1 00:00:00,400 --> 00:00:05,000 Last week, Deepseek, a Chinese AI company, released a new 2 00:00:05,000 --> 00:00:08,840 reasoning model that turned out to be comparable to models made 3 00:00:08,840 --> 00:00:13,360 by firms like Open AI, Google, Meta, and Anthropic. 4 00:00:13,680 --> 00:00:17,440 The big difference is that Deepseek claims to have built 5 00:00:17,440 --> 00:00:21,880 their model at a fraction of the cost big tech firms have spent. 6 00:00:22,320 --> 00:00:28,000 AUS export ban on Nvidia's best AI chips means that Deepseek has 7 00:00:28,000 --> 00:00:31,440 done what many thought was impossible possible building and 8 00:00:31,440 --> 00:00:35,440 training one of the most impressive AI models using the 9 00:00:35,480 --> 00:00:38,200 outdated chips available in China. 10 00:00:38,680 --> 00:00:42,400 The new model was announced in a white paper in December and 11 00:00:42,400 --> 00:00:46,400 released a few weeks ago. Over the weekend, people started 12 00:00:46,400 --> 00:00:50,600 paying a lot of attention to how good the new model was, and on 13 00:00:50,600 --> 00:00:54,680 Monday when the stock market opened, the NASDAQ opened down 14 00:00:54,680 --> 00:00:59,840 around 3 1/2 percent, with NVIDIA declining by 17%. 15 00:01:00,240 --> 00:01:05,120 While 17% sounds like a big number, to put that in context, 16 00:01:05,360 --> 00:01:09,880 this was a $600 billion decline in market cap, which is more 17 00:01:09,880 --> 00:01:13,920 than the entire market cap of ExxonMobil, which not so long 18 00:01:13,920 --> 00:01:16,360 ago was the biggest company in the world. 19 00:01:16,880 --> 00:01:22,120 While NVIDIA and the US mega cap tech companies got most of the 20 00:01:22,120 --> 00:01:25,400 news coverage, they weren't the only stocks hit. 21 00:01:25,720 --> 00:01:29,480 The realisation that a reasoning model could be built on such a 22 00:01:29,480 --> 00:01:33,640 tight budget raised doubts about the scale of spending we've seen 23 00:01:33,640 --> 00:01:36,760 from big tech. While big tech suddenly seemed 24 00:01:36,760 --> 00:01:40,240 less insulated from competition than they were previously 25 00:01:40,240 --> 00:01:43,760 believed to be, their price declines on Monday were quite 26 00:01:43,760 --> 00:01:46,440 modest. The hardest hit stocks other 27 00:01:46,440 --> 00:01:50,120 than media were the ones expected to benefit most from 28 00:01:50,120 --> 00:01:52,280 the emerging data center economy. 29 00:01:52,640 --> 00:01:56,320 Utilities like Constellation Energy, who announced a few 30 00:01:56,320 --> 00:02:00,320 months ago that they were reopening 3 Mile Island to power 31 00:02:00,320 --> 00:02:04,920 Microsoft data centres, along with other electrical utilities 32 00:02:04,920 --> 00:02:08,600 near data center hotspots, all fell hard. 33 00:02:08,919 --> 00:02:12,600 Just a few days earlier, the Stargate project had been 34 00:02:12,600 --> 00:02:16,800 announced with huge fanfare where plans for up to 20 large 35 00:02:17,240 --> 00:02:21,160 AI data centers were announced in the United States with an 36 00:02:21,160 --> 00:02:26,600 initial investment of $100 billion and plans for U to $500 37 00:02:26,600 --> 00:02:32,320 billion by 2029. GE Vernova, an energy equipment 38 00:02:32,320 --> 00:02:37,320 manufacturer, Eton, a power management company, Oracle, who 39 00:02:37,320 --> 00:02:41,520 just announced a huge data center investment, and Broadcom 40 00:02:41,680 --> 00:02:45,920 who sell advanced networking equipment to data centers all 41 00:02:45,920 --> 00:02:50,080 got hit in the sell off. Energy, commodities, copper and 42 00:02:50,080 --> 00:02:53,960 mining stocks were all hit too. There are a few things that you 43 00:02:53,960 --> 00:02:58,080 could read into the sell off 1 is just that investors now 44 00:02:58,080 --> 00:03:01,400 believe that less AI infrastructure will be needed, 45 00:03:01,680 --> 00:03:04,560 but the sell off could instead be telling us that the 46 00:03:04,560 --> 00:03:08,480 infrastructure just won't be as concentrated as was expected 47 00:03:08,680 --> 00:03:12,640 with a few huge data centres owned and run by the mega cap 48 00:03:12,640 --> 00:03:16,280 tech companies. The emergence of Deepseek caused 49 00:03:16,280 --> 00:03:19,680 investors to question whether AI will be a winner takes all 50 00:03:19,680 --> 00:03:23,120 business model like a lot of tech innovation has been in the 51 00:03:23,120 --> 00:03:27,120 past, or if it's easily replicated and we'll instead see 52 00:03:27,120 --> 00:03:31,640 lots of different models run in smaller data centers all over 53 00:03:31,640 --> 00:03:35,000 the world. The idea up until about a week 54 00:03:35,000 --> 00:03:39,760 ago was that someone would achieve a huge lead in AIA, bit 55 00:03:39,760 --> 00:03:42,720 like Google did in search and own the market. 56 00:03:43,120 --> 00:03:46,440 Google became the dominant search engine back in around 57 00:03:46,440 --> 00:03:49,880 2002 and is still dominant today. 