1 00:00:00,360 --> 00:00:03,800 The companies at the centre of the AI boom have been busy 2 00:00:03,800 --> 00:00:06,160 investing billions of dollars in each other. 3 00:00:06,360 --> 00:00:09,120 I'm sure you've seen the spaghetti diagrams in the media 4 00:00:09,120 --> 00:00:13,120 recently showing how companies like Open AI are investing in 5 00:00:13,120 --> 00:00:16,600 their chip suppliers, or how chip manufacturers like NVIDIA 6 00:00:16,720 --> 00:00:20,160 are investing in their customers, enabling them to buy 7 00:00:20,160 --> 00:00:23,200 more chips. I first noticed how strange 8 00:00:23,200 --> 00:00:26,640 these deals were back in March when Core Weave, a company who 9 00:00:26,640 --> 00:00:30,840 buys chips from NVIDIA, puts them in data centres and rents 10 00:00:30,840 --> 00:00:35,640 out compute, filed to go public. It's IPO prospectus revealed 11 00:00:35,800 --> 00:00:39,080 that NVIDIA owned about 5% of the company. 12 00:00:39,280 --> 00:00:43,240 When investor interest seemed tepid after a long IPO drought, 13 00:00:43,440 --> 00:00:48,720 NVIDIA offered to anchor the deal at $40 a share with a $250 14 00:00:48,720 --> 00:00:52,520 million order. Bryce Elder described the deal 15 00:00:52,520 --> 00:00:57,160 in the FT at the time as an Aura Boris, an ancient symbol of a 16 00:00:57,160 --> 00:01:00,080 snake or a dragon eating its own tail. 17 00:01:00,360 --> 00:01:03,720 A similar metaphor might be an extension cord plugged into 18 00:01:03,720 --> 00:01:05,840 itself. If you don't know much about 19 00:01:05,840 --> 00:01:09,760 electricity, that might look like a perpetual energy machine. 20 00:01:09,960 --> 00:01:13,760 But trust me, I've tried it out and no matter how you configure 21 00:01:13,760 --> 00:01:16,240 it, it won't power your home appliances. 22 00:01:16,600 --> 00:01:20,240 You just need outside energy to get things going. 23 00:01:20,640 --> 00:01:24,400 That rough speaking is the current state of AI 24 00:01:24,400 --> 00:01:28,400 infrastructure financing. While the sheer number and size 25 00:01:28,400 --> 00:01:32,480 of these deals have convinced some investors that the AI value 26 00:01:32,480 --> 00:01:36,160 chain is developing rapidly, others are concerned by the 27 00:01:36,160 --> 00:01:38,960 circularity. Two companies sit near the 28 00:01:38,960 --> 00:01:43,520 centre of nearly every diagram, Open AI and NVIDIA. 29 00:01:44,040 --> 00:01:48,240 Each is likely trying to ensure that everyone in the ecosystem, 30 00:01:48,240 --> 00:01:52,120 from suppliers to customers to cloud providers, has a vested 31 00:01:52,120 --> 00:01:57,040 interest in their success. Open AI recently announced a 32 00:01:57,040 --> 00:02:00,360 $300 billion cloud infrastructure agreement with 33 00:02:00,360 --> 00:02:04,920 Oracle, a $10 billion custom chip partnership with Broadcom, 34 00:02:05,120 --> 00:02:08,960 and strategic alliances with major memory suppliers. 35 00:02:09,160 --> 00:02:13,920 According to UBS analysts, Open a is memory commitments alone 36 00:02:13,920 --> 00:02:16,800 account for half of the world's current capacity. 37 00:02:17,440 --> 00:02:22,560 NVIDIA, meanwhile, pledged up to $100 billion in investment to 38 00:02:22,560 --> 00:02:27,600 Open AI, who will in turn buy millions of Nvidia's AI graphics 39 00:02:27,600 --> 00:02:30,440 cards. AMD is also in on the game. 40 00:02:30,640 --> 00:02:34,680 Open AI agreed to buy 10s of billions of dollars worth of AMD 41 00:02:34,680 --> 00:02:40,000 chips, and in return, AMD gave Open AI the right to buy 10% of 42 00:02:40,000 --> 00:02:44,960 its stock for one cent per share, contingent on AMD hitting 43 00:02:44,960 --> 00:02:49,280 certain share price targets and Open AI deploying its chips. 44 00:02:49,720 --> 00:02:53,560 The way Matt Levine explained the deal at the time was that if 45 00:02:53,560 --> 00:02:56,480 Open AI announces a big partnership with a public 46 00:02:56,480 --> 00:02:59,600 company, that company's stock price goes up. 47 00:02:59,800 --> 00:03:03,760 So Open AI could just pay for the chips in cash, receive 48 00:03:03,760 --> 00:03:06,800 stock, and when the deal is announced, the stock would 49 00:03:06,800 --> 00:03:10,840 rocket, effectively reimbursing Open AI for its purchase. 