19,919 --> 00:03:22,160 Speaker 3: and then they step back and they're like, Okay, we're 57 00:03:22,200 --> 00:03:26,440 Speaker 3: not involved anymore. It's between you guys. And so the 58 00:03:26,600 --> 00:03:30,679 Speaker 3: party that filed the claimant, so Bloomberg here, has ten 59 00:03:30,760 --> 00:03:35,680 Speaker 3: business days to file proof of lawsuits. So they if 60 00:03:35,720 --> 00:03:38,560 Speaker 3: they submit to YouTube proof that they are suing us, 61 00:03:38,560 --> 00:03:40,560 Speaker 3: which thus far as we record this, they have not, 62 00:03:42,400 --> 00:03:45,760 Speaker 3: then at that stage YouTube would keep the video offline 63 00:03:45,920 --> 00:03:47,640 Speaker 3: and they kind of let the two parties resolve that 64 00:03:47,680 --> 00:03:50,920 Speaker 3: amongst themselves. If they don't submit any proof of a 65 00:03:51,000 --> 00:03:53,440 Speaker 3: lawsuit or they don't file a lawsuit and then show 66 00:03:53,480 --> 00:03:56,760 Speaker 3: YouTube that they've done so, again, they haven't at this point. 67 00:03:56,880 --> 00:04:00,000 Speaker 3: Then basically in ten business days the video goes back up. 68 00:04:01,000 --> 00:04:03,320 Speaker 3: We get fucked over for the ten days in between, 69 00:04:03,800 --> 00:04:08,360 Speaker 3: and then the money, in theory should be released back 70 00:04:08,360 --> 00:04:10,200 Speaker 3: to us, the ad revenue that was made from it, 71 00:04:10,600 --> 00:04:14,200 Speaker 3: if it follows that course. But the biggest damage that's 72 00:04:14,240 --> 00:04:17,480 Speaker 3: done is from the loss of momentum in those you know, 73 00:04:17,720 --> 00:04:21,440 Speaker 3: total time period maybe thirteen days or so. But why would. 74 00:04:21,200 --> 00:04:24,880 Speaker 2: Bloomberg do this? Are they really that attached to videos 75 00:04:24,880 --> 00:04:26,840 Speaker 2: of Donald Trump? Or is it something else? 76 00:04:27,720 --> 00:04:31,640 Speaker 3: I mean, I yeah, we talked about this in our 77 00:04:31,720 --> 00:04:34,239 Speaker 3: video where we flew out to try and speak with them. 78 00:04:36,360 --> 00:04:40,760 Speaker 3: I have a number of theories. We don't know their 79 00:04:40,839 --> 00:04:44,479 Speaker 3: motive precisely, but you know, there's a few strange things 80 00:04:44,520 --> 00:04:48,599 Speaker 3: going on where first of all, I've noticed that they 81 00:04:48,640 --> 00:04:52,960 Speaker 3: have not claimed any or struck any that I'm aware 82 00:04:53,000 --> 00:04:56,480 Speaker 3: of the re uploads of the video. So that's interesting 83 00:04:56,520 --> 00:04:59,400 Speaker 3: because there's hundreds of them. It was just hours that 84 00:04:59,400 --> 00:05:03,839 Speaker 3: want top initially. Secondly, we have used that clip before 85 00:05:04,279 --> 00:05:08,839 Speaker 3: and that hasn't been stricken and so specifically in this 86 00:05:08,920 --> 00:05:14,440 Speaker 3: black market video. And yeah, the theories we came up 87 00:05:14,440 --> 00:05:17,880 Speaker 3: with were there's some competing content from Bloomberg where they 88 00:05:17,920 --> 00:05:22,120 Speaker 3: attempted to do what we did. It was really a 89 00:05:22,160 --> 00:05:26,440 Speaker 3: failed attempt. They tried to find smuggled ai GPUs in 90 00:05:26,520 --> 00:05:28,359 Speaker 3: China as far as they kind of presented it in 91 00:05:28,440 --> 00:05:30,719 Speaker 3: data centers, they were unable to do so they couldn't 92 00:05:30,720 --> 00:05:33,480 Speaker 3: get any access. They filmed sand in the desert. They 93 00:05:33,520 --> 00:05:39,280 Speaker 3: go home. There's also that's all they go. Yeah, they 94 00:05:39,400 --> 00:05:42,000 Speaker 3: drive around at a desert, they point the camera at 95 00:05:42,000 --> 00:05:45,000 Speaker 3: some buildings that are being constructed. They say, you know, 96 00:05:45,760 --> 00:05:49,080 Speaker 3: there's ai GPUs in them their hills, and then that's 97 00:05:49,120 --> 00:05:51,240 Speaker 3: about it. 98 00:05:51,520 --> 00:05:54,600 Speaker 2: It's kind of funny that you you appear to on 99 00:05:54,640 --> 00:05:57,760 Speaker 2: your own and with your with your team obviously able 100 00:05:57,800 --> 00:06:00,960 Speaker 2: to get this access to these large legacy media people don't, 101 00:06:01,000 --> 00:06:03,440 Speaker 2: which it's meant to be the other way around based 102 00:06:03,480 --> 00:06:04,479 Speaker 2: on what everyone says. 103 00:06:05,040 --> 00:06:08,720 Speaker 4: So what is it the makes your approach. 104 00:06:08,440 --> 00:06:11,400 Speaker 2: Different to bloombugs, Like why did you get stuff that 105 00:06:11,440 --> 00:06:12,919 Speaker 2: they didn't? 106 00:06:13,040 --> 00:06:17,000 Speaker 3: I think rather than just the desert, right, Yeah, Well, 107 00:06:17,080 --> 00:06:20,039 Speaker 3: to be fair, we did not film the desert, so 108 00:06:20,040 --> 00:06:25,040 Speaker 3: they did have an exclusive there. But I think it 109 00:06:25,040 --> 00:06:27,919 Speaker 3: comes down to the approach where we looked at it 110 00:06:28,000 --> 00:06:30,679 Speaker 3: like this is an interesting concept, this idea of black 111 00:06:30,720 --> 00:06:36,000 Speaker 3: market GPUs, But from my point of view, we approached 112 00:06:36,000 --> 00:06:38,600 Speaker 3: it as well, let's just see what are the people 113 00:06:38,640 --> 00:06:43,560 Speaker 3: involved in this chain think and do they think it's 114 00:06:43,680 --> 00:06:45,960 Speaker 3: a black market or do they just think it's a market. 115 00:06:47,720 --> 00:06:50,320 Speaker 3: And a lot of the people we worked with were just, 116 00:06:51,040 --> 00:06:53,760 Speaker 3: I think, amused at the concept of being involved in 117 00:06:53,800 --> 00:06:56,760 Speaker 3: the media process and wanted to see how it works. 118 00:06:57,560 --> 00:07:00,080 Speaker 3: And so I honestly, I think a lot of that access, 119 00:07:00,320 --> 00:07:02,159 Speaker 3: especially to the people who are just sort of like 120 00:07:02,320 --> 00:07:05,960 Speaker 3: normal people. They're not these big, you know, executives at 121 00:07:06,000 --> 00:07:10,040 Speaker 3: some huge corporation. They're they're middleman and they're people transacting 122 00:07:10,080 --> 00:07:14,680 Speaker 3: GPUs on the ground. I think for them, if you 123 00:07:14,760 --> 00:07:19,360 Speaker 3: go into it without this free supposition of some kind 124 00:07:19,400 --> 00:07:26,280 Speaker 3: of particular i don't know, leaning or ethics of whether 125 00:07:26,640 --> 00:07:28,960 Speaker 3: what they're doing is or isn't okay or whatever, or. 126 00:07:29,000 --> 00:07:31,880 Speaker 2: Even expecting what you'll find exactly. 