58 00:03:50,240 --> 00:03:54,760 The question is whether AI will workout like Search Oregon not. 59 00:03:55,280 --> 00:03:59,640 The first AI reasoning model, known as O One, was released by 60 00:03:59,640 --> 00:04:03,320 Open AI last September. It was different to the prior 61 00:04:03,320 --> 00:04:06,960 models because it used a chain of thought approach to solving 62 00:04:06,960 --> 00:04:10,880 complex problems by breaking a big problem down to its 63 00:04:10,880 --> 00:04:14,800 constituent parts, then testing a number of approaches to 64 00:04:14,800 --> 00:04:18,920 solving each part in the background before presenting an 65 00:04:18,959 --> 00:04:23,440 answer along with the chain of logic that led to that answer to 66 00:04:23,440 --> 00:04:28,000 the user, it not only gave better answers, but users got to 67 00:04:28,000 --> 00:04:31,960 see how the model thinks and decide if they agree with it or 68 00:04:31,960 --> 00:04:35,560 not. As soon as O One was released, 69 00:04:35,760 --> 00:04:39,720 competitors were rushing to catch up with Google releasing a 70 00:04:39,720 --> 00:04:43,080 competing model a few months later in December. 71 00:04:43,400 --> 00:04:47,240 The thing is though, that Alibaba, the Chinese tech giant, 72 00:04:47,360 --> 00:04:51,480 had actually beaten Google by releasing their reasoning model 73 00:04:51,480 --> 00:04:56,800 called QWQ ahead of Google. Not only did Alibaba get there 74 00:04:56,800 --> 00:05:01,040 faster, but they published the model under an open license, 75 00:05:01,200 --> 00:05:05,040 meaning that anyone could dig through it to see how it works. 76 00:05:05,560 --> 00:05:10,120 This is very different to Open AI, who, despite the name of the 77 00:05:10,120 --> 00:05:13,440 company, keep the workings of their model secret. 78 00:05:13,840 --> 00:05:17,000 So let's look at whether we should believe that Deepseek 79 00:05:17,000 --> 00:05:21,280 built their model for $5.6 million, what Chinese 80 00:05:21,280 --> 00:05:25,840 competition means for big tech, for NVIDIA, and for the future 81 00:05:25,840 --> 00:05:29,120 of AI. And is this a Sputnik moment? 82 00:05:29,880 --> 00:05:34,640 Deep Seek is an interesting AI company in that it isn't part of 83 00:05:34,640 --> 00:05:37,920 a huge tech firm, nor is it VC funded. 84 00:05:38,120 --> 00:05:41,600 It was originally part of a Chinese quantitative hedge fund 85 00:05:41,600 --> 00:05:46,760 called High Flyer, and was spun out as a separate unit in 2023. 86 00:05:47,200 --> 00:05:51,520 Deepseek released a number of models since then, making the 87 00:05:51,520 --> 00:05:56,120 code open source under the MIT license, which puts very few 88 00:05:56,120 --> 00:06:01,200 restrictions on reuse, allowing users to modify the code even 89 00:06:01,200 --> 00:06:06,680 for proprietary commercial use. Despite its low cost, Deepseeks 90 00:06:06,680 --> 00:06:10,920 scores on AI performance benchmarks show that it's as 91 00:06:10,920 --> 00:06:15,160 good, if not better than the latest cutting edge models from 92 00:06:15,160 --> 00:06:19,560 the top US firms. It's almost as good as Open AIS 93 00:06:19,600 --> 00:06:24,320 01 model in the Artificial Analysis Quality Index, an 94 00:06:24,320 --> 00:06:29,320 independent AI analysis ranking and it beats Google Anthropic 95 00:06:29,320 --> 00:06:33,200 and Made As models. They released a large language 96 00:06:33,200 --> 00:06:38,320 model in December called V3 and then a reasoning model called R1 97 00:06:38,320 --> 00:06:42,480 on the 20th of January, both of which got positive reviews in 98 00:06:42,480 --> 00:06:45,320 industry publications like Semi Analysis. 99 00:06:45,720 --> 00:06:50,520 An Economist article a few days later on how China's AI labs 100 00:06:50,680 --> 00:06:54,040 were significantly better than anyone outside of China was 101 00:06:54,040 --> 00:06:58,240 giving them credit for got a lot of attention over the weekend, 102 00:06:58,440 --> 00:07:02,520 and by Monday morning people were questioning how necessary 103 00:07:02,520 --> 00:07:06,920 it was to have access to Nvidia's most expensive chips. 104 00:07:07,400 --> 00:07:13,880 NVIDIA gets to sell its H100 chips at a 1000% markup because 105 00:07:13,880 --> 00:07:18,040 of the belief that if you use the second best chip, you've no 106 00:07:18,040 --> 00:07:20,840 chance of ever catching up in AI. 107 00:07:21,200 --> 00:07:25,880 The emergence of Deepseek changed the AI CapEx narrative. 108 00:07:26,400 --> 00:07:30,760 Being a Chinese model, Deepseek does appear to be heavily 109 00:07:30,760 --> 00:07:34,280 censored, avoiding topics that are considered politically 110 00:07:34,280 --> 00:07:36,480 sensitive for the government of China. 111 00:07:36,800 --> 00:07:40,040 Users have, of course, had fun trying to trick it into 112 00:07:40,040 --> 00:07:43,840 discussing the Tiananmen Square massacre, the independence of 113 00:07:43,840 --> 00:07:48,480 Taiwan, and into making comparisons between Xi Jinping 114 00:07:48,480 --> 00:07:51,640 and Winnie the Pooh. This is not so different to the 115 00:07:51,640 --> 00:07:55,800 way that Grok appears to be hard coded to speak well of Elon 116 00:07:55,800 --> 00:07:59,920 Musk, praising his relatively slender build. 117 00:08:00,600 --> 00:08:04,440 Deepseek is not the only Chinese AI model. 