1 00:00:02,480 --> 00:00:08,320 Speaker 1: More media. Hello and welcome to this week's Bear Offline Monologue. 2 00:00:08,360 --> 00:00:18,759 Speaker 1: I'm your host ed Ze Trump, and today I'm talking 3 00:00:18,800 --> 00:00:22,040 Speaker 1: about the great parasite of AI and how it's making 4 00:00:22,079 --> 00:00:24,599 Speaker 1: everything more expensive. And to be clear, before we go 5 00:00:24,600 --> 00:00:27,360 Speaker 1: any further, what you're seeing in these remarkable sales from 6 00:00:27,520 --> 00:00:29,920 Speaker 1: video and all the RAM companies in Broadcommon and all 7 00:00:30,000 --> 00:00:33,080 Speaker 1: them isn't the measure of industry demand, but of speculative 8 00:00:33,120 --> 00:00:36,040 Speaker 1: purchasing based on the vegas the projections. Because the actual 9 00:00:36,080 --> 00:00:38,919 Speaker 1: demand for AI, the actual revenue generated by this industry 10 00:00:39,040 --> 00:00:42,080 Speaker 1: is kind of pathetic. And honestly, we don't really have 11 00:00:42,159 --> 00:00:45,320 Speaker 1: any breason to build these data centers. Nobody has trouble 12 00:00:45,360 --> 00:00:48,520 Speaker 1: accessing these services. I mean, it's just being done because 13 00:00:48,520 --> 00:00:50,199 Speaker 1: they've got nothing else to do. And I know that 14 00:00:50,320 --> 00:00:53,880 Speaker 1: sounds kind of flippant, but really is what's happening to 15 00:00:53,880 --> 00:00:56,279 Speaker 1: make that's worse. The more the bubble inflates, the more 16 00:00:56,280 --> 00:00:59,400 Speaker 1: expensive everything gets, the more data centers that get funded, 17 00:00:59,560 --> 00:01:02,360 Speaker 1: the more that are ordered, The more resources this takes, 18 00:01:02,360 --> 00:01:04,360 Speaker 1: and the more the RAM companies can crank up the 19 00:01:04,400 --> 00:01:08,920 Speaker 1: prices and fuck over the customer. You see, aigpus is 20 00:01:08,920 --> 00:01:11,720 Speaker 1: a special kind of memory, high bandwidth memory that takes 21 00:01:11,800 --> 00:01:13,920 Speaker 1: up more of the limited amount of fab space where 22 00:01:14,120 --> 00:01:17,720 Speaker 1: eske Heinex, Micron and Samsung build their RAM. These are 23 00:01:17,760 --> 00:01:21,240 Speaker 1: also the only three companies that really can build HBM. 24 00:01:21,360 --> 00:01:24,480 Speaker 1: It's a triopoly, I guess you'd call them. And because 25 00:01:24,560 --> 00:01:27,760 Speaker 1: in Nvidia broad common, other chip manufacturers quite literally can't 26 00:01:27,800 --> 00:01:31,160 Speaker 1: build the GPUs that hyperscalers crave without them, their memory 27 00:01:31,160 --> 00:01:35,360 Speaker 1: companies can charge effectively whatever they want. Aigpus have become 28 00:01:35,400 --> 00:01:38,800 Speaker 1: a massive amount of guaranteed revenue. Is the AI bubble inflates, 29 00:01:39,120 --> 00:01:41,600 Speaker 1: leaving less fab space for the regular RAM, creating a 30 00:01:41,640 --> 00:01:44,080 Speaker 1: supply shortage, or put another way, a reason for the 31 00:01:44,120 --> 00:01:47,039 Speaker 1: memory companies to bump up the prices aggressively. And yeah, 32 00:01:47,080 --> 00:01:49,160 Speaker 1: they're allowed to do that. There's nobody to stop them. 33 00:01:49,200 --> 00:01:51,559 Speaker 1: It's not like we have regulatory authorities that would bother 34 00:01:51,920 --> 00:01:54,360 Speaker 1: I should also add that in Vidia is one of 35 00:01:54,400 --> 00:01:57,120 Speaker 1: the largest, if not the largest HBM buyer in the world. 