50 00:03:11,040 --> 00:03:14,840 Everyone wins. Amazon has its own version of 51 00:03:14,840 --> 00:03:18,080 the loop. It invested more than $8 billion 52 00:03:18,080 --> 00:03:22,080 in Anthropic, the company behind the Clawed chatbot, and in 53 00:03:22,080 --> 00:03:26,040 return, Anthropic committed to using Amazon as its primary 54 00:03:26,040 --> 00:03:29,240 cloud provider. That means training and running 55 00:03:29,240 --> 00:03:33,680 it's models on Amazon's custom AI chips, renting compute from 56 00:03:34,280 --> 00:03:38,840 AWS, and integrating Clawed into Amazon Bedrock, the company's 57 00:03:38,840 --> 00:03:43,040 enterprise AI platform. In effect, Amazon is funding a 58 00:03:43,080 --> 00:03:47,360 company that will use its chips run on its cloud and help sell 59 00:03:47,360 --> 00:03:50,680 services. And now Google is getting in in 60 00:03:50,680 --> 00:03:53,760 the loop, too. And Tropic just announced a deal 61 00:03:53,920 --> 00:03:58,600 to access up to 1,000,000 of Google's TP US, bringing over a 62 00:03:58,600 --> 00:04:02,280 GW of compute capacity online by 2026. 63 00:04:02,520 --> 00:04:05,880 The arrangement is worth 10s of billions of dollars and 64 00:04:05,880 --> 00:04:09,160 positions Google as both the major investor and 65 00:04:09,160 --> 00:04:12,920 infrastructure provider. And Tropic says it chose 66 00:04:13,040 --> 00:04:16,000 Google's chips for their efficiency and performance. 67 00:04:16,200 --> 00:04:20,480 But the deal also reduces it's reliance on NVIDIA and Amazon. 68 00:04:20,800 --> 00:04:24,440 Google has already invested $3 billion in Anthropic. 69 00:04:24,640 --> 00:04:29,120 Amazon has pledged 8 billion. Both companies now provide cloud 70 00:04:29,120 --> 00:04:32,200 services, custom chips and strategic capital. 71 00:04:32,440 --> 00:04:36,920 Anthropic insists that it's just pursuing a multi cloud strategy, 72 00:04:37,160 --> 00:04:40,560 but it's hard to ignore how deeply entangled it has become 73 00:04:40,680 --> 00:04:45,200 with all three of the largest US cloud providers, each of whom 74 00:04:45,200 --> 00:04:48,480 now has a financial interest in its success. 75 00:04:48,880 --> 00:04:52,640 Then there's Elon Musk, who seems to believe that the best 76 00:04:52,640 --> 00:04:56,000 way to build artificial general intelligence is to have his 77 00:04:56,000 --> 00:05:00,560 companies date each other. His AI startup XAI acquire 78 00:05:00,560 --> 00:05:04,800 Twitter or the Everything app, which supplies real time data to 79 00:05:04,800 --> 00:05:09,120 Grog, his chatbot, sometimes referred to as Mecca Hitler. 80 00:05:09,520 --> 00:05:14,560 Tesla, his electric car company, uses the chatbot in it's cars 81 00:05:14,560 --> 00:05:17,920 and possibly it's robots, which are coming next year. 82 00:05:18,440 --> 00:05:23,120 Musk owns a majority stake in XAI, which recently bought 83 00:05:23,120 --> 00:05:26,520 Twitter from him. He owns a minority stake in 84 00:05:26,520 --> 00:05:31,560 Tesla and now wants Tesla shareholders to invest in XAI. 85 00:05:31,920 --> 00:05:36,160 It's not incestuous, exactly, but we'd have to get Errol Musk 86 00:05:36,160 --> 00:05:40,360 to explain why it's OK. The whole thing is starting to 87 00:05:40,360 --> 00:05:44,400 look less like a tech boom and more like a Mobius strip made of 88 00:05:44,400 --> 00:05:48,760 venture capital and electricity. And the electricity part isn't a 89 00:05:48,760 --> 00:05:54,320 metaphor either. McKinsey forecasts $5.2 trillion 90 00:05:54,320 --> 00:05:58,840 in CapEx for chips, data centres and energy over the next five 91 00:05:58,840 --> 00:06:02,120 years alone. Bain says that we'll need to see 92 00:06:02,120 --> 00:06:07,080 $2 trillion in annual revenue from AI companies just to 93 00:06:07,080 --> 00:06:12,440 justify that spending open. AI has about $13 billion in 94 00:06:12,440 --> 00:06:16,480 revenues today and is essentially a money furnace, and 95 00:06:16,480 --> 00:06:19,000 Tropic is a smaller money furnace. 