118 00:08:04,680 --> 00:08:09,760 Alibaba, Tencent, Byte Dance, and Moon Shot all have models 119 00:08:09,920 --> 00:08:14,520 that are slowly catching up with US peers, most importantly by 120 00:08:14,520 --> 00:08:19,160 beating them in cost efficiency. Because of the US export 121 00:08:19,160 --> 00:08:24,000 restrictions that were placed on advanced AI chips, Chinese AI 122 00:08:24,000 --> 00:08:27,480 companies were forced to innovate with more efficient 123 00:08:27,480 --> 00:08:31,120 algorithms, architecture, and training strategies. 124 00:08:31,360 --> 00:08:35,360 According to the Deep Seq white paper, their model was trained 125 00:08:35,360 --> 00:08:41,520 using NVIDIA H 800 GPUs, which are similar to the H 100 but 126 00:08:41,520 --> 00:08:45,960 specifically tailored for the Chinese market to comply with US 127 00:08:46,040 --> 00:08:50,240 export restrictions. According to Reuters, the main 128 00:08:50,240 --> 00:08:55,120 thing NVIDIA changed in the H800 was that it reduced the chip to 129 00:08:55,120 --> 00:09:00,160 chip data transfer rate to around half that of the H100. 130 00:09:00,680 --> 00:09:05,600 In October 2023, the US government banned the export of 131 00:09:05,600 --> 00:09:10,440 H8 hundreds as well. Despite having access to worse 132 00:09:10,440 --> 00:09:14,760 chips, Deepseek managed to complete training in just two 133 00:09:14,760 --> 00:09:19,520 months at a cost of $5.6 million, a fraction of the sums 134 00:09:19,520 --> 00:09:23,360 reportedly spent by Open AI, Google, and Meta. 135 00:09:23,960 --> 00:09:28,040 Another reason that China was slow to develop AI chat models, 136 00:09:28,040 --> 00:09:31,440 according to The Economist, is that they worried about how 137 00:09:31,440 --> 00:09:35,680 sensors in China would react to models that might hallucinate 138 00:09:35,840 --> 00:09:39,720 and provide either incorrect information or come out with 139 00:09:39,720 --> 00:09:43,480 politically dangerous statements that could get the developers in 140 00:09:43,480 --> 00:09:46,120 trouble. The Chinese authorities 141 00:09:46,120 --> 00:09:50,720 eventually issued regulations to foster the AI industry and 142 00:09:50,720 --> 00:09:55,040 models started to be built, usually based on Meta's open 143 00:09:55,040 --> 00:09:59,760 source Llama model. the US chip restrictions are likely 144 00:09:59,760 --> 00:10:03,800 responsible for the efficiency of deep seats model, which 145 00:10:03,800 --> 00:10:07,400 didn't come from one huge innovation, but instead from a 146 00:10:07,400 --> 00:10:11,280 series of small improvements which when combined made a 147 00:10:11,280 --> 00:10:14,320 massive difference. The Deep seq white paper 148 00:10:14,320 --> 00:10:18,000 explains a lot of the technical details, like how they used 149 00:10:18,000 --> 00:10:23,200 float 8 bit numbers instead of 16 to speed up training and save 150 00:10:23,200 --> 00:10:25,880 memory. The problem with doing that is 151 00:10:25,880 --> 00:10:29,680 that you can lose a lot of of detail and so then they used 152 00:10:29,680 --> 00:10:33,160 other smart techniques to keep the training accurate. 153 00:10:33,760 --> 00:10:38,000 Deepseek used a mixture of experts model, which means that 154 00:10:38,000 --> 00:10:42,440 rather than training one large model, they trained 10s of 155 00:10:42,440 --> 00:10:47,000 smaller ones on more specific data that then get switched on 156 00:10:47,000 --> 00:10:50,040 or off as needed. A lot of their focus was on 157 00:10:50,040 --> 00:10:54,840 reducing communication overhead, both between nodes and within 158 00:10:54,840 --> 00:10:58,040 nodes. The server farm was reconfigured 159 00:10:58,160 --> 00:11:02,440 to let individual chips speak to each other more efficiently. 160 00:11:02,800 --> 00:11:08,000 After Deepseeks LLM was trained, it was then fine-tuned on output 161 00:11:08,000 --> 00:11:11,120 from the reasoning model, learning how to mimic its 162 00:11:11,120 --> 00:11:15,160 quality at a lower cost. A lot of this reminds me of 163 00:11:15,160 --> 00:11:18,520 older coders who I've worked with who learned how to write 164 00:11:18,520 --> 00:11:20,800 software on much simpler computers. 165 00:11:21,240 --> 00:11:25,160 The capacity constraints meant that they wrote very efficient 166 00:11:25,160 --> 00:11:27,680 code. They were often scornful of the 167 00:11:27,680 --> 00:11:31,280 bloated code written by younger programmers who never had to 168 00:11:31,280 --> 00:11:35,800 worry about efficiency. Chinese AI engineers faced with 169 00:11:35,800 --> 00:11:40,280 less efficient GP US focused on more efficient code and found 170 00:11:40,280 --> 00:11:43,280 smart ways of working around the constraints. 171 00:11:44,120 --> 00:11:48,800 Thanks to the efficiencies they found it cost around $56,000,000 172 00:11:48,800 --> 00:11:53,240 to train the new model, or about 110th of what it cost Meta to 173 00:11:53,240 --> 00:11:58,480 train their Llama model. That $5.6 million price tag has 174 00:11:58,480 --> 00:12:01,880 been getting a lot of attention, but if you read the technical 175 00:12:01,880 --> 00:12:06,240 document, this was just the cost of training, and Deep Seek are 176 00:12:06,240 --> 00:12:09,840 clear that this wasn't the overall cost of development. 