127 00:07:32,320 --> 00:07:34,760 Speaker 3: Yeah, I think if you go in with an open 128 00:07:34,760 --> 00:07:36,360 Speaker 3: mind of like we're just gonna hang out with this 129 00:07:36,400 --> 00:07:39,960 Speaker 3: guy for a day and just observe his job, people 130 00:07:39,960 --> 00:07:42,120 Speaker 3: are you know, they're pretty happy to share that. 131 00:07:42,360 --> 00:07:46,680 Speaker 2: So, yeah, how did you actually start those? So you 132 00:07:46,960 --> 00:07:48,480 Speaker 2: you came up with the idea that you wanted to 133 00:07:48,520 --> 00:07:52,000 Speaker 2: see how elicit GPUs making it into China. Did you 134 00:07:52,360 --> 00:07:54,800 Speaker 2: have contacts in China or Hong Kong and so on 135 00:07:54,960 --> 00:07:57,760 Speaker 2: that you talk to in advance and they kind of say, oh, 136 00:07:57,800 --> 00:07:59,480 Speaker 2: if you come here, we can show you this. What 137 00:08:00,280 --> 00:08:01,560 Speaker 2: was the process? 138 00:08:02,120 --> 00:08:06,040 Speaker 3: Yeah, the process was really cool. We're still kind of 139 00:08:06,080 --> 00:08:10,400 Speaker 3: learning how to iterate on this, but each time we're 140 00:08:11,160 --> 00:08:14,720 Speaker 3: The big thing we learned this time for this investigation 141 00:08:15,840 --> 00:08:20,800 Speaker 3: was really just trying to that first contact matters a lot, 142 00:08:21,080 --> 00:08:22,920 Speaker 3: and being able to work with that first person we 143 00:08:22,960 --> 00:08:25,640 Speaker 3: find to then find their contacts and kind of follow 144 00:08:25,680 --> 00:08:28,440 Speaker 3: the chain down the line. And so the first guy 145 00:08:28,480 --> 00:08:31,240 Speaker 3: we found is a professor at the Chinese University of 146 00:08:31,240 --> 00:08:37,760 Speaker 3: Hong Kong. He buys and uses these high end GPUs 147 00:08:37,800 --> 00:08:42,520 Speaker 3: in servers for educational purposes and research purposes. And as 148 00:08:42,520 --> 00:08:46,600 Speaker 3: soon as we made contact with him, we were able 149 00:08:46,640 --> 00:08:50,160 Speaker 3: to work with someone who first of all speaks you know, 150 00:08:50,360 --> 00:08:54,439 Speaker 3: native level excellent English, so that certainly helped. And then 151 00:08:54,480 --> 00:08:57,720 Speaker 3: secondly who has a lot of local knowledge and contacts. 152 00:08:58,040 --> 00:09:00,280 Speaker 3: And I think a lot of it is just kind 153 00:09:00,280 --> 00:09:06,520 Speaker 3: of knowing which, in our case, which link in the 154 00:09:06,600 --> 00:09:09,760 Speaker 3: chain you need next to make the story makes sense. 155 00:09:09,840 --> 00:09:13,119 Speaker 3: So in his case, he's a user of these GPUs. 156 00:09:13,600 --> 00:09:16,200 Speaker 3: So next we need is some kind of supplier, you know, 157 00:09:16,200 --> 00:09:17,439 Speaker 3: and then we need their supplier. 158 00:09:17,679 --> 00:09:22,200 Speaker 2: Yeah, so what it's all quite complex. I did watch 159 00:09:22,240 --> 00:09:24,680 Speaker 2: the video and it was it was what three three 160 00:09:24,720 --> 00:09:28,600 Speaker 2: hours or so or more even really wonderful. And so 161 00:09:29,120 --> 00:09:31,679 Speaker 2: what exactly are the GPUs that are restricted? Was it 162 00:09:31,880 --> 00:09:35,480 Speaker 2: just aigpus? Was it like consumer level ones as well? 163 00:09:35,880 --> 00:09:37,960 Speaker 2: What is exactly being handed around? 164 00:09:39,400 --> 00:09:45,800 Speaker 3: It's both. So on the high end side, there's some 165 00:09:46,640 --> 00:09:51,240 Speaker 3: smuggling and presence of these data center class GPUs, so 166 00:09:51,280 --> 00:09:55,840 Speaker 3: that'd be things like EH one hundreds, Hopper one hundreds, 167 00:09:55,880 --> 00:09:59,040 Speaker 3: A one hundreds. We actually physically saw several of those 168 00:10:00,000 --> 00:10:04,480 Speaker 3: and things like that. On the consumer side, there aren't 169 00:10:04,520 --> 00:10:06,920 Speaker 3: as many restricted consumer GPUs, but there are a few. 170 00:10:06,960 --> 00:10:10,480 Speaker 3: So the RTX fifty ninety is one of them that 171 00:10:10,520 --> 00:10:13,720 Speaker 3: has thirty two gigabytes of video memories. That's very valuable 172 00:10:13,720 --> 00:10:17,480 Speaker 3: for these use cases. The ARTAX forty ninety actually is 173 00:10:17,520 --> 00:10:21,319 Speaker 3: another one, and that one's kind of amusing because it's 174 00:10:21,440 --> 00:10:27,520 Speaker 3: so GPU right, yes, consumer GPU, yeah, and last gen 175 00:10:27,600 --> 00:10:32,600 Speaker 3: consumer GPU at that, but it's easily modified to carry 176 00:10:32,679 --> 00:10:37,160 Speaker 3: double the video memory as the actual official end video 177 00:10:37,360 --> 00:10:42,319 Speaker 3: spec and the skew so they can increase it from 178 00:10:42,320 --> 00:10:44,880 Speaker 3: twenty four to forty eight gigabytes of memory, which is 179 00:10:45,080 --> 00:10:50,440 Speaker 3: extremely valuable for these training and LLM type tasks where 180 00:10:51,520 --> 00:10:53,319 Speaker 3: it may be a last gen So it's technically a 181 00:10:53,320 --> 00:10:56,640 Speaker 3: little bit slower, but if you can fit it in memory, 182 00:10:56,800 --> 00:10:58,559 Speaker 3: you can still run it, versus if you can't fit 183 00:10:58,600 --> 00:11:00,040 Speaker 3: it in memory of the model you made, not be 184 00:11:00,080 --> 00:11:02,880 Speaker 3: able to run it at all. And so those are 185 00:11:02,920 --> 00:11:06,480 Speaker 3: the consumer cards the forty ninety. That was interesting because 186 00:11:07,040 --> 00:11:11,360 Speaker 3: one of the guys we spoke to who is effectively 187 00:11:11,400 --> 00:11:15,680 Speaker 3: a buyer and seller of these band GPUs, he was 188 00:11:15,720 --> 00:11:20,240 Speaker 3: asking me on camera. He was seeking confirmation of well, 189 00:11:20,679 --> 00:11:23,560 Speaker 3: but are you sure the fifty nineties banned? Yeah? And 190 00:11:23,600 --> 00:11:26,280 Speaker 3: I said, yeah, I'm sure. Here it is on the list. 191 00:11:27,200 --> 00:11:31,560 Speaker 3: And he was so confused because he said, but they're 192 00:11:31,600 --> 00:11:35,560 Speaker 3: just all over the market next door. We walk over 193 00:11:35,600 --> 00:11:37,160 Speaker 3: there and there's dozens of them. 194 00:11:37,320 --> 00:11:51,440 Speaker 2: So it's so strange as well, because it doesn't seem 195 00:11:51,480 --> 00:11:54,640 Speaker 2: like you basically walked off the plane and had a 196 00:11:54,679 --> 00:11:58,280 Speaker 2: price sheet from just a guy, Like it wasn't clear 197 00:11:58,320 --> 00:12:00,920 Speaker 2: how you got the price sheet exactly. Perhaps you can't say, 198 00:12:00,960 --> 00:12:03,480 Speaker 2: but it was just Yeah, these things a band other 199 00:12:03,520 --> 00:12:05,160 Speaker 2: than the fact we have just like guys who sell 200 00:12:05,240 --> 00:12:09,280 Speaker 2: them literally everywhere. Yeah, getting in. 201 00:12:10,679 --> 00:12:12,840 Speaker 3: So there's a couple of ways. The one of the 202 00:12:12,840 --> 00:12:15,640 Speaker 3: ones that we showed in the video is really interesting. 