96 00:06:19,200 --> 00:06:23,960 NVIDIA is very profitable, but not $100 billion profitable. 97 00:06:24,200 --> 00:06:27,760 So the question becomes, who's going to pay for all of this? 98 00:06:28,320 --> 00:06:31,760 The interconnected nature of these deals, the reason we need 99 00:06:31,760 --> 00:06:36,040 spaghetti diagrams to understand them, has raised concerns about 100 00:06:36,040 --> 00:06:39,600 circular financing. Companies are investing in each 101 00:06:39,600 --> 00:06:42,800 other, buying each others products and pushing up each 102 00:06:42,800 --> 00:06:46,440 others stock prices. Investors are now asking whether 103 00:06:46,440 --> 00:06:50,880 these interdependencies could pose risks if AI demand or 104 00:06:50,880 --> 00:06:54,440 monetization falls short of investor expectations. 105 00:06:54,920 --> 00:06:57,960 The AI industry's investment structure is starting to 106 00:06:57,960 --> 00:07:01,640 resemble something we've seen before, just not in Silicon 107 00:07:01,640 --> 00:07:04,560 Valley. In post war Japan, large 108 00:07:04,560 --> 00:07:08,400 industrial groups known as kiritsu were built usually 109 00:07:08,400 --> 00:07:12,560 around banks and trading houses, with companies taking stakes in 110 00:07:12,560 --> 00:07:15,440 each other and coordinating supply chains. 111 00:07:15,720 --> 00:07:19,840 South Korea's Chai Ball System followed a similar pattern, but 112 00:07:19,840 --> 00:07:22,440 with families in control rather than banks. 113 00:07:22,880 --> 00:07:26,760 These models weren't about competition, they were, at least 114 00:07:26,760 --> 00:07:29,920 initially, about survival. In capital constrained 115 00:07:29,920 --> 00:07:32,520 economies. It seemed to make sense to have 116 00:07:32,520 --> 00:07:36,320 tight financial relationships with the businesses you relied 117 00:07:36,320 --> 00:07:39,240 upon so that your supply chains were secure. 118 00:07:39,920 --> 00:07:43,760 The kiritsu and chibal models were often criticized for 119 00:07:43,760 --> 00:07:47,680 obscuring financial risk, misallocating capital and 120 00:07:47,680 --> 00:07:52,360 propping up uncompetitive firms. When Japanese asset bubble burst 121 00:07:52,360 --> 00:07:57,040 in the 1990's, the tangled web of cross holdings made it almost 122 00:07:57,040 --> 00:08:02,280 impossible to unwind bad beds. Today's AI giants are by no 123 00:08:02,280 --> 00:08:06,440 means short on capital, but they are assembling these webs of 124 00:08:06,440 --> 00:08:10,200 mutual dependence. The question is whether today's 125 00:08:10,200 --> 00:08:13,600 AI giants are building a similarly fragile structure, 126 00:08:13,680 --> 00:08:17,080 which looks stable from the outside but depends on a 127 00:08:17,080 --> 00:08:20,600 constant influx of new capital to keep the lights on. 128 00:08:21,040 --> 00:08:24,480 When we look at the numbers, they almost seem made-up. 129 00:08:24,720 --> 00:08:28,280 Open AI Stargate project, announced this January at a 130 00:08:28,280 --> 00:08:32,880 White House event, is a $500 billion plan to build 10 131 00:08:32,880 --> 00:08:37,480 gigawatts of AI data center capacity across the US It's 132 00:08:37,480 --> 00:08:41,640 firstly kind of crazy that we're talking about data centers in 133 00:08:41,640 --> 00:08:44,840 terms of gigawatts. According to the US Department 134 00:08:44,840 --> 00:08:49,200 of Energy website, a typical nuclear power plant produces one 135 00:08:49,200 --> 00:08:53,480 GW of power on average, and that's enough electricity to 136 00:08:53,480 --> 00:08:56,760 power a million typical US households. 137 00:08:57,000 --> 00:09:03,000 The typical US household is 2.6, so 10 gigawatts is enough power 138 00:09:03,000 --> 00:09:06,000 for 26,000,000 average Americans. 139 00:09:06,400 --> 00:09:09,520 Open AI isn't just building Stargate. 