177 00:12:10,120 --> 00:12:13,440 In order to reach the point of training the model, they had to 178 00:12:13,440 --> 00:12:17,000 spend possibly hundreds of millions of dollars working out 179 00:12:17,000 --> 00:12:19,960 how to get there and how to build the necessary 180 00:12:19,960 --> 00:12:23,000 infrastructure. And once they knew what to do, 181 00:12:23,200 --> 00:12:27,000 they then spent 5.6 million, $1,000,000 on compute. 182 00:12:27,280 --> 00:12:31,720 So the overall cost was much higher, but still significantly 183 00:12:31,720 --> 00:12:36,520 lower than the amount being spent by major USAI companies. 184 00:12:37,440 --> 00:12:41,440 The thing is that now the deep seek have shown the way, these 185 00:12:41,440 --> 00:12:45,120 efficiencies will significantly reduce the cost to those who 186 00:12:45,120 --> 00:12:48,800 follow in their footsteps. But that still doesn't mean that 187 00:12:48,800 --> 00:12:52,560 you can do the same and build an advanced AI model with $6 188 00:12:52,560 --> 00:12:56,760 million. Open AI are now saying that they 189 00:12:56,760 --> 00:13:00,280 have found evidence that Deepseek used their proprietary 190 00:13:00,280 --> 00:13:04,680 models to train Deepseek, having told the Financial Times that 191 00:13:04,680 --> 00:13:08,880 they had seen some evidence of distillation, which they suspect 192 00:13:08,880 --> 00:13:12,400 came from China. Distillation is a technique to 193 00:13:12,400 --> 00:13:16,120 get better performance on smaller models by using the 194 00:13:16,120 --> 00:13:19,760 outputs from larger ones, allowing them to achieve similar 195 00:13:19,760 --> 00:13:23,880 results on specific tasks at a much lower cost. 196 00:13:24,320 --> 00:13:27,960 Many have pointed out that this is the pot calling the kettle 197 00:13:27,960 --> 00:13:32,880 black, as Open AI have already been accused of building ChatGPT 198 00:13:33,040 --> 00:13:36,640 by using online content that they didn't have the rights to. 199 00:13:36,880 --> 00:13:41,120 Open AI is in fact the subject of multiple lawsuits, including 200 00:13:41,120 --> 00:13:44,760 one from the New York Times, who claimed that Open AI built 201 00:13:44,760 --> 00:13:49,240 ChatGPT in part by downloading millions of their articles 202 00:13:49,240 --> 00:13:52,680 without permission. People across China are, of 203 00:13:52,680 --> 00:13:56,560 course, cheering the success of Deep Seek and its founder, who 204 00:13:56,560 --> 00:13:59,720 have made this great achievement in the face of US tech 205 00:13:59,720 --> 00:14:02,600 restrictions. There are all sorts of memes 206 00:14:02,600 --> 00:14:05,680 doing the rounds of the shock waves that sent through Silicon 207 00:14:05,680 --> 00:14:10,080 Valley and Wall Street. As Angela Zhang wrote in the FT, 208 00:14:10,280 --> 00:14:14,640 the inconvenient truth for U.S. policy makers is that strict 209 00:14:14,640 --> 00:14:19,160 export controls forced Chinese tech companies to become more 210 00:14:19,160 --> 00:14:22,720 self reliant, spurring breakthroughs that might not 211 00:14:22,720 --> 00:14:26,680 have occurred otherwise. She says that this episode lays 212 00:14:26,680 --> 00:14:30,640 bare the limits of technology sanctions, which may deliver 213 00:14:30,640 --> 00:14:34,320 short term disruptions, but their impact diminishes over 214 00:14:34,320 --> 00:14:37,400 time as other countries innovate and adapt. 215 00:14:38,400 --> 00:14:42,280 The rise of Deep Seek is a reminder that constraints can 216 00:14:42,280 --> 00:14:46,720 sometimes fuel innovation. With unlimited access to money, 217 00:14:46,840 --> 00:14:51,360 Mehta has so far spent more on GP US than the US government 218 00:14:51,360 --> 00:14:54,160 spent on the entire Manhattan Project. 219 00:14:54,320 --> 00:14:58,960 When adjusted for inflation, Open AI has been burning through 220 00:14:58,960 --> 00:15:04,200 more than $5 billion per year and projected by 2029 they'll be 221 00:15:04,200 --> 00:15:07,200 spending almost $40 billion a year. 222 00:15:07,800 --> 00:15:11,640 The deep seek story, more than anything else, breaks the AI 223 00:15:11,640 --> 00:15:15,120 CapEx narrative, where the biggest firms need to fight for 224 00:15:15,120 --> 00:15:18,560 resources, which are mostly NVIDIA GPU's. 225 00:15:18,920 --> 00:15:22,680 The belief was that the company that could spend the most was 226 00:15:22,680 --> 00:15:28,520 most likely to win the AI race. Jensen Wong of NVIDIA recently 227 00:15:28,520 --> 00:15:32,480 said on an earnings call that he expected the data center 228 00:15:32,480 --> 00:15:37,320 building frenzy to last at least U until the end of the decade. 229 00:15:37,840 --> 00:15:41,440 Up until now, Nvidia's looked like a money printing machine 230 00:15:41,640 --> 00:15:44,800 where they can sell their highest performing chips to the 231 00:15:44,800 --> 00:15:47,200 highest bidder at a massive markup. 