36 00:01:57,240 --> 00:01:59,880 Speaker 1: Fifty five to sixty five percent of all high bandwidth 37 00:01:59,920 --> 00:02:03,880 Speaker 1: memory sales go to Nvidia, and this makes it actually 38 00:02:03,880 --> 00:02:06,400 Speaker 1: a victim of this same inflation because memory is getting 39 00:02:06,400 --> 00:02:09,920 Speaker 1: more expensive across the board, even for high bandwidth memory. 40 00:02:10,600 --> 00:02:12,960 Speaker 1: Gold and Sachs reports that high bandwidth memory is on 41 00:02:13,040 --> 00:02:15,320 Speaker 1: course to increase by as much as ninety percent year 42 00:02:15,360 --> 00:02:18,160 Speaker 1: over year in twenty twenty seven. A linear increase in 43 00:02:18,240 --> 00:02:21,359 Speaker 1: costs for every single in Vidio GPU and every single 44 00:02:21,440 --> 00:02:24,760 Speaker 1: new one that one of the four hyperscalers OATHS insists 45 00:02:24,800 --> 00:02:27,000 Speaker 1: on pre ordering only serves to inflate a bubble for 46 00:02:27,000 --> 00:02:30,120 Speaker 1: effectively no reason. They're not making a profit. Shit, they 47 00:02:30,120 --> 00:02:33,000 Speaker 1: won't even talk about revenues. Maybe they're not really making 48 00:02:33,040 --> 00:02:36,080 Speaker 1: that much outside of open AI and anthropics compute, like 49 00:02:36,120 --> 00:02:40,280 Speaker 1: I've been saying. But good lord, good lord, good golly, 50 00:02:40,360 --> 00:02:42,840 Speaker 1: Miss Molly, do these data centers take up a shit 51 00:02:42,880 --> 00:02:45,840 Speaker 1: ton of RAM? Analysts estimate that by the time that 52 00:02:45,840 --> 00:02:49,040 Speaker 1: in Video ships its first next generation Vera Roubin GPUs, 53 00:02:49,120 --> 00:02:51,720 Speaker 1: it'll cost eighteen dollars and forty cents per gigabyte of 54 00:02:51,800 --> 00:02:54,840 Speaker 1: high bandwidth memory sold two hundred and eighty eight gigabytes 55 00:02:54,880 --> 00:02:57,040 Speaker 1: at a time. In fact, let me give you some 56 00:02:57,040 --> 00:02:59,160 Speaker 1: scale as to how much RAM and a data center 57 00:02:59,200 --> 00:03:02,639 Speaker 1: actually needs, because it's an astonishing amount. The current generation 58 00:03:02,760 --> 00:03:05,680 Speaker 1: GB three hundred racks the vast seventy two GPUs servers 59 00:03:05,760 --> 00:03:07,720 Speaker 1: that fill out modern data centers, and they have a 60 00:03:07,720 --> 00:03:11,440 Speaker 1: CPU for every two GPUs have about twenty point seven 61 00:03:11,919 --> 00:03:17,480 Speaker 1: terabytes of high bandwidth memory and seventeen terabytes of LPDDR 62 00:03:17,639 --> 00:03:20,360 Speaker 1: five X RAM. And by the way, that's the RAM 63 00:03:20,360 --> 00:03:22,919 Speaker 1: that you find in smartphones and mobile devices, which is 64 00:03:22,960 --> 00:03:26,240 Speaker 1: also contributing to the supply chain shortages and driving up 65 00:03:26,280 --> 00:03:29,160 Speaker 1: the price. Anyway, a giga what data center is about 66 00:03:29,320 --> 00:03:32,520 Speaker 1: four nine hundred and thirty three of these raks, working 67 00:03:32,520 --> 00:03:34,920 Speaker 1: out to about one point ninety four billion dollars in 68 00:03:35,000 --> 00:03:38,320 Speaker 1: high bandwidth memory alone for a single data center, a 69 00:03:38,400 --> 00:03:41,120 Speaker 1: number that's certain to increase across the board, even for 70 00:03:41,160 --> 00:03:44,040 Speaker 1: current generation gear because the price of high bandwidth memory 71 00:03:44,080 --> 00:03:47,040 Speaker 1: is going up. I haven't even included the many separate 72 00:03:47,080 --> 00:03:50,720 Speaker 1: computers and different machines inside the data center that require RAM, 73 00:03:50,760 --> 00:03:54,080 Speaker 1: two entire racks of power management gear that themselves require 74 00:03:54,200 --> 00:03:58,480 Speaker 1: RAM or HVAC systems alone. They're effectively servers unto