140 00:09:09,680 --> 00:09:13,880 The Financial Times pointed out that the six GW deal announced 141 00:09:13,880 --> 00:09:18,200 with AMD is enough energy to power Singapore for a year. 142 00:09:18,480 --> 00:09:23,120 But there are other deals too. All in, Open AI has committed to 143 00:09:23,120 --> 00:09:28,240 building 23 gigawatts of new data centre capacity, which they 144 00:09:28,240 --> 00:09:32,360 say will cost well over a trillion dollars to develop and 145 00:09:32,360 --> 00:09:36,440 it seems would require 23 nuclear power stations to power 146 00:09:36,440 --> 00:09:39,760 up. In Texas, where several Stargate 147 00:09:39,760 --> 00:09:44,680 sites are planned, electricity demand is rising so quickly that 148 00:09:44,680 --> 00:09:49,360 some operators are installing on site gas turbines and exploring 149 00:09:49,360 --> 00:09:53,200 nuclear partnerships to avoid waiting for grid hookups. 150 00:09:53,440 --> 00:09:58,200 The XAI data center in South Memphis is running gas turbines 151 00:09:58,200 --> 00:10:02,200 with no emissions controls and no permits, creating enough 152 00:10:02,200 --> 00:10:06,160 pollution that, according to Politico, the area surrounding 153 00:10:06,160 --> 00:10:09,840 it leads Tennessee in asthma hospitalizations. 154 00:10:10,480 --> 00:10:12,960 There's not just one or two of these firms. 155 00:10:13,200 --> 00:10:16,440 All of the big tech firms in the United States and a bunch of 156 00:10:16,440 --> 00:10:20,200 additional firms in China and elsewhere are building out AI 157 00:10:20,200 --> 00:10:23,360 capability. So as I mentioned earlier, 158 00:10:23,480 --> 00:10:28,680 McKinsey now estimates that $5.2 trillion in CapEx will be needed 159 00:10:28,680 --> 00:10:33,400 by 20-30 just to build the data centers required for the 160 00:10:33,400 --> 00:10:37,480 projected AI workloads. On top of that, data centers 161 00:10:37,480 --> 00:10:40,920 powering traditional IT applications are exected to 162 00:10:40,920 --> 00:10:45,600 require $1.5 trillion in capital expenditures, meaning that we're 163 00:10:45,600 --> 00:10:49,920 talking about almost $7 trillion in projected data center 164 00:10:49,920 --> 00:10:52,200 spending over the next five years. 165 00:10:53,000 --> 00:10:56,760 The firms are not generating sufficient revenues to justify 166 00:10:56,760 --> 00:11:00,440 that spending and don't appear to have a path to profitability 167 00:11:00,440 --> 00:11:03,120 planned out yet. For a technology that was 168 00:11:03,120 --> 00:11:06,440 supposed to make scientific breakthroughs, cure diseases, 169 00:11:06,440 --> 00:11:10,120 and maybe even replace human cognition, A surprising amount 170 00:11:10,120 --> 00:11:14,320 of AI output looks like slop, and sometimes worse. 171 00:11:14,760 --> 00:11:19,320 Open AI Sora can generate realistic video, but the most 172 00:11:19,320 --> 00:11:23,360 viral clips so far have been deepfakes of Taylor Swift and 173 00:11:23,360 --> 00:11:25,920 SpongeBob as a character in Breaking Bad. 174 00:11:25,920 --> 00:11:30,080 Yo Sponge, This stuff looks extra crystalline, like 175 00:11:30,200 --> 00:11:34,240 restaurant quality. Then there's Elon Musk's XAI, 176 00:11:34,400 --> 00:11:38,040 which has built an anime girlfriend chatbot, which many 177 00:11:38,040 --> 00:11:41,600 feel is an improvement over the Hitler one, which should 178 00:11:41,600 --> 00:11:45,040 hopefully keep basement dwellers occupied for the foreseeable 179 00:11:45,040 --> 00:11:47,920 future. There's also a cartoony red 180 00:11:47,920 --> 00:11:51,040 panda version if that's what you're into. 181 00:11:51,880 --> 00:11:56,480 As easy as it is to make fun of this, there are many less widely 182 00:11:56,480 --> 00:12:00,160 discussed breakthroughs. The 2024 Nobel Prize for 183 00:12:00,160 --> 00:12:04,440 Chemistry went to two Google DeepMind researchers for their 184 00:12:04,440 --> 00:12:08,800 pioneering work on AI powered protein folding, which promises 185 00:12:08,800 --> 00:12:12,880 to expedite drug discovery and development and is already being 186 00:12:12,880 --> 00:12:15,760 used to combat cancers and other diseases. 