203 00:12:16,800 --> 00:12:19,559 Speaker 3: To the very end of the process. We actually this 204 00:12:19,600 --> 00:12:23,079 Speaker 3: is one of the reasons we delayed initial publication. We 205 00:12:23,080 --> 00:12:27,560 Speaker 3: were able to, with help of a viewer, locate someone 206 00:12:27,600 --> 00:12:30,520 Speaker 3: you could actually classify as a smuggler in the US. 207 00:12:30,679 --> 00:12:34,360 Speaker 3: So someone from China in the US. 208 00:12:34,960 --> 00:12:35,440 Speaker 5: And. 209 00:12:37,000 --> 00:12:41,400 Speaker 3: He buys GPUs from people in the US. He drives 210 00:12:41,440 --> 00:12:45,240 Speaker 3: around all over and buys most through forty nineties, which 211 00:12:45,240 --> 00:12:48,600 Speaker 3: are in a high demand in China, and then he'll 212 00:12:48,600 --> 00:12:52,440 Speaker 3: strip those downs that could be removing the cooler. I 213 00:12:52,480 --> 00:12:56,160 Speaker 3: don't think this particular guy disolders them. I've heard stories 214 00:12:56,160 --> 00:12:59,040 Speaker 3: of some people desoldering GPUs and mailing them that would 215 00:12:59,040 --> 00:13:02,000 Speaker 3: reduce your shipping costs a lot, increase your margin. But 216 00:13:02,120 --> 00:13:05,080 Speaker 3: he buys the forty nineties ships the boards back his 217 00:13:05,160 --> 00:13:08,720 Speaker 3: profits about three hundred dollars US per board before taxes. 218 00:13:08,920 --> 00:13:11,640 Speaker 3: It wasn't a one hundred percent clear if he does or 219 00:13:11,679 --> 00:13:17,600 Speaker 3: doesn't pay taxes. And then at that point, once they 220 00:13:17,880 --> 00:13:23,920 Speaker 3: are either shipped into Hong Kong or Macau directly if 221 00:13:23,960 --> 00:13:27,400 Speaker 3: they feel gutsy enough, or indirectly through a third country, 222 00:13:27,400 --> 00:13:31,000 Speaker 3: a middle country where there is no export control, you know, 223 00:13:31,040 --> 00:13:33,840 Speaker 3: then they arrive and they get redistributed from there. There's 224 00:13:33,880 --> 00:13:37,680 Speaker 3: sometimes hand carried as well, like by students, international students 225 00:13:37,720 --> 00:13:39,840 Speaker 3: who just fly back with a fifty ninety. They bought 226 00:13:39,840 --> 00:13:41,960 Speaker 3: it best Buy or something, and they can double their 227 00:13:42,000 --> 00:13:43,880 Speaker 3: money roughly if they're lucky. 228 00:13:44,120 --> 00:13:47,040 Speaker 2: That's so, it's wacky that there is such a big 229 00:13:47,080 --> 00:13:50,280 Speaker 2: black market, but I guess the export bends just created it, 230 00:13:50,320 --> 00:13:55,200 Speaker 2: as did this whole AI, this crazed AI moment. Did 231 00:13:55,320 --> 00:13:58,000 Speaker 2: in fact, while you were over over in Hong Kong, 232 00:13:58,040 --> 00:14:00,199 Speaker 2: did you actually go to China or can't remember? It 233 00:14:00,280 --> 00:14:02,880 Speaker 2: was just in home? So did you? This is an 234 00:14:02,960 --> 00:14:06,160 Speaker 2: anecdotal thing. How was the pressure of AI there? How 235 00:14:06,200 --> 00:14:09,400 Speaker 2: big was the conversation the advertisement? Did you see it 236 00:14:09,880 --> 00:14:11,800 Speaker 2: on the same level? It's fine if you didn't, I'm 237 00:14:11,800 --> 00:14:12,400 Speaker 2: just curious. 238 00:14:12,640 --> 00:14:15,200 Speaker 3: I saw it more in Taiwan, which we went to 239 00:14:15,240 --> 00:14:18,040 Speaker 3: as well. So yeah, after the China trip, we went 240 00:14:18,080 --> 00:14:20,840 Speaker 3: over to Taiwan, and I don't think it made it 241 00:14:20,840 --> 00:14:23,720 Speaker 3: into the cut, but we shot a short clip of 242 00:14:24,840 --> 00:14:27,359 Speaker 3: just a bus, like a public bus with a gigantic 243 00:14:27,920 --> 00:14:32,080 Speaker 3: AI the laptop advertisement on it. So I would say 244 00:14:32,560 --> 00:14:34,720 Speaker 3: it's kind of at least there, it's similar to what 245 00:14:34,720 --> 00:14:39,200 Speaker 3: you see here where it's you know, if you're anywhere 246 00:14:39,200 --> 00:14:42,440 Speaker 3: in the vicinity of a best Buy or a city 247 00:14:42,440 --> 00:14:46,360 Speaker 3: that has technology dealers, then you're gonna see AI ads 248 00:14:46,800 --> 00:14:50,600 Speaker 3: In China, Yeah, I saw it in the tech markets 249 00:14:50,600 --> 00:14:54,040 Speaker 3: for sure, some banners, but it just it. I don't know. 250 00:14:54,120 --> 00:14:58,440 Speaker 3: That's I only understand Chinese if I'm actively trying to 251 00:14:58,520 --> 00:14:59,960 Speaker 3: understand it, you know. So it's I think a lot 252 00:15:00,120 --> 00:15:03,160 Speaker 3: like it's filtered out for me, right, Yeah, I was good. 253 00:15:03,360 --> 00:15:05,400 Speaker 2: Yeah, I assume you speak some You seem to speak 254 00:15:05,440 --> 00:15:08,360 Speaker 2: some Chinese. I can't speak anything but English, so I 255 00:15:08,600 --> 00:15:10,240 Speaker 2: could not correct your Chinese, doll. 256 00:15:10,400 --> 00:15:12,800 Speaker 3: You do speak more proper English than I do, though, 257 00:15:12,840 --> 00:15:13,160 Speaker 3: so I'll. 258 00:15:13,080 --> 00:15:16,480 Speaker 2: Get Yeah, yeah, I speak all kinds of it. These days. 259 00:15:16,280 --> 00:15:19,240 Speaker 2: It's not going well. People criticize my accent all the time. 260 00:15:19,800 --> 00:15:23,600 Speaker 2: So as far as how the government sets these limits, 261 00:15:23,600 --> 00:15:25,640 Speaker 2: one thing he did really eloquently towards the beginning of 262 00:15:25,680 --> 00:15:27,200 Speaker 2: the video as well, was he kind of went through 263 00:15:27,240 --> 00:15:30,720 Speaker 2: how arbitrary the limits seemed to be and how in 264 00:15:30,840 --> 00:15:34,440 Speaker 2: Video seems very capable of adapting around them to the 265 00:15:34,440 --> 00:15:36,600 Speaker 2: point they almost doesn't make sense. 266 00:15:37,400 --> 00:15:43,000 Speaker 3: Yeah, yeah, they they've modified the regulations a few times. 267 00:15:43,280 --> 00:15:50,680 Speaker 3: And it's it's so this particular part I don't really 268 00:15:50,720 --> 00:15:53,560 Speaker 3: blame and video for for this, which is just trying 269 00:15:53,600 --> 00:15:57,360 Speaker 3: to you know, the government sets some regulations and a 270 00:15:57,440 --> 00:16:00,480 Speaker 3: company says, okay, we'll comply with these, and then they 271 00:16:00,520 --> 00:16:03,000 Speaker 3: comply with them, and then the government kind of says 272 00:16:03,000 --> 00:16:04,000 Speaker 3: no weight not like that. 273 00:16:04,480 --> 00:16:08,240 Speaker 2: Yeah, that's very standard for pretty much or regulation. It 274 00:16:08,320 --> 00:16:10,400 Speaker 2: maybe it's good or bad what have you, but it's not. 275 00:16:11,720 --> 00:16:13,640 Speaker 2: It happens all the time with companies. 