themselves, 75 00:03:59,360 --> 00:04:02,240 Speaker 1: and every the data center is basically a giant warehouse 76 00:04:02,280 --> 00:04:05,240 Speaker 1: of different RAM dependent devices and machines, each controlled by 77 00:04:05,280 --> 00:04:08,200 Speaker 1: other RAM dependent machines or to do one thing. AI 78 00:04:08,320 --> 00:04:11,040 Speaker 1: that's yet to prove it even has a meaningful business 79 00:04:11,080 --> 00:04:13,720 Speaker 1: model or one that I don't know makes more money 80 00:04:13,720 --> 00:04:17,240 Speaker 1: than it loses. And that high bandwidth memory, by the way, 81 00:04:17,279 --> 00:04:19,680 Speaker 1: is also not particularly useful for anything else, at least 82 00:04:19,680 --> 00:04:21,960 Speaker 1: not at the scale that's being built, meaning that once 83 00:04:22,000 --> 00:04:24,440 Speaker 1: there's a noted Kapex pullback, somebody is going to be 84 00:04:24,480 --> 00:04:26,080 Speaker 1: left with a bag, by which I mean somebody is 85 00:04:26,120 --> 00:04:28,640 Speaker 1: going to either have too much high bandwidth memory or 86 00:04:28,640 --> 00:04:30,560 Speaker 1: gear that they can't sell with it put on the 87 00:04:30,600 --> 00:04:33,680 Speaker 1: side that won't be useful or really sellable to someone else. 88 00:04:34,320 --> 00:04:37,920 Speaker 1: And I must be clear how much more of this 89 00:04:38,480 --> 00:04:40,920 Speaker 1: they are selling than usual. Back in twenty twenty two, 90 00:04:41,160 --> 00:04:45,440 Speaker 1: they sold about six hundred million gigabytes of high bandwidth memory, 91 00:04:45,600 --> 00:04:48,159 Speaker 1: so sounds like a lot until you realize they're expecting 92 00:04:48,200 --> 00:04:51,000 Speaker 1: to sell seven trillion or more of it in twenty 93 00:04:51,040 --> 00:04:55,280 Speaker 1: twenty six. That's that's a problem. That's a problem because 94 00:04:55,279 --> 00:04:57,040 Speaker 1: the more of it they sell, the more expensive it gets. 95 00:04:57,080 --> 00:05:00,599 Speaker 1: But that's they're going to have years worth this crap 96 00:05:00,720 --> 00:05:03,320 Speaker 1: just either sitting around or being attached to in video 97 00:05:03,400 --> 00:05:06,160 Speaker 1: GPUs that can't sell to anyone. And the longer it 98 00:05:06,200 --> 00:05:09,360 Speaker 1: takes for Capex to retract, the more expensive everything becomes. 99 00:05:09,400 --> 00:05:12,360 Speaker 1: As a result, the more GPUs that get sold, the 100 00:05:12,360 --> 00:05:14,960 Speaker 1: more capacity that gets put towards high bandwidth memory, and 101 00:05:15,000 --> 00:05:18,240 Speaker 1: the more that Micron, Skhienex and Samson could charge for it, 102 00:05:18,480 --> 00:05:21,480 Speaker 1: which makes it more expensive to buy aigpus, which increases 103 00:05:21,520 --> 00:05:24,320 Speaker 1: the amount that hyperscalats are spending on AI capex for 104 00:05:24,360 --> 00:05:27,640 Speaker 1: effectively the same amount of gear. The longer that Hyperscale 105 00:05:27,640 --> 00:05:29,960 Speaker 1: has sustain this pace, the larger the return needs to be. 106 00:05:30,120 --> 00:05:32,080 Speaker 1: And at this point none of them have disclosed their 107 00:05:32,080 --> 00:05:35,080 Speaker 1: actual AI revenues, which heavily suggests that there's yet to 108 00:05:35,120 --> 00:05:37,440 Speaker 1: be even a single dollar of profit, and the actual 109 00:05:37,839 --> 00:05:41,120 Speaker 1: quarterly revenues are pretty piss poor. Yet the more they commit, 110 00:05:41,160 --> 00:05:43,600 Speaker 1: the more committed they have to be. Pulling back at 111 00:05:43,640 --> 00:05:45,640 Speaker 1: this point will prove to the markets that all of 112 00:05:45,640 --> 00:05:49,760 Speaker 1: the hyperscalats have committed far too much to too much capacity. 