232 00:15:47,480 --> 00:15:51,520 The market has had really high expectations of NVIDIA, and 233 00:15:51,640 --> 00:15:55,880 NVIDIA has managed to surpass them both in terms of sales 234 00:15:55,880 --> 00:16:00,640 growth and profitability. Just last week, Sam Altman 235 00:16:00,800 --> 00:16:05,600 announced the Stargate project, where he secured a $500 billion 236 00:16:05,600 --> 00:16:10,400 commitment to building an AI data centre empire thanks to 237 00:16:10,400 --> 00:16:14,640 backing from SoftBank, Oracle and an Abu Dhabi government 238 00:16:14,640 --> 00:16:17,280 fund. You have to imagine that the 239 00:16:17,280 --> 00:16:21,760 news about Deepseek made a few Silicon Valley investors nervous 240 00:16:21,760 --> 00:16:25,440 this week, as they've been piling money into AI at an 241 00:16:25,440 --> 00:16:29,880 unprecedented rate. It's worth no that no one got to 242 00:16:29,880 --> 00:16:34,040 see the prices of the firms most impacted by the announcement on 243 00:16:34,040 --> 00:16:38,280 Monday, like Open AI and Anthropic, as they are all 244 00:16:38,280 --> 00:16:41,160 privately held and not actively traded. 245 00:16:41,720 --> 00:16:45,400 As someone pointed out on Twitter, all of Silicon Valley's 246 00:16:45,400 --> 00:16:50,800 next big things of the last 15 years like NFTS, Web 3, the 247 00:16:50,800 --> 00:16:55,000 metaverse, and virtual reality have been utterly rejected by 248 00:16:55,000 --> 00:16:59,120 the market, and now they're all in on generative AI and 249 00:16:59,120 --> 00:17:03,240 desperately need it to work. Monday will not have been a fun 250 00:17:03,240 --> 00:17:07,000 day in Silicon Valley. Despite the cheerful tweets that 251 00:17:07,000 --> 00:17:11,960 they published, Monday's shock by no means was an indication 252 00:17:11,960 --> 00:17:14,560 that investment in AI is drying up. 253 00:17:14,880 --> 00:17:19,079 This Thursday was announced that SoftBank is in talks to invest 254 00:17:19,079 --> 00:17:24,560 as much as $25 billion into Open AI, and this is on top of the 255 00:17:24,560 --> 00:17:27,160 money they've already committed to Stargate. 256 00:17:27,520 --> 00:17:31,680 According to the FT, SoftBank could spend more than $40 257 00:17:31,680 --> 00:17:34,760 billion on its partnership with Open AI. 258 00:17:35,400 --> 00:17:39,000 Elliott Management, on the other hand, wrote in a recent letter 259 00:17:39,000 --> 00:17:43,000 to investors that the artificial intelligence boom and high 260 00:17:43,000 --> 00:17:47,440 equity market valuation seen today are signs of investors 261 00:17:47,440 --> 00:17:49,840 acting like a crowd of sports betters. 262 00:17:50,120 --> 00:17:55,440 So not everyone is a believer. Artificial General Intelligence, 263 00:17:55,440 --> 00:18:00,520 or AGI, is a type of artificial intelligence that matches or 264 00:18:00,520 --> 00:18:04,800 surpasses human cognitive capabilities across a wide range 265 00:18:04,800 --> 00:18:08,520 of tasks. This contrasts with narrow AI, 266 00:18:08,720 --> 00:18:11,280 which is limited to specific tasks. 267 00:18:11,800 --> 00:18:15,640 For the last few years, Big Tech has been warning us of the 268 00:18:15,640 --> 00:18:20,320 dangers of AGI while working as hard as they can to achieve it. 269 00:18:20,800 --> 00:18:24,520 I've mostly been skeptical about both their warnings and their 270 00:18:24,520 --> 00:18:28,560 claims that they're close to achieving AGII think it's mostly 271 00:18:28,560 --> 00:18:31,960 a marketing scheme where they get a lot of attention by 272 00:18:31,960 --> 00:18:35,080 claiming to be on the verge of discovering something really 273 00:18:35,080 --> 00:18:39,000 dangerous, which might also be really profitable. 274 00:18:39,320 --> 00:18:42,880 We're told that these tech Bros are the only people who can be 275 00:18:42,880 --> 00:18:46,880 trusted with this dangerous technology when, based on the 276 00:18:46,880 --> 00:18:50,120 news, it's not clear that they can even be trusted with their 277 00:18:50,120 --> 00:18:53,880 own system. Now that the AI race is heating 278 00:18:53,880 --> 00:18:58,000 up, it's not obvious that we'll hear a whole lot more about AI 279 00:18:58,000 --> 00:19:02,360 safety, especially if leading models are open source, widely 280 00:19:02,360 --> 00:19:05,440 available, and can be modified by users. 281 00:19:06,320 --> 00:19:10,680 Throughout Monday morning, Deepseek experienced outages 282 00:19:10,840 --> 00:19:14,000 which they said were caused by high traffic, and they 283 00:19:14,000 --> 00:19:16,760 temporarily limited registrations. 284 00:19:17,000 --> 00:19:21,240 Even still, it quickly became the most downloaded free app on 285 00:19:21,240 --> 00:19:25,040 Apple's App Store, overtaking ChatGPT. 286 00:19:25,800 --> 00:19:29,800 As with other Chinese apps, U.S. politicians have been quick to 287 00:19:29,800 --> 00:19:34,440 raise security and privacy concerns, and both the US Navy 288 00:19:34,440 --> 00:19:38,320 and Congress banned employees from downloading the app on 289 00:19:38,320 --> 00:19:41,080 their phones. But luckily they still have 290 00:19:41,080 --> 00:19:45,760 TikTok so they should be fine. At present, there's no reason to 291 00:19:45,760 --> 00:19:48,880 expect Deep Seek to be the long term winner. 