276 00:16:14,200 --> 00:16:17,520 Speaker 3: Yes, And in this case, one of the things that 277 00:16:17,600 --> 00:16:23,560 Speaker 3: as the I guess AI world developed and the applications, 278 00:16:24,280 --> 00:16:28,360 Speaker 3: one of the things that changed was this understanding that 279 00:16:28,760 --> 00:16:32,960 Speaker 3: memory capacity is basically the most important aspect of it. 280 00:16:32,960 --> 00:16:36,080 Speaker 3: And the initial versions of the formula didn't consider memory 281 00:16:36,400 --> 00:16:40,920 Speaker 3: in any scenario. What I remember, it was based on 282 00:16:41,000 --> 00:16:45,880 Speaker 3: the marketed flops of performance on the spec sheet multiplied 283 00:16:46,160 --> 00:16:50,040 Speaker 3: against the bit length of the operation or something like that, 284 00:16:50,480 --> 00:16:56,040 Speaker 3: and they later added memory bandwidth. Memory capacity though is 285 00:16:56,080 --> 00:16:58,920 Speaker 3: still as far as I'm aware right now, is still 286 00:16:58,960 --> 00:17:01,400 Speaker 3: not considered in the formula. 287 00:17:01,520 --> 00:17:04,119 Speaker 2: Yeah, it just it feels just arbitrary. But I imagine 288 00:17:04,160 --> 00:17:07,840 Speaker 2: also it's quite difficult to find a way to do this, 289 00:17:08,200 --> 00:17:10,760 Speaker 2: especially when the people setting the regulation might not have 290 00:17:10,800 --> 00:17:12,720 Speaker 2: the best handle on computers. 291 00:17:13,880 --> 00:17:16,600 Speaker 3: Yeah, I think I think the best way to do 292 00:17:16,640 --> 00:17:18,800 Speaker 3: this would probably be if they wanted to and I'm 293 00:17:18,840 --> 00:17:21,320 Speaker 3: not I don't have a particular one way or the other, 294 00:17:21,359 --> 00:17:23,159 Speaker 3: but if they wanted to control it, I think the 295 00:17:23,160 --> 00:17:26,120 Speaker 3: best way is probably some kind of benchmark or set 296 00:17:26,160 --> 00:17:28,199 Speaker 3: of benchmarks, just like you would do for a review, 297 00:17:29,000 --> 00:17:33,200 Speaker 3: where you know, you say, okay, if it exceeds some 298 00:17:33,240 --> 00:17:36,240 Speaker 3: performance threshold in this real world use case or multiple 299 00:17:36,240 --> 00:17:40,280 Speaker 3: of them in aggregate, then it's control, it's export controlled. 300 00:17:40,320 --> 00:17:43,560 Speaker 3: But instead they just kind of calculate based on the 301 00:17:43,600 --> 00:17:46,040 Speaker 3: spec sheet and the companies that make these things. Obviously, 302 00:17:46,960 --> 00:17:51,119 Speaker 3: you know, full appropriate credit to Nvidia. The company is 303 00:17:51,200 --> 00:17:54,320 Speaker 3: extremely confident with making GPUs. They know what they're doing, 304 00:17:55,119 --> 00:17:58,159 Speaker 3: and they're going to know a lot more about how 305 00:17:58,160 --> 00:18:03,560 Speaker 3: to tweak those dials to comply then a government agency 306 00:18:03,560 --> 00:18:05,159 Speaker 3: will so. 307 00:18:05,160 --> 00:18:08,040 Speaker 2: So on the subject of Nvidia knowing things, do you 308 00:18:08,040 --> 00:18:10,760 Speaker 2: think that do they seem aware of any of this, 309 00:18:11,440 --> 00:18:13,879 Speaker 2: like any of any of this smuggling happening. Do they 310 00:18:13,920 --> 00:18:16,280 Speaker 2: seem to Did you get any comment out of them? 311 00:18:17,640 --> 00:18:20,520 Speaker 3: We asked a lot of the people in the chain 312 00:18:20,880 --> 00:18:27,120 Speaker 3: what their perception was of Nvidia's awareness, and so the 313 00:18:27,119 --> 00:18:30,719 Speaker 3: professor at the Chinese University of Hong Kong, he had 314 00:18:30,720 --> 00:18:32,159 Speaker 3: a great answer when I asked him, do you think 315 00:18:32,280 --> 00:18:35,639 Speaker 3: Nvidia knows? And he kind of he paused for a second, 316 00:18:35,720 --> 00:18:38,000 Speaker 3: and then he said, how could they not? You know, 317 00:18:38,240 --> 00:18:42,119 Speaker 3: like these things are very expensive. You've got thirty plus 318 00:18:42,119 --> 00:18:45,520 Speaker 3: thousand dollars GPUs and it just seems like they would 319 00:18:45,520 --> 00:18:48,840 Speaker 3: all be tracked. So what was the term? 320 00:18:48,920 --> 00:18:53,600 Speaker 2: There was a specific terms like kind of like basically 321 00:18:53,600 --> 00:18:55,280 Speaker 2: equivalent of ignoring him. 322 00:18:55,840 --> 00:18:59,840 Speaker 3: Yeah, so the Chinese has done eg and b egn, 323 00:19:00,000 --> 00:19:03,320 Speaker 3: which means open one eye, close one eye, or turn 324 00:19:03,359 --> 00:19:04,760 Speaker 3: a blind eye? Is the idiot? 325 00:19:04,840 --> 00:19:07,680 Speaker 2: Yeah, And that's awesome. 326 00:19:08,119 --> 00:19:11,040 Speaker 3: It's just very cool. Yeah, it's super I mean the 327 00:19:11,400 --> 00:19:16,000 Speaker 3: coolest part about that was multiple people said that exact 328 00:19:16,200 --> 00:19:22,800 Speaker 3: idiom and it's it's just it's such a specific choice 329 00:19:22,840 --> 00:19:25,719 Speaker 3: of words. It'd be the same as if you asked 330 00:19:25,760 --> 00:19:28,320 Speaker 3: like three people a question, you know, they all said, oh, 331 00:19:28,400 --> 00:19:30,840 Speaker 3: they turn a blind eye. At some point you're like, Okay, 332 00:19:30,880 --> 00:19:36,040 Speaker 3: this seems to be a really consistent belief here, But yeah, 333 00:19:36,040 --> 00:19:40,639 Speaker 3: I think from a video standpoint, it's whether or not 334 00:19:40,680 --> 00:19:45,720 Speaker 3: they want to actually control what gets in. It's hard 335 00:19:45,760 --> 00:19:49,560 Speaker 3: to say. Certainly, from a pure financial standpoint. The more 336 00:19:49,640 --> 00:19:53,360 Speaker 3: gps they sell, the more money they make that they 337 00:19:53,359 --> 00:19:55,040 Speaker 3: probably don't really care who they sell them to as 338 00:19:55,080 --> 00:19:56,680 Speaker 3: long as they sell them to someone and then where 339 00:19:56,680 --> 00:19:59,320 Speaker 3: they end up after that. I mean, maybe to the 340 00:19:59,359 --> 00:20:01,280 Speaker 3: extent that video doesn't want to piss off the United 341 00:20:01,320 --> 00:20:05,719 Speaker 3: States government. They you know, they try to comply as 342 00:20:05,800 --> 00:20:08,520 Speaker 3: much as they reasonably can. But how much do they 343 00:20:08,520 --> 00:20:13,480 Speaker 3: actually try to control ingress into a country that's export controlled. 