187 00:12:16,160 --> 00:12:20,000 A number of my viewers think of me as being anti tech and anti 188 00:12:20,160 --> 00:12:23,800 AI as I've made fun of many of the more ridiculous claims out 189 00:12:23,800 --> 00:12:26,600 of Silicon Valley. And there are a lot of them to 190 00:12:26,600 --> 00:12:30,840 keep up with, like the Hyperloop, the metaverse, AI 191 00:12:30,840 --> 00:12:34,240 enhanced water bottles, the general usefulness of the 192 00:12:34,240 --> 00:12:38,240 blockchain, and passing off short term office rentals with 193 00:12:38,240 --> 00:12:41,520 free beer. As a tech business, there are 194 00:12:41,600 --> 00:12:46,000 plenty of uses for AI which don't involve generating slop, 195 00:12:46,120 --> 00:12:51,560 but people do seem to love slop. Open AI is not profitable. 196 00:12:51,680 --> 00:12:55,200 It's spending much more money than it brings in in revenue and 197 00:12:55,200 --> 00:12:59,560 is doing so at a pace that would give most CFOs post traumatic 198 00:12:59,560 --> 00:13:03,080 Stress disorder To fund its infrastructure build out. 199 00:13:03,240 --> 00:13:07,440 The company has secured a $4 billion revolving credit line 200 00:13:07,560 --> 00:13:11,720 from a consortium of banks. This is very unusual, unusual. 201 00:13:11,840 --> 00:13:15,480 Historically high grow tech firms raised capital through 202 00:13:15,480 --> 00:13:19,640 equity, especially if they were burning cash, as lenders like to 203 00:13:19,640 --> 00:13:24,200 see predictable earnings. The shift from equity to debt 204 00:13:24,320 --> 00:13:27,880 and from public listing to private investment is happening 205 00:13:27,880 --> 00:13:31,640 across the sector, where data center providers are borrowing 206 00:13:31,640 --> 00:13:36,000 against assets like racks of GPU's, which might quickly 207 00:13:36,000 --> 00:13:40,360 become obsolete. This creates a strange dynamic. 208 00:13:40,560 --> 00:13:43,880 The companies building the infrastructure are borrowing to 209 00:13:43,880 --> 00:13:48,120 serve customers who are also borrowing or being subsidized by 210 00:13:48,120 --> 00:13:50,960 their investors. The whole system appears to be 211 00:13:50,960 --> 00:13:55,880 leveraged on optimism for now. The money is flowing, and for 212 00:13:55,880 --> 00:13:59,560 users, as I argued in my Blitz scaling video from a few years 213 00:13:59,560 --> 00:14:03,960 ago, it probably makes sense to make the most of these expensive 214 00:14:04,120 --> 00:14:06,880 AI tools that we're currently getting for free. 215 00:14:07,120 --> 00:14:10,640 It's not clear how long that can last for though, or who'll be 216 00:14:10,640 --> 00:14:13,920 left holding the bag if AI providers can't flip to 217 00:14:13,920 --> 00:14:18,040 profitability. The GPU rental market is already 218 00:14:18,040 --> 00:14:21,680 showing early signs of stress and the build out is only 219 00:14:21,680 --> 00:14:25,080 getting going. According to FT Alphaville, the 220 00:14:25,080 --> 00:14:30,720 price to rent Nvidia's B200 chip has dropped from $3.20 an hour 221 00:14:30,880 --> 00:14:34,920 to $2.80 per hour in just a few months. 222 00:14:35,200 --> 00:14:39,480 Older chips like the A-100 are now available for as little as 223 00:14:39,480 --> 00:14:43,280 $0.40 per hour. That's below break even for many 224 00:14:43,280 --> 00:14:46,360 operators, even under ideal conditions. 225 00:14:46,920 --> 00:14:50,560 They calculate in the article that a cluster of eight chips, 226 00:14:50,560 --> 00:14:54,840 which would have cost around $200,000 five years ago and has 227 00:14:54,840 --> 00:14:58,600 a five year useful life, would need to have generated about 228 00:14:58,600 --> 00:15:02,640 $4.00 an hour in rental fees just to break even. 229 00:15:02,920 --> 00:15:09,080 Back in 2020, the average rental price for an A-100 was $2.40 an 230 00:15:09,080 --> 00:15:11,320 hour. That's now fallen to around 231 00:15:11,360 --> 00:15:14,960 $1.65 per hour. And to make it worse, the 232 00:15:14,960 --> 00:15:18,880 average is being skewed by high hyperscalers who are continuing 233 00:15:18,880 --> 00:15:22,520 to charge more than $4.00 when their competitors go as low as 234 00:15:22,520 --> 00:15:25,880 $0.40. If demand for all of this 235 00:15:25,880 --> 00:15:29,320 infrastructure doesn't materialize, it could become 236 00:15:29,320 --> 00:15:32,400 stranded. Data centres built for five 237 00:15:32,400 --> 00:15:35,240 years of peak usage might sit half empty. 