292 00:19:49,040 --> 00:19:52,880 Firstly because it's too much of A security risk in countries 293 00:19:53,040 --> 00:19:56,640 that are worried about Chinese influence, but mostly because 294 00:19:56,640 --> 00:20:00,920 it's way too early to know who the overall winner will be or 295 00:20:00,920 --> 00:20:02,960 even how many winners they'll be. 296 00:20:03,240 --> 00:20:06,120 It's quite possible that in different parts of the world, 297 00:20:06,120 --> 00:20:10,200 different AI models will be used because countries just don't 298 00:20:10,200 --> 00:20:15,360 trust each other's technology. If we go back to the late 1990s, 299 00:20:15,520 --> 00:20:19,280 it was very clear at the time that the Internet and e-commerce 300 00:20:19,280 --> 00:20:23,920 would be huge, but most of the big.com companies of the time 301 00:20:23,960 --> 00:20:26,720 have disappeared. If you'd bought all of the 302 00:20:26,720 --> 00:20:31,040 Internet companies in 1999, you would have owned winners like 303 00:20:31,040 --> 00:20:35,240 eBay and Amazon, but you would also have had a bunch of other 304 00:20:35,240 --> 00:20:38,240 companies that have since failed so that you would have been 305 00:20:38,240 --> 00:20:41,560 better off just buying a diversified index fund. 306 00:20:41,800 --> 00:20:44,800 Despite being entirely right about the growth of the 307 00:20:44,800 --> 00:20:48,760 Internet, this is what makes technology investing so 308 00:20:48,760 --> 00:20:52,000 difficult. First mover advantage, which we 309 00:20:52,000 --> 00:20:56,120 all get excited about, often doesn't matter in the long run. 310 00:20:56,400 --> 00:21:00,040 The early search engines all fell into irrelevance when 311 00:21:00,040 --> 00:21:02,640 Google was released. There were companies like 312 00:21:02,640 --> 00:21:06,040 Myspace and Friendster, which were exactly the same as 313 00:21:06,040 --> 00:21:08,840 Facebook but didn't catch on as well. 314 00:21:09,280 --> 00:21:12,760 Going back further, a ton of money went into railroad stocks 315 00:21:12,760 --> 00:21:17,320 in the 1840s and they almost all went bust after building out way 316 00:21:17,320 --> 00:21:21,200 too much infrastructure. There's no reason to think that 317 00:21:21,200 --> 00:21:25,520 any of the leading AI companies today will even be around in a 318 00:21:25,520 --> 00:21:28,200 decade. Most of them are burning money 319 00:21:28,200 --> 00:21:31,280 today with no real path to profitability. 320 00:21:32,040 --> 00:21:36,680 Now, if you wanted to bet on Internet growth in the 1990s but 321 00:21:36,680 --> 00:21:40,280 thought it was safer to bet on infrastructure stocks rather 322 00:21:40,280 --> 00:21:43,880 than the riskier.com companies, you would have invested in 323 00:21:43,880 --> 00:21:48,720 companies like Cisco, Corning, JDS, Uniphase, and Loosened, 324 00:21:48,960 --> 00:21:52,040 none of which turned out to be great investments. 325 00:21:52,440 --> 00:21:55,840 You can be right about the growth of a company or a sector, 326 00:21:56,200 --> 00:21:59,000 but what you pay for a stock still matters. 327 00:21:59,240 --> 00:22:02,720 And back then, some of these companies may have been great, 328 00:22:02,880 --> 00:22:06,000 but you were paying too much for growth that never came. 329 00:22:06,560 --> 00:22:10,280 Sun Microsystems was an infrastructure play in the late 330 00:22:10,280 --> 00:22:12,640 90s. They marketed themselves 331 00:22:12,640 --> 00:22:17,560 asthe.in.com at the time to highlight their central role in 332 00:22:17,560 --> 00:22:21,080 the growth of the Internet. At its peak, the company was 333 00:22:21,080 --> 00:22:26,120 valued at 10 times revenues. The bubble popped in early 2000 334 00:22:26,240 --> 00:22:28,880 and the Internet stocks all got crushed. 335 00:22:29,120 --> 00:22:33,200 Mostofthe.com stocks were a website and a business plan with 336 00:22:33,200 --> 00:22:37,160 no earnings whatsoever, but the infrastructure companies were 337 00:22:37,160 --> 00:22:42,200 real businesses. In 2002 after the crash, Cott 338 00:22:42,200 --> 00:22:46,360 Mcneely, the cofounder of Sun Microsystems, gave an interview 339 00:22:46,360 --> 00:22:50,200 to Business Week where he asked what were investors thinking. 340 00:22:50,640 --> 00:22:55,640 He said at 10 times revenues, to give you a 10 year payback, I 341 00:22:55,640 --> 00:23:00,200 have to pay you 100% of revenues for 10 straight years in 342 00:23:00,200 --> 00:23:03,120 dividends. That assumes that I can get that 343 00:23:03,120 --> 00:23:06,880 by my shareholders. That assumes I have 0 cost of 344 00:23:06,880 --> 00:23:10,520 goods sold, which is very hard for a computer company. 345 00:23:10,880 --> 00:23:15,840 That assumes 0 expenses, which is really hard with 39,000 346 00:23:15,840 --> 00:23:19,720 employees. That assumes I pay no taxes, 347 00:23:19,720 --> 00:23:23,280 which is very hard. And that assumes that you pay no 348 00:23:23,280 --> 00:23:26,720 taxes on your dividends, which is kind of illegal. 349 00:23:27,080 --> 00:23:31,320 And that assumes with zero warranty for the next 10 years, 350 00:23:31,480 --> 00:23:34,480 I can maintain the current revenue run rate. 351 00:23:34,840 --> 00:23:37,840 Now, having thought through that, would any of you like to 352 00:23:37,840 --> 00:23:43,160 buy my stock at $64? Do you realize how ridiculous 353 00:23:43,160 --> 00:23:45,680 those basic assumptions are? He asked. 