344 00:20:13,480 --> 00:20:17,800 Speaker 3: It's hard to say, But certainly I think they are 345 00:20:17,840 --> 00:20:19,960 Speaker 3: aware of this happening, and I just I have a 346 00:20:20,000 --> 00:20:24,080 Speaker 3: really hard time believing that they are unaware their gps 347 00:20:24,119 --> 00:20:26,320 Speaker 3: are making it into China when they shouldn't be. 348 00:20:26,760 --> 00:20:28,879 Speaker 2: And I imagine it must be quite hard to stop. 349 00:20:28,960 --> 00:20:31,080 Speaker 2: Like I hate to have any sympathy for in video, 350 00:20:31,520 --> 00:20:33,840 Speaker 2: but it does. It does feel like something that would 351 00:20:33,880 --> 00:20:37,480 Speaker 2: be kind of a moving target. Because one thing that 352 00:20:37,560 --> 00:20:42,000 Speaker 2: was consistent across the whole video was the ingenuity of 353 00:20:42,040 --> 00:20:45,520 Speaker 2: the people involved. It does make you see like American 354 00:20:45,720 --> 00:20:48,439 Speaker 2: entrepreneurs sometimes and they're like, oh, yeah, it's tough to 355 00:20:48,840 --> 00:20:51,159 Speaker 2: you know that the hours are very tough, and you know, 356 00:20:51,480 --> 00:20:54,080 Speaker 2: I was covering all not and all of these guys 357 00:20:54,080 --> 00:20:55,760 Speaker 2: were just like, yeah, I've got a thing in the 358 00:20:55,760 --> 00:20:58,560 Speaker 2: back of my car full of wires. Yeah I've got 359 00:20:58,840 --> 00:20:59,880 Speaker 2: I have a whole rig. 360 00:21:01,240 --> 00:21:04,720 Speaker 3: Yeah it's crazy, and that is I mean again, you know, 361 00:21:05,480 --> 00:21:13,320 Speaker 3: from that particular side of things, I don't suspect. I mean, look, 362 00:21:13,320 --> 00:21:16,840 Speaker 3: it's a government regulation, and if the government wants to 363 00:21:16,880 --> 00:21:19,639 Speaker 3: regulate it, it's kind of up to them. Obviously the 364 00:21:19,640 --> 00:21:22,240 Speaker 3: companies need to comply with it, but I wouldn't expect 365 00:21:22,240 --> 00:21:26,879 Speaker 3: a company, even Nvidia, to send in any kind of enforcement. 366 00:21:27,280 --> 00:21:29,960 Speaker 3: Maybe what they would do is if they wanted to 367 00:21:30,080 --> 00:21:34,040 Speaker 3: enforce it, you know, they might identify, Okay, this purchaser, 368 00:21:35,480 --> 00:21:39,080 Speaker 3: there's evidence of this purchaser exporting to China anyway, and 369 00:21:39,119 --> 00:21:41,199 Speaker 3: so we're just gonna stop selling to them or something. 370 00:21:41,240 --> 00:21:44,080 Speaker 3: But that's kind of I think the most you might 371 00:21:44,119 --> 00:21:48,600 Speaker 3: expect of them. But yeah, as far as the ingenuity, 372 00:21:49,119 --> 00:21:50,880 Speaker 3: it was pretty crazy. Like the guy in the US 373 00:21:50,920 --> 00:21:53,760 Speaker 3: who buys the cards, he he just got a Prius 374 00:21:53,800 --> 00:21:57,760 Speaker 3: with a massive like battery bank in the back, like 375 00:21:57,840 --> 00:22:00,440 Speaker 3: a ups with a lithium ion battery in it. He's 376 00:22:00,440 --> 00:22:02,600 Speaker 3: got a test bench hooked up to it, an ATX 377 00:22:02,640 --> 00:22:05,480 Speaker 3: test bench. Pops the trunk open. He's got a spare 378 00:22:05,560 --> 00:22:07,680 Speaker 3: license plate in the trunk for some reason. I don't 379 00:22:07,760 --> 00:22:12,359 Speaker 3: know why. Uh, that's not to ask. Yeah, I don't 380 00:22:12,359 --> 00:22:15,200 Speaker 3: need to know. And uh yeah. And if you want 381 00:22:15,200 --> 00:22:17,359 Speaker 3: to sell him your card, you're just some American new 382 00:22:17,359 --> 00:22:21,480 Speaker 3: found him on Facebook Marketplace. He at the time we 383 00:22:21,800 --> 00:22:24,600 Speaker 3: were talking with him, he was paying two thousand dollars flat. 384 00:22:25,160 --> 00:22:27,560 Speaker 3: He makes three hundred when he sells it, and you 385 00:22:27,600 --> 00:22:30,680 Speaker 3: bring them the card, meet him wherever he tests it, 386 00:22:31,400 --> 00:22:34,119 Speaker 3: and in the back of the you know, in the 387 00:22:34,160 --> 00:22:36,600 Speaker 3: trunk of the car, he just runs a benchmark, make 388 00:22:36,640 --> 00:22:39,960 Speaker 3: sure it's good, and then buys it and ships it 389 00:22:40,000 --> 00:22:44,639 Speaker 3: back or carries it back, depending on how secure he feels. 390 00:22:44,320 --> 00:22:47,160 Speaker 3: And then also to your point of ingenuity, the shops 391 00:22:47,920 --> 00:22:50,600 Speaker 3: that are capable of modifying cards, where they're just they're 392 00:22:50,680 --> 00:22:53,120 Speaker 3: repair shops, and they just happen to be really good 393 00:22:53,160 --> 00:23:00,000 Speaker 3: at board level BGA or ballgrid array device repair or 394 00:23:00,119 --> 00:23:03,400 Speaker 3: or swapping, so they can pull memory modules. They can 395 00:23:03,480 --> 00:23:08,000 Speaker 3: add memory modules, swap the GPU between PCBs, and they're 396 00:23:08,040 --> 00:23:10,560 Speaker 3: able to keep Silicon in service where you know, they're 397 00:23:10,560 --> 00:23:13,000 Speaker 3: just repair shops, so it's kind of like from their perspective, look, 398 00:23:13,000 --> 00:23:16,280 Speaker 3: they're not doing anything illegal in China, they're just keeping 399 00:23:16,280 --> 00:23:19,840 Speaker 3: boards in service. But for people who need these high 400 00:23:19,840 --> 00:23:23,040 Speaker 3: capacity GPUs for AI tasks and they having to be 401 00:23:23,080 --> 00:23:26,880 Speaker 3: export controlled, a shop like this would be capable of 402 00:23:26,960 --> 00:23:31,200 Speaker 3: taking a card that has died, you know, through through 403 00:23:31,320 --> 00:23:35,639 Speaker 3: use or just needs to be upfit to be better, 404 00:23:36,240 --> 00:23:38,960 Speaker 3: and they can do that for pretty cheap and basically 405 00:23:39,000 --> 00:23:40,879 Speaker 3: in a couple hours. 406 00:23:40,680 --> 00:23:44,160 Speaker 2: And it's not illegal to sell them in China once 407 00:23:44,240 --> 00:23:47,000 Speaker 2: they are there, right, They've like that is not a 408 00:23:47,440 --> 00:23:49,879 Speaker 2: like China is not prosecuting people for any of this. 409 00:23:50,840 --> 00:23:54,800 Speaker 3: Correct that it is not illegal to own or sell 410 00:23:55,600 --> 00:24:00,880 Speaker 3: these GPUs or buy them in China. It's legal if 411 00:24:00,920 --> 00:24:05,240 Speaker 3: you are an American company, if you're Nvidia or your 412 00:24:05,920 --> 00:24:08,199 Speaker 3: best buyer or a partner or whatever. It would be 413 00:24:08,200 --> 00:24:12,719 Speaker 3: illegal to sell the export controlled cards to a company 414 00:24:12,800 --> 00:24:17,680 Speaker 3: or an entity or person in China. But once it's there, yeah, 415 00:24:17,720 --> 00:24:20,320 Speaker 3: there's no real control over it. The only the only 416 00:24:20,359 --> 00:24:24,880 Speaker 3: place that we found that China steps in is if 417 00:24:24,920 --> 00:24:29,399 Speaker 3: there's some kind of undeclared ingress through port. So like, 418 00:24:29,440 --> 00:24:34,320 Speaker 3: for example, there were CPUs and GPUs at various points 419 00:24:34,320 --> 00:24:37,199 Speaker 3: in the last several years that were smuggled in with 420 00:24:37,280 --> 00:24:40,280 Speaker 3: creates of live lobsters or prosthetic baby bombs. 