238 00:15:35,520 --> 00:15:40,360 The FTP suggests that many pandemic era GPU's will end up 239 00:15:40,360 --> 00:15:43,720 in liquidation, never having earned back their cost. 240 00:15:44,200 --> 00:15:48,840 There is a precedent to this. Telecom firms in the early 2000s 241 00:15:48,840 --> 00:15:52,480 built out fiber optic networks that were never used. 242 00:15:52,680 --> 00:15:56,880 Railways in the 19th century similarly laid track to nowhere, 243 00:15:57,040 --> 00:16:01,480 much of which was later removed. The AI industry is now laying 244 00:16:01,480 --> 00:16:05,720 down gigawatts of compute, betting that someone will not 245 00:16:05,720 --> 00:16:09,160 just show up to use it, but actually pay to use it. 246 00:16:09,400 --> 00:16:13,080 If they don't, the fallout won't be limited to a few start-ups. 247 00:16:13,240 --> 00:16:17,040 It'll hit lenders, landlords, and the Public Utilities that 248 00:16:17,040 --> 00:16:20,080 signed up to support the boom without necessarily 249 00:16:20,080 --> 00:16:22,480 understanding the bet that they were making. 250 00:16:22,760 --> 00:16:27,600 Nvidia's stock market valuation is based on the idea that demand 251 00:16:27,600 --> 00:16:31,800 for its chips is massive and will keep rising, not just this 252 00:16:31,800 --> 00:16:33,800 year, but for the foreseeable future. 253 00:16:34,080 --> 00:16:38,040 The question we have to ask is how much of that demand is real 254 00:16:38,160 --> 00:16:41,520 and how much is driven by Nvidia's investments in other 255 00:16:41,520 --> 00:16:45,160 companies. Open AI is buying and renting 256 00:16:45,200 --> 00:16:49,520 billions of dollars worth of NVIDIA chips, Nvidia's investing 257 00:16:49,520 --> 00:16:54,280 in Open AI, Cor Weave rents NVIDIA chips to Open AI, and 258 00:16:54,400 --> 00:16:56,880 NVIDIA owns a stake in Cor Weave. 259 00:16:57,200 --> 00:17:01,080 The same dollars are circulating through the system, possibly 260 00:17:01,080 --> 00:17:04,359 inflating purchase orders and revenue projections. 261 00:17:04,599 --> 00:17:07,920 It's hard to tell where the demand ends and the subsidy 262 00:17:07,920 --> 00:17:10,720 begins. The circularity makes it 263 00:17:10,720 --> 00:17:14,319 difficult to assess the quality of revenues, and that's why 264 00:17:14,319 --> 00:17:17,560 there are so many people asking if we're in an AI bub. 265 00:17:18,000 --> 00:17:21,280 If Nvidia's biggest customers are also it's investment 266 00:17:21,280 --> 00:17:25,040 targets, and those customers are using Nvidia's money to buy 267 00:17:25,040 --> 00:17:29,320 Nvidia's products, then the margins may not be quite what 268 00:17:29,320 --> 00:17:32,360 they seem. There's also the question of how 269 00:17:32,360 --> 00:17:34,880 well this infrastructure is being used. 270 00:17:35,200 --> 00:17:39,960 Open AI claims to have 700 million weekly users, but only 271 00:17:39,960 --> 00:17:43,920 5% are paying customers. Most of the revenue in the 272 00:17:43,920 --> 00:17:47,800 sector comes from enterprise contracts, not individual 273 00:17:47,800 --> 00:17:51,000 subscriptions. And even among business users, 274 00:17:51,160 --> 00:17:54,440 the success rate of AI pilot projects is low. 275 00:17:54,720 --> 00:17:57,640 McKinsey puts it at less than 15%. 