354 00:23:45,920 --> 00:23:49,240 He went on to say you don't need any transparency. 355 00:23:49,240 --> 00:23:52,600 You don't need any footnotes. What were you thinking? 356 00:23:53,240 --> 00:23:58,240 Now NVIDIA is an amazing and hugely profitable company that 357 00:23:58,240 --> 00:24:02,440 transformed itself from a maker of video game graphic cards to 358 00:24:02,440 --> 00:24:05,600 the biggest company in the world over the last few years. 359 00:24:05,880 --> 00:24:09,840 To quote Jim Reed at Deutsche Bank, it's gone from last 12 360 00:24:09,840 --> 00:24:14,480 month earnings of around $4 billion two years ago to around 361 00:24:14,480 --> 00:24:18,600 $63 billion in the last quarterly release. 362 00:24:18,960 --> 00:24:22,840 He points out that for context, this is around half the total 363 00:24:22,840 --> 00:24:27,440 earnings made by listed stocks in each of the UK, Germany and 364 00:24:27,440 --> 00:24:31,480 France over the last 12 months. And while they're not really 365 00:24:31,480 --> 00:24:36,080 growing, NVIDIA is forecast to continue to see significant 366 00:24:36,080 --> 00:24:40,280 earnings growth. While the stock fell 17% on 367 00:24:40,280 --> 00:24:44,400 Monday, which sounds like a big deal, that just took it back to 368 00:24:44,400 --> 00:24:48,560 its stock price from October, So not such a big deal for long 369 00:24:48,560 --> 00:24:52,040 term investors. The problem is, as Reed points 370 00:24:52,040 --> 00:24:56,400 out, that the AI industry is embryonic, and it's almost 371 00:24:56,400 --> 00:25:00,360 impossible to know how it will develop or what competition 372 00:25:00,360 --> 00:25:04,400 current winners might face, even if you fully believe in its 373 00:25:04,400 --> 00:25:07,080 potential to drive future productivity. 374 00:25:07,640 --> 00:25:11,600 While Sun Microsystems was trading at 10 times revenues in 375 00:25:11,600 --> 00:25:18,280 1999, NVIDIA is trading at 27 times revenues today, meaning it 376 00:25:18,280 --> 00:25:23,480 really is priced for perfection. All of the big tech stocks are 377 00:25:23,520 --> 00:25:27,920 very expensive today, but it's not the same as duringthe.com 378 00:25:27,920 --> 00:25:32,400 bubble in that the big US tech stocks are very profitable and 379 00:25:32,400 --> 00:25:35,840 have been responsible for most of the earnings growth in the 380 00:25:35,840 --> 00:25:40,760 S&P 500, and their growth has vaulted the United States ahead 381 00:25:40,760 --> 00:25:44,440 of the rest of the world. It's not just big tech that's 382 00:25:44,440 --> 00:25:48,040 expensive though. All large cap U.S. stocks are 383 00:25:48,040 --> 00:25:51,760 expensive. Costco, the US retailer, has a 384 00:25:51,760 --> 00:25:56,000 higher PE ratio than Amazon, Microsoft, or Meta. 385 00:25:56,400 --> 00:26:00,720 Investors probably shouldn't be overly optimistic about further 386 00:26:00,720 --> 00:26:04,880 multiple expansion. The fact that Deepseek was able 387 00:26:04,880 --> 00:26:08,840 to build a reasoning model with Nvidia's older, slower chips 388 00:26:09,040 --> 00:26:13,640 suggests that the doors open to other competitors, not just in 389 00:26:13,640 --> 00:26:17,360 building AI models, but also in building chips. 390 00:26:17,600 --> 00:26:22,080 It's worth noting that NVIDIA H100 chips aren't just used in 391 00:26:22,080 --> 00:26:25,400 data centres, they're also used to make handbags. 392 00:26:25,600 --> 00:26:30,360 But the handbags are subject to export control, so do be careful 393 00:26:30,360 --> 00:26:33,480 with that. If you look at some of the big 394 00:26:33,480 --> 00:26:36,520 tech stocks on Monday, when the market was panicking about 395 00:26:36,520 --> 00:26:40,440 Deepseek, you'll see that they didn't really decline very much. 396 00:26:40,440 --> 00:26:43,040 In fact, I think Apple was even up a bit. 397 00:26:43,480 --> 00:26:47,720 Most of the big tech stocks are at or near their highs and deep 398 00:26:47,720 --> 00:26:52,320 seeks efficiency might actually be good for big tech as it means 399 00:26:52,320 --> 00:26:55,520 that they might not need to spend nearly as much on building 400 00:26:55,520 --> 00:26:58,800 their AI models as was previously expected. 401 00:26:59,040 --> 00:27:02,360 And it also means that we might not be in a winner takes all 402 00:27:02,360 --> 00:27:08,360 model like it was for Internet search, even if AI training no 403 00:27:08,360 --> 00:27:12,360 longer requires spending as much on NVIDIA chips as was being 404 00:27:12,360 --> 00:27:15,840 planned for. It's not all bad news for NVIDIA 405 00:27:16,080 --> 00:27:19,960 as reasoning models which break down a problem and solve it step 406 00:27:19,960 --> 00:27:24,280 by step on the fly work quite differently to large language 407 00:27:24,280 --> 00:27:27,160 models. As reasoning models get better 408 00:27:27,320 --> 00:27:29,840 the more processing power you throw at them. 409 00:27:30,280 --> 00:27:33,600 This is a process called inference time compute. 