421 00:24:40,320 --> 00:24:41,160 Speaker 4: But was that real? 422 00:24:41,359 --> 00:24:42,520 Speaker 2: Was the lobes the thing real? 423 00:24:43,000 --> 00:24:45,960 Speaker 3: The lobster things real? Despite what Nvidia said. 424 00:24:59,640 --> 00:25:04,040 Speaker 4: Hell yeah, I love the So what actually describe the full. 425 00:25:03,800 --> 00:25:06,159 Speaker 2: Loves the situation? I think everyone wants will want to 426 00:25:06,160 --> 00:25:06,880 Speaker 2: hear it over lobs. 427 00:25:08,000 --> 00:25:15,080 Speaker 3: Yes, So the loster situation as it were is basically 428 00:25:15,119 --> 00:25:21,680 Speaker 3: there was a Hong Kong uh port, so customs intercepted 429 00:25:21,800 --> 00:25:25,040 Speaker 3: a white van I think it was. There's like a 430 00:25:25,119 --> 00:25:27,880 Speaker 3: customs post they post photos of the things that they 431 00:25:28,080 --> 00:25:32,200 Speaker 3: are able to snipe being brought in improperly. So Hong 432 00:25:32,280 --> 00:25:35,919 Speaker 3: Kong's a free trade port, but there's still certain declarations 433 00:25:35,960 --> 00:25:38,920 Speaker 3: you have to make and sometimes people won't make those. 434 00:25:39,520 --> 00:25:42,480 Speaker 3: And so there was a van bringing in crates of 435 00:25:43,080 --> 00:25:48,879 Speaker 3: undeclared live lobsters driving on I think it was the 436 00:25:49,040 --> 00:25:51,960 Speaker 3: Macau or the jew Hai Bridge or something. But anyway, 437 00:25:52,480 --> 00:25:57,080 Speaker 3: they got intercepted and alongside the live lobsters were GPUs 438 00:25:58,000 --> 00:26:00,680 Speaker 3: and they were older, so I think they were quadros, 439 00:26:00,720 --> 00:26:05,280 Speaker 3: but they're brought it all being brought in improperly. And yeah, 440 00:26:05,359 --> 00:26:07,439 Speaker 3: so customs post the thing. They're like, we found this, 441 00:26:07,600 --> 00:26:09,119 Speaker 3: you know, and I think that's just kind of like 442 00:26:09,160 --> 00:26:11,960 Speaker 3: the here's us doing our job and maybe scared people 443 00:26:12,000 --> 00:26:19,160 Speaker 3: off or whatever. And then later AI startup Anthropic released 444 00:26:19,160 --> 00:26:24,640 Speaker 3: a statement talking about how there needed to be more 445 00:26:26,160 --> 00:26:31,120 Speaker 3: strictly enforced export control in the US, and they mentioned 446 00:26:31,200 --> 00:26:37,320 Speaker 3: smuggling of GPUs with lobsters, and Vidia responded and I 447 00:26:37,320 --> 00:26:39,000 Speaker 3: don't have their statement in front of me right now, 448 00:26:39,040 --> 00:26:42,440 Speaker 3: but it was something along the lines of it would 449 00:26:42,480 --> 00:26:46,919 Speaker 3: be better if American you know, AI companies would focus 450 00:26:46,920 --> 00:26:49,760 Speaker 3: on innovation rather than the part I do remember was 451 00:26:50,080 --> 00:26:56,200 Speaker 3: tell tall tales of prosthetic baby bump smuggling or GPU's 452 00:26:56,200 --> 00:27:00,159 Speaker 3: sensitive electronics being smuggled with live lobsters. Turns out that 453 00:27:00,280 --> 00:27:02,280 Speaker 3: is not a tall tale. That is a thing that happened. 454 00:27:02,760 --> 00:27:06,600 Speaker 2: So, yes, this is the thing people say. Jensen Wong 455 00:27:06,720 --> 00:27:09,840 Speaker 2: is cool. If you're cool, you say, yeah, well, our 456 00:27:09,920 --> 00:27:12,360 Speaker 2: chips like, we don't like this happening. This is bad, 457 00:27:12,400 --> 00:27:14,160 Speaker 2: but doing everything we can't stop it. But yeah, they're 458 00:27:14,200 --> 00:27:17,000 Speaker 2: doing James bondshit to get our GPUs, Like, come on, 459 00:27:17,200 --> 00:27:21,280 Speaker 2: you could just say that, like they do seem a 460 00:27:21,320 --> 00:27:24,240 Speaker 2: little sensitive, and we're we're recording this the day before 461 00:27:24,280 --> 00:27:26,359 Speaker 2: their earnings, and I got no idea what's going to happen, 462 00:27:26,720 --> 00:27:29,600 Speaker 2: No one does. Everyone's a bit everyone's a bit worried 463 00:27:29,640 --> 00:27:32,879 Speaker 2: about it. So it's I think it. The China market 464 00:27:32,960 --> 00:27:36,440 Speaker 2: is going to get extremely interesting because they've got the 465 00:27:36,600 --> 00:27:38,560 Speaker 2: H twenty band. But now the H twenty band has 466 00:27:38,600 --> 00:27:41,920 Speaker 2: been lifted, and now we're in this weird spot where 467 00:27:42,960 --> 00:27:45,600 Speaker 2: I don't know if you've heard anything about this yourself, 468 00:27:45,680 --> 00:27:49,000 Speaker 2: where there are suggestions that the Chinese government is actively 469 00:27:49,040 --> 00:27:51,360 Speaker 2: saying not to buy in video GPUs. If you heard 470 00:27:51,359 --> 00:27:52,160 Speaker 2: anything about that. 471 00:27:52,600 --> 00:27:57,720 Speaker 3: Yes, So the Chinese government, I believe it's called the 472 00:27:57,800 --> 00:28:02,679 Speaker 3: Cyberspace be a security agency or something like that, but 473 00:28:03,600 --> 00:28:07,440 Speaker 3: one of their government agencies put out a statement following 474 00:28:08,080 --> 00:28:12,040 Speaker 3: separate statement from a state owned or at least state 475 00:28:12,840 --> 00:28:15,879 Speaker 3: you know, controlled in some capacity media report where the 476 00:28:15,880 --> 00:28:21,360 Speaker 3: statement said or at least floated the possibility that Nvidia's 477 00:28:21,520 --> 00:28:26,240 Speaker 3: H twenties may have some form of spyware in them. 478 00:28:26,800 --> 00:28:31,520 Speaker 3: And so the suggestions we're tracking or government back doors, 479 00:28:31,840 --> 00:28:34,320 Speaker 3: and so that was kind of the accusation, and Vidia 480 00:28:35,119 --> 00:28:38,040 Speaker 3: of course posted a statement saying absolutely not, this is 481 00:28:38,080 --> 00:28:42,080 Speaker 3: not a thing, and you know, just to be fair here, 482 00:28:42,280 --> 00:28:44,880 Speaker 3: there's not currently, as far as we're aware of, any 483 00:28:45,040 --> 00:28:48,360 Speaker 3: firm evidence of this. But I'm also sure that governments 484 00:28:48,360 --> 00:28:52,880 Speaker 3: know things about you know, about products from other nations 485 00:28:52,920 --> 00:28:55,680 Speaker 3: that we would never know anyway, So who knows. But 486 00:28:55,800 --> 00:28:57,840 Speaker 3: that's where it stood. Nvidia said, we don't have these. 