276 00:17:58,080 --> 00:18:01,560 We're not seeing the mass AI driven layoffs that many were 277 00:18:01,560 --> 00:18:05,400 predicting a few years ago. Labor data shows no clear 278 00:18:05,400 --> 00:18:09,280 relationship between AI deployment and trends in 279 00:18:09,360 --> 00:18:12,960 employment, other than for freelance graphic designers and 280 00:18:12,960 --> 00:18:16,480 copywriters, who have seen sharp declines since the arrival of 281 00:18:16,480 --> 00:18:20,720 ChatGPT and some junior coding jobs, which have been in 282 00:18:20,720 --> 00:18:24,760 decline. Not long ago, the complaint was 283 00:18:24,920 --> 00:18:27,800 that American companies were no longer investing. 284 00:18:27,960 --> 00:18:31,880 They were hoarding cash or using it to buy back stock and just 285 00:18:31,880 --> 00:18:34,640 avoiding risk. Now, the complaint is that 286 00:18:34,640 --> 00:18:38,200 they're investing too much and possibly in the wrong things. 287 00:18:38,880 --> 00:18:43,360 The circular deals are big, but they're not overwhelmingly so. 288 00:18:43,640 --> 00:18:47,600 The open AI NVIDIA deal, as an example, should account for 289 00:18:47,600 --> 00:18:53,640 around 13% of Nvidia's expected 2026 revenue, according to UPS. 290 00:18:53,880 --> 00:18:57,640 And that's assuming the full GW deployment goes ahead. 291 00:18:58,000 --> 00:19:01,800 That would mean 50 to $60 billion in total capital 292 00:19:01,800 --> 00:19:05,880 investment, with NVIDIA receiving $35 billion of it 293 00:19:05,880 --> 00:19:08,600 back. NVIDIA says that it might 294 00:19:08,600 --> 00:19:13,600 reinvest $10 billion into open AI, but only if monetization 295 00:19:13,600 --> 00:19:16,640 keeps pace. That's a performance based 296 00:19:16,640 --> 00:19:19,720 approach, which is smart and gives plenty of room to back 297 00:19:19,720 --> 00:19:22,480 out. The financial health of the big 298 00:19:22,480 --> 00:19:26,880 players is solid too. The mega cap US tech firms are 299 00:19:26,880 --> 00:19:31,400 expected to generate over $200 billion in free cash flow next 300 00:19:31,400 --> 00:19:35,560 year alone, even after CapEx. That's enough to fund the 301 00:19:35,560 --> 00:19:39,480 infrastructure build out without leaning too hard on debt or 302 00:19:39,480 --> 00:19:43,400 requiring new financing. The balance sheets are strong 303 00:19:43,520 --> 00:19:46,600 and the earnings are real. This isn't the same as the 304 00:19:46,680 --> 00:19:50,760 telecom bubble. Valuations are elevated, but 305 00:19:50,760 --> 00:19:55,000 once again not absurd. In the late 1990s, Internet 306 00:19:55,000 --> 00:19:58,040 stocks traded at 60 times forward earnings. 307 00:19:58,280 --> 00:20:03,080 Today's AI leaders are closer to 35 times, and the ones everyone 308 00:20:03,080 --> 00:20:06,080 is excited about actually have earnings. 309 00:20:06,360 --> 00:20:09,160 The market isn't pricing in infinite growth. 310 00:20:09,320 --> 00:20:13,400 It's pricing in a bet that AI will be big and that the 311 00:20:13,400 --> 00:20:15,720 companies building it will make a lot of money. 312 00:20:16,440 --> 00:20:19,440 Now, that bed might not pay off immediately. 313 00:20:19,680 --> 00:20:24,520 Monetization so far has been slow, Adoption is uneven, and 314 00:20:24,520 --> 00:20:28,320 some parts of the value chain, especially cloud renters and AI 315 00:20:28,320 --> 00:20:30,840 labs, are more exposed than others. 316 00:20:31,040 --> 00:20:34,920 But the fundamentals are better than they were in past cycles, 317 00:20:35,120 --> 00:20:37,880 and the investment strategies are more cautious. 318 00:20:38,560 --> 00:20:41,800 There's one constraint that doesn't show up on balance 319 00:20:41,800 --> 00:20:43,480 sheets. Electricity. 320 00:20:43,840 --> 00:20:46,880 Open AI. Stargate project alone will 321 00:20:46,880 --> 00:20:50,760 require 10 gigawatts of power, which as I mentioned is around 322 00:20:50,760 --> 00:20:55,480 10 nuclear power stations. The full build out just for open 323 00:20:55,600 --> 00:20:59,720 AI, not for the others, is expected to need 23. 