410 00:27:33,880 --> 00:27:37,720 So you don't just need a big data center to train the models 411 00:27:37,720 --> 00:27:41,840 anymore, you also need one to run them, as the more processing 412 00:27:41,840 --> 00:27:45,760 power these models have access to, the smarter they get. 413 00:27:46,080 --> 00:27:49,920 The same chips needed for training AI are also used in 414 00:27:49,920 --> 00:27:54,080 inference data centres now. Deepseek is more efficient at 415 00:27:54,080 --> 00:27:57,600 inference than the other models too, and can use the cheaper 416 00:27:57,600 --> 00:28:00,960 NVIDIA chips, but it's still smarter when it has more 417 00:28:00,960 --> 00:28:04,600 processing power, so there's still a reason to believe that 418 00:28:04,600 --> 00:28:08,440 there's plenty of demand for NVIDIA chips as long as the tech 419 00:28:08,440 --> 00:28:12,560 doesn't change again. As a last point, as this video 420 00:28:12,800 --> 00:28:16,200 might be getting a bit too long, you've possibly heard a lot of 421 00:28:16,200 --> 00:28:20,240 people talking about Jevons Paradox this week, an economic 422 00:28:20,240 --> 00:28:23,240 idea that's mostly applied to energy usage. 423 00:28:23,800 --> 00:28:27,320 Jevin's paradox is that improvements in efficiency lead 424 00:28:27,320 --> 00:28:30,360 to more use of a resource, not less. 425 00:28:30,760 --> 00:28:34,440 A good example would be that as car engines have become more 426 00:28:34,440 --> 00:28:38,840 efficient over time, instead of us using less fuel, we've just 427 00:28:38,840 --> 00:28:42,080 got more powerful and bigger cars than ever before. 428 00:28:42,520 --> 00:28:46,680 Another example would be the better home insulation in Europe 429 00:28:46,720 --> 00:28:50,640 led to people installing much bigger windows in new houses, 430 00:28:50,880 --> 00:28:54,920 offsetting the efficiency gains. The reason people are talking 431 00:28:54,920 --> 00:28:59,560 about Jevin's Paradox and AI is that Satya Nadella, the CEO of 432 00:28:59,560 --> 00:29:04,120 Microsoft, posted on Twitter in reaction to the deep sea model. 433 00:29:04,480 --> 00:29:08,800 Jevin's paradox strikes again. As AI gets more efficient and 434 00:29:08,800 --> 00:29:13,000 accessible, we'll see its use skyrocket, turning it into a 435 00:29:13,000 --> 00:29:15,800 commodity we just can't get enough of. 436 00:29:16,520 --> 00:29:20,600 Now, whether this paradox applies to AI or not really 437 00:29:20,600 --> 00:29:23,240 depends on how much demand there is. 438 00:29:23,480 --> 00:29:27,800 If adoption is being held back by price, then efficiency gains 439 00:29:27,800 --> 00:29:31,640 should lead to greater use. I can see this being the case 440 00:29:31,640 --> 00:29:35,520 for businesses who have commercial uses for AI, and the 441 00:29:35,520 --> 00:29:39,040 cheaper it is, the more commercial uses they might find. 442 00:29:39,400 --> 00:29:43,840 But today, most people who are using AI tools like ChatGPT are 443 00:29:43,840 --> 00:29:47,800 using the free versions, and so their usage isn't really being 444 00:29:47,800 --> 00:29:50,920 held back by the cost because they're not paying anything. 445 00:29:51,200 --> 00:29:54,160 So how much more are they likely to use it? 446 00:29:54,680 --> 00:29:58,840 I guess the question is whether AI will become a service we all 447 00:29:58,840 --> 00:30:03,000 value and pay for, or if it'll end up like e-mail where most 448 00:30:03,000 --> 00:30:05,680 people just want the basic or free version. 449 00:30:06,120 --> 00:30:11,240 A recent US survey found that 80% of American businesses said 450 00:30:11,240 --> 00:30:15,480 that they don't use AI because it's either difficult to use or 451 00:30:15,480 --> 00:30:17,760 irrelevant to their line of business. 452 00:30:18,200 --> 00:30:21,000 That could, of course, all change, but that is the 453 00:30:21,000 --> 00:30:24,920 situation right now and it is nothing to do with the cost of 454 00:30:24,920 --> 00:30:27,960 the tech. We shouldn't overstate the 455 00:30:27,960 --> 00:30:32,160 market's reaction to Deep Seek. On Monday, the NASDAQ fell about 456 00:30:32,160 --> 00:30:37,160 3%, which is a normal bad day and by no means a panic. 457 00:30:37,480 --> 00:30:41,080 The appearance of Deep Seek isn't so much about whether 458 00:30:41,080 --> 00:30:43,640 China will catch up in AI or not. 459 00:30:43,880 --> 00:30:48,600 It's more about how easy it is to catch up and if any of these 460 00:30:48,720 --> 00:30:51,440 AI companies have a defensible mode. 461 00:30:51,760 --> 00:30:55,360 If the models are easily replicated, it'll be difficult 462 00:30:55,360 --> 00:30:58,600 to charge a lot for them and it just becomes a commodity 463 00:30:58,600 --> 00:31:01,840 business, like selling mobile phone contracts. 464 00:31:02,280 --> 00:31:06,520 One way or another, the AI bubble, if there is one, may 465 00:31:06,520 --> 00:31:09,280 still pop, but it didn't pop this week. 466 00:31:09,760 --> 00:31:12,000 Thanks for tuning into this week's podcast. 467 00:31:12,120 --> 00:31:15,240 If you found it interesting, please send a link to a friend 468 00:31:15,440 --> 00:31:19,400 as there's no podcast algorithm and they just grow by word of 469 00:31:19,400 --> 00:31:21,920 mouth. Have a great day and talk to you 470 00:31:21,920 --> 00:31:23,240 again soon. Bye.