487 00:28:58,240 --> 00:29:02,680 Speaker 3: China's agency said to companies in China that they should 488 00:29:02,720 --> 00:29:06,760 Speaker 3: be wary of using H twenties and advised against purchasing 489 00:29:06,800 --> 00:29:10,520 Speaker 3: them for fear of some kind of backdoor or spyware. 490 00:29:11,760 --> 00:29:15,320 Speaker 2: So how realistic would it be for Actually, did you 491 00:29:15,360 --> 00:29:18,600 Speaker 2: hear any scuttle about around this of switching to Huawei 492 00:29:18,640 --> 00:29:19,440 Speaker 2: chips or something? 493 00:29:19,920 --> 00:29:22,600 Speaker 3: We did ask? Yeah, I think we might have only 494 00:29:22,680 --> 00:29:26,360 Speaker 3: kept one of those questions in the actual video, but 495 00:29:26,400 --> 00:29:34,000 Speaker 3: I did ask a lot of people Huawei from the users' perspectives, 496 00:29:34,040 --> 00:29:38,080 Speaker 3: the people we spoke to, they none of the people 497 00:29:38,080 --> 00:29:41,479 Speaker 3: we spoke to are using Huawei components right now, so 498 00:29:41,640 --> 00:29:45,360 Speaker 3: they're getting more powerful from what I understand, I think 499 00:29:45,400 --> 00:29:47,920 Speaker 3: there's a few hurdles. One of them is. It's kind 500 00:29:47,960 --> 00:29:51,160 Speaker 3: of the same reason where the companies aren't even really 501 00:29:51,240 --> 00:29:54,640 Speaker 3: using AMD and they've been doing this for a while, 502 00:29:54,720 --> 00:29:58,120 Speaker 3: whereas Huawei is pretty new. And a lot of that 503 00:29:58,160 --> 00:30:03,040 Speaker 3: comes down to Kuda, where Nvidia is the proprietor of Kuda, 504 00:30:03,200 --> 00:30:08,040 Speaker 3: which is a library they can use to accelerate tasks 505 00:30:08,080 --> 00:30:10,720 Speaker 3: in which a lot of software is written to use, 506 00:30:10,760 --> 00:30:12,360 Speaker 3: I mean even just rendering a video. 507 00:30:12,880 --> 00:30:17,400 Speaker 2: It's the coding language to get accelerate computing, right, correct, Yeah, 508 00:30:17,480 --> 00:30:19,560 Speaker 2: specifically on Nvidia GPUs. 509 00:30:19,120 --> 00:30:22,800 Speaker 3: Though, yes, it specifically works on Nvidia and nothing else, 510 00:30:23,560 --> 00:30:26,440 Speaker 3: and so AMD would have to use i don't know, 511 00:30:26,480 --> 00:30:29,800 Speaker 3: open cl or something else, depending on like for video encoding, 512 00:30:29,840 --> 00:30:34,000 Speaker 3: you maybe use open CL. But that's the biggest limitation 513 00:30:34,120 --> 00:30:37,600 Speaker 3: where a lot of the sort of so called AI 514 00:30:38,120 --> 00:30:42,720 Speaker 3: tasks are reliant on Kuda to operate at any reasonable speed. 515 00:30:43,200 --> 00:30:46,240 Speaker 3: And so until there's a big push to move away 516 00:30:46,240 --> 00:30:50,920 Speaker 3: from that, it's Nvidia GPUs all the way down, you know. So, 517 00:30:51,560 --> 00:30:54,680 Speaker 3: and maybe the push is export control. Maybe the groups 518 00:30:54,680 --> 00:30:58,000 Speaker 3: in China that need a higher volume of GPUs than 519 00:30:58,040 --> 00:31:01,120 Speaker 3: they're able to easily get of actually say okay, let's 520 00:31:01,160 --> 00:31:05,440 Speaker 3: just let's put all of our resources on trying to 521 00:31:05,560 --> 00:31:08,040 Speaker 3: use open Seattle or use something out of or PyTorch 522 00:31:08,160 --> 00:31:10,720 Speaker 3: or whatever, so that may be the direction that goes. 523 00:31:11,760 --> 00:31:14,280 Speaker 2: Yeah, it just feel it does actually feel like Couder 524 00:31:14,440 --> 00:31:17,720 Speaker 2: is the thing that needs to be broken here for 525 00:31:17,840 --> 00:31:20,080 Speaker 2: better or for worse. I'm just saying that there's a 526 00:31:20,120 --> 00:31:22,479 Speaker 2: lots of talk about oh, making more powerful chips here 527 00:31:22,520 --> 00:31:24,560 Speaker 2: and there with AMD or Huawei, but it feels like 528 00:31:24,640 --> 00:31:27,640 Speaker 2: couder really is the thing that they actually need to change. 529 00:31:28,040 --> 00:31:30,560 Speaker 2: And I guess the probably explaining to the markets what 530 00:31:30,680 --> 00:31:33,320 Speaker 2: Couder is is difficult enough, but it feels that that 531 00:31:33,400 --> 00:31:35,440 Speaker 2: if they could move past Kuda, that would be the 532 00:31:35,480 --> 00:31:36,440 Speaker 2: actual thing that we're doing. 533 00:31:36,960 --> 00:31:39,480 Speaker 3: Yeah, it's kind of like the X eighty six situation 534 00:31:39,760 --> 00:31:42,840 Speaker 3: where getting off of X eighty six. 535 00:31:43,400 --> 00:31:46,080 Speaker 2: Just for the listeners, what do you what do you 536 00:31:46,160 --> 00:31:46,560 Speaker 2: mean by that? 537 00:31:46,920 --> 00:31:50,880 Speaker 3: Yeah, So your average computer is very likely an X 538 00:31:50,960 --> 00:31:55,000 Speaker 3: eighty six computer, which is a specific micro architecture that 539 00:31:55,040 --> 00:31:58,280 Speaker 3: has to do with how software runs and interacts with 540 00:31:58,800 --> 00:32:01,680 Speaker 3: the hardware and the other so around the system. And 541 00:32:02,080 --> 00:32:05,560 Speaker 3: so it's kind of like a prerequisite for you might 542 00:32:05,640 --> 00:32:10,440 Speaker 3: have an application that requires Windows to be compatible, and 543 00:32:11,240 --> 00:32:14,640 Speaker 3: this is another one of those types of foundational prerequisites. Okay, 544 00:32:14,720 --> 00:32:18,000 Speaker 3: X eighty six as opposed to for example, ARM and 545 00:32:18,480 --> 00:32:20,480 Speaker 3: ARM laptops made kind of a splash a while ago 546 00:32:20,520 --> 00:32:24,840 Speaker 3: because they were trying to basically assert that, hey, look 547 00:32:24,880 --> 00:32:27,320 Speaker 3: we can run stuff that typically would only run on 548 00:32:27,320 --> 00:32:30,440 Speaker 3: an old X eighty six micro architecture and now we 549 00:32:30,480 --> 00:32:34,040 Speaker 3: can run it on ARM. And so it's that is 550 00:32:34,240 --> 00:32:40,680 Speaker 3: something where Intel an AMD, although mostly Intel have a 551 00:32:41,320 --> 00:32:45,720 Speaker 3: vested interest in keeping everything on x eighty six, and 552 00:32:45,800 --> 00:32:48,720 Speaker 3: likewise on Vidia has an interest in keeping things on Cuda. 553 00:32:49,200 --> 00:32:51,959 Speaker 3: From an end user standpoint, there are benefits to this 554 00:32:52,040 --> 00:32:55,320 Speaker 3: where Kuda, for me, for rendering a video, it just 555 00:32:55,320 --> 00:32:58,720 Speaker 3: works better than something else, right, right, And so it's 556 00:32:58,800 --> 00:33:00,720 Speaker 3: kind of it is a chain or the egg problem 557 00:33:00,760 --> 00:33:02,920 Speaker 3: where it's it's kind of like, well, okay, you know, 558 00:33:02,960 --> 00:33:06,240 Speaker 3: on one hand, you don't want a company to