324 00:21:00,000 --> 00:21:04,200 For context, the last new nuclear reactor in the United 325 00:21:04,200 --> 00:21:08,120 States took more than a decade to complete and came online in 326 00:21:08,120 --> 00:21:11,480 20 24. There are no new nuclear sites 327 00:21:11,480 --> 00:21:15,600 currently under construction. Permitting for solar and wind 328 00:21:15,720 --> 00:21:19,240 has been tightened and tariffs have raised costs for those 329 00:21:19,240 --> 00:21:22,800 power sources. Even fast tracked projects face 330 00:21:22,800 --> 00:21:26,280 multi year delays. Some developers are installing 331 00:21:26,280 --> 00:21:30,440 gas turbines on site just to avoid waiting for grid 332 00:21:30,440 --> 00:21:33,400 connections. The chips might arrive on 333 00:21:33,400 --> 00:21:36,480 schedule, the electricity probably won't. 334 00:21:37,000 --> 00:21:40,480 High private market valuations like we're seeing for firms like 335 00:21:40,480 --> 00:21:45,840 Open, AIXAI, and Anthropic only makes sense if one of the men's 336 00:21:45,840 --> 00:21:49,200 up dominating the space. That's what tech investors are 337 00:21:49,200 --> 00:21:53,120 expecting as that's what's happened in the past with Google 338 00:21:53,240 --> 00:21:57,560 dominating search, Amazon dominating e-commerce, Meta 339 00:21:57,560 --> 00:22:01,680 dominating the metaverse and those glasses that Zuckerberg. 340 00:22:01,760 --> 00:22:04,480 It works. If AI turns out to be a winner 341 00:22:04,480 --> 00:22:08,560 take all market then paying up for the winner could be a great 342 00:22:08,560 --> 00:22:12,800 investment, but owning the mall might not as a bunch of them 343 00:22:12,800 --> 00:22:15,400 could fail. If you invested in all of the 344 00:22:15,400 --> 00:22:19,120 big search engines in the mid 1990s it wouldn't have worked 345 00:22:19,120 --> 00:22:23,400 out for you as Google arrived late but then dominated search. 346 00:22:23,760 --> 00:22:28,560 The DeepSeek moment earlier this year caused a bit of panic in AI 347 00:22:28,840 --> 00:22:32,120 as it showed that models can possibly be replicated, located 348 00:22:32,120 --> 00:22:35,760 quickly and cheaply. Elon Musk's rapid deployment of 349 00:22:35,760 --> 00:22:39,680 Grok showed the same thing. These systems might require a 350 00:22:39,680 --> 00:22:43,800 lot of really expensive R&D, but they may not be very hard to 351 00:22:43,800 --> 00:22:46,200 copy. And if the models are all 352 00:22:46,200 --> 00:22:50,040 roughly the same, then the market may not reward anyone 353 00:22:50,040 --> 00:22:52,440 player. Instead of a big winner and a 354 00:22:52,440 --> 00:22:57,160 monopoly, we might see a very competitive market for AI tools 355 00:22:57,240 --> 00:22:59,440 where none have any pricing power. 356 00:23:00,000 --> 00:23:03,880 On top of all of that, there's the question of who profits. 357 00:23:04,040 --> 00:23:07,040 It might not be the model builders, it might be the 358 00:23:07,040 --> 00:23:10,960 businesses that use the models. AI could end up boosting 359 00:23:10,960 --> 00:23:14,640 productivity across the economy, while the labs themselves 360 00:23:14,640 --> 00:23:19,000 struggle to monetize. So while this might be a bubble, 361 00:23:19,200 --> 00:23:22,600 the fundamentals of the biggest companies involved are stronger 362 00:23:22,600 --> 00:23:25,600 than in past bubbles. The investment strategies are 363 00:23:25,600 --> 00:23:28,800 also more cautious, where a lot of the big deals that have been 364 00:23:28,800 --> 00:23:33,440 announced leave lots of room for backing out, but the outcome is 365 00:23:33,440 --> 00:23:37,120 still really uncertain. Someone has to pay for all of 366 00:23:37,120 --> 00:23:40,880 this and it's not clear who wins or if anyone does. 367 00:23:41,600 --> 00:23:43,720 Thanks for tuning into this week's podcast. 368 00:23:43,840 --> 00:23:46,080 If you found it interesting, do subscribe. 369 00:23:46,240 --> 00:23:49,480 And I'd really appreciate it if you forwarded a link to a friend 370 00:23:49,680 --> 00:23:52,600 as there's not really an algorithm for podcasts, they 371 00:23:52,600 --> 00:23:55,480 just grow based on word of mouth recommendations. 372 00:23:55,720 --> 00:23:58,280 Have a great week and talk to you again soon. 373 00:23:58,400 --> 00:23:58,720 Bye.