develop 559 00:33:06,280 --> 00:33:10,240 Speaker 3: the monopoly only because there's some kind of anti competitive 560 00:33:10,280 --> 00:33:13,320 Speaker 3: practice that's keeping them in place. On the other hand, 561 00:33:13,400 --> 00:33:15,760 Speaker 3: you don't want to punish a company either for inventing 562 00:33:15,800 --> 00:33:18,720 Speaker 3: something that's just better. And that The problem we run 563 00:33:18,720 --> 00:33:22,360 Speaker 3: into with software is the developers need to choose what 564 00:33:22,440 --> 00:33:25,480 Speaker 3: to support, and a lot of times they choose n 565 00:33:25,560 --> 00:33:28,600 Speaker 3: video and maybe not something else, because if you have 566 00:33:28,640 --> 00:33:30,680 Speaker 3: limited resources, you're going to go with what ninety percent 567 00:33:30,720 --> 00:33:32,800 Speaker 3: of the market has. It's kind of a self fulfilling, 568 00:33:32,840 --> 00:33:36,440 Speaker 3: you know, it to feedback loop basically. So to expand 569 00:33:36,440 --> 00:33:38,440 Speaker 3: it to something like Huawei, they would have to break 570 00:33:38,480 --> 00:33:40,200 Speaker 3: that kuda. I don't know if you want to call 571 00:33:40,240 --> 00:33:43,720 Speaker 3: it a moat or something like that, but or a 572 00:33:43,760 --> 00:33:45,680 Speaker 3: walled garden I guess either one of those or a 573 00:33:45,680 --> 00:33:48,720 Speaker 3: walled garden with a moat around it. But they would 574 00:33:48,760 --> 00:33:52,120 Speaker 3: have to break that down to really become more viable 575 00:33:52,400 --> 00:33:53,560 Speaker 3: and increase the performance. 576 00:33:54,160 --> 00:33:56,680 Speaker 2: So before we go, did you catch any black Well 577 00:33:56,720 --> 00:33:58,880 Speaker 2: out there or any signs that black Well would even 578 00:33:58,960 --> 00:33:59,600 Speaker 2: make it out there? 579 00:34:00,320 --> 00:34:03,400 Speaker 3: Yeah, So in addition to the fifteen nineties which are Blackwell, 580 00:34:03,720 --> 00:34:07,880 Speaker 3: as far as like the high end server stuff, we 581 00:34:08,040 --> 00:34:12,040 Speaker 3: did work with. So there's two companies we worked with here. 582 00:34:12,920 --> 00:34:16,239 Speaker 3: One of them we physically visited in Taiwan, and there's 583 00:34:16,239 --> 00:34:20,799 Speaker 3: another one we were talking to in Stapore. And what 584 00:34:20,840 --> 00:34:25,400 Speaker 3: we learned is that there is some intermediary basically passed 585 00:34:25,400 --> 00:34:29,160 Speaker 3: throughs that can happen through for example, a Taiwanese testing 586 00:34:29,200 --> 00:34:34,080 Speaker 3: agency where a Chinese company might ask them to purchase 587 00:34:34,840 --> 00:34:39,640 Speaker 3: Blackwell servers or high end Hopper servers and bring them 588 00:34:39,680 --> 00:34:44,520 Speaker 3: into Taiwan under their company performed testing or pre testing 589 00:34:44,520 --> 00:34:49,200 Speaker 3: on them, assembly, whatever, packaging, whether it's a facade or not, 590 00:34:49,400 --> 00:34:52,400 Speaker 3: basically the stated purposes, make sure it all works, and 591 00:34:52,440 --> 00:34:55,920 Speaker 3: then once it does, forward that shipment to their customer, 592 00:34:56,120 --> 00:34:59,759 Speaker 3: the Taiwanese company's customer in China, and so that would 593 00:34:59,760 --> 00:35:02,839 Speaker 3: allow them to potentially get into China. And so in 594 00:35:02,840 --> 00:35:05,319 Speaker 3: that situation, yeah, we were showing a room filled with 595 00:35:05,680 --> 00:35:10,479 Speaker 3: for example GB or Grace Blackwell servers that were being 596 00:35:10,600 --> 00:35:12,000 Speaker 3: tested and prepped. 597 00:35:13,400 --> 00:35:16,560 Speaker 2: Yeah, well I'm gonna I think we should wrap it there, Steve. So, 598 00:35:17,320 --> 00:35:19,320 Speaker 2: the status of the video right now as we record 599 00:35:19,360 --> 00:35:22,040 Speaker 2: this is it's offline. But by the time this goes live, 600 00:35:22,480 --> 00:35:24,520 Speaker 2: you know, in I think September the third, it will 601 00:35:24,520 --> 00:35:28,640 Speaker 2: go up. Hopefully it's back on. Where can people support. 602 00:35:28,320 --> 00:35:31,560 Speaker 3: You gamers and access just the YouTube channel is the 603 00:35:31,600 --> 00:35:36,080 Speaker 3: best place to go. And yeah, hopefully it's back up 604 00:35:36,080 --> 00:35:38,080 Speaker 3: within It should be within a couple of days of 605 00:35:38,120 --> 00:35:43,920 Speaker 3: you posting this, unless Bloomberg submits a files a lawsuit, 606 00:35:44,200 --> 00:35:46,200 Speaker 3: in which case they'll hear about you. Everyone will hear 607 00:35:46,200 --> 00:35:46,719 Speaker 3: about it. 608 00:35:47,160 --> 00:35:48,959 Speaker 2: Oh and no hear about it from me as well. 609 00:35:48,960 --> 00:35:52,200 Speaker 2: This is utter bullshit what Bloomberg is doing. It's ridiculous. 610 00:35:52,600 --> 00:35:54,759 Speaker 2: But you have been listening to Better Offline. I'm at 611 00:35:54,840 --> 00:35:57,799 Speaker 2: Zetron Steve. Thank you so much for joining us, and 612 00:35:58,080 --> 00:35:59,960 Speaker 2: yeah we can do something in a couple months as well, 613 00:36:00,160 --> 00:36:02,040 Speaker 2: but always push to having you ma'am. 614 00:36:02,160 --> 00:36:04,680 Speaker 3: Thank you right, thanks so much. 615 00:36:12,880 --> 00:36:14,560 Speaker 1: Thank you for listening to Better Offline. 616 00:36:14,680 --> 00:36:17,120 Speaker 2: The editor and composer of the Better Offline theme song 617 00:36:17,160 --> 00:36:19,840 Speaker 2: is Matasowski. You can check out more of his music 618 00:36:19,840 --> 00:36:23,520 Speaker 2: and audio projects at Mattasowski dot com, m A T 619 00:36:23,520 --> 00:36:27,960 Speaker 2: T O S O W s ki dot com. You 620 00:36:28,000 --> 00:36:30,479 Speaker 2: can email me at easy at Better offline dot com 621 00:36:30,600 --> 00:36:32,880 Speaker 2: or visit Better Offline dot com to find more podcast 622 00:36:32,960 --> 00:36:36,320 Speaker 2: links and of course my newsletter. I also really recommend 623 00:36:36,320 --> 00:36:38,239 Speaker 2: you go to chat dot Where's Youreed dot at to 624 00:36:38,320 --> 00:36:40,680 Speaker 2: visit the discord, and go to our slash. 625 00:36:40,400 --> 00:36:43,520 Speaker 1: Better Offline to check out our reddit. Thank you so 626 00:36:43,640 --> 00:36:44,400 Speaker 1: much for listening. 627 00:36:45,200 --> 00:36:47,920 Speaker 5: Better Offline is a production of cool Zone Media. For 628 00:36:48,040 --> 00:36:51,200 Speaker 5: more from cool Zone Media, visit our website cool Zonemedia 629 00:36:51,239 --> 00:36:54,120 Speaker 5: dot com, or check us out on the iHeartRadio app, 630 00:36:54,160 --> 00:37:17,239 Speaker 5: Apple Podcasts, or wherever you get your podcasts,