113 00:05:49,880 --> 00:05:52,760 Speaker 1: Yet not pulling back means that hyperscalers will continue to 114 00:05:52,800 --> 00:05:55,400 Speaker 1: turn their free cash flows negative in pursuit of some 115 00:05:55,640 --> 00:05:59,359 Speaker 1: indeterminate goal. It's a vicious cycle, made worse by the 116 00:05:59,360 --> 00:06:02,240 Speaker 1: fact that every very spin of the Capex wheel increases 117 00:06:02,279 --> 00:06:05,279 Speaker 1: the price of just about every consumer electronic in the world, 118 00:06:05,520 --> 00:06:08,760 Speaker 1: creating a market wide inflation for what amounts to a 119 00:06:08,800 --> 00:06:11,640 Speaker 1: speculative asset bubble, and will leave them sitting with a 120 00:06:11,640 --> 00:06:14,200 Speaker 1: bunch of GPUs they don't really need now and certainly 121 00:06:14,240 --> 00:06:17,479 Speaker 1: won't in the future. What great works are these data 122 00:06:17,560 --> 00:06:21,119 Speaker 1: centers in search for? What possible achievement can you point 123 00:06:21,120 --> 00:06:24,360 Speaker 1: to at this time that remotely justifies the current expenditures, 124 00:06:24,520 --> 00:06:27,320 Speaker 1: let alone the trillion plus more plan for the future. 125 00:06:28,160 --> 00:06:30,640 Speaker 1: This is the vicious cycle of the AI bubble. The 126 00:06:30,640 --> 00:06:33,040 Speaker 1: more data centers they build, the more expensive it becomes 127 00:06:33,080 --> 00:06:35,680 Speaker 1: to build them. Because of the cartel of memory manufacturers 128 00:06:35,680 --> 00:06:38,080 Speaker 1: that adjust the prices up at basically whatever rate they 129 00:06:38,120 --> 00:06:40,680 Speaker 1: want to and hyperscalers that have no idea what to 130 00:06:40,680 --> 00:06:44,240 Speaker 1: do other than spending more money. RAM is ineffectively everything, 131 00:06:44,240 --> 00:06:48,080 Speaker 1: and thus everything is going to get more expensive from here. Samsung, 132 00:06:48,240 --> 00:06:50,360 Speaker 1: sk high Inex and Micron have committed more than two 133 00:06:50,360 --> 00:06:53,200 Speaker 1: trillion dollars to building out more capacity, even though this 134 00:06:53,240 --> 00:06:56,000 Speaker 1: bubble has been inflated based on building high bandwidth memory 135 00:06:56,000 --> 00:06:59,400 Speaker 1: that is being built at a scale beyond any reasonable 136 00:06:59,480 --> 00:07:02,880 Speaker 1: level of intry demand. Will that capacity because you can 137 00:07:02,880 --> 00:07:06,120 Speaker 1: build other things in the same RAM capacity could theoretically 138 00:07:06,160 --> 00:07:07,880 Speaker 1: be used for something else. The question is how they 139 00:07:07,920 --> 00:07:10,200 Speaker 1: intend to pay for it, and how many of these 140 00:07:10,240 --> 00:07:12,640 Speaker 1: actual commitments they've made with their own money, and indeed, 141 00:07:12,720 --> 00:07:15,360 Speaker 1: how many of the commitments made to them by hyperscalers 142 00:07:15,400 --> 00:07:19,080 Speaker 1: buy in video can be canceled? And I expect we 143 00:07:19,120 --> 00:07:21,400 Speaker 1: see a large and more grotesque version of the massive 144 00:07:21,400 --> 00:07:24,400 Speaker 1: write offs and losses across memory manufacturers that we saw 145 00:07:24,440 --> 00:07:27,200 Speaker 1: when the supply chain crisis ended in twenty twenty three, 146 00:07:27,280 --> 00:07:30,320 Speaker 1: after the supply chain crisis that started in twenty twenty one. 147 00:07:31,240 --> 00:07:34,240 Speaker 1: And when the time comes, know who to blame. Sam Altman, 148 00:07:34,480 --> 00:07:38,440 Speaker 1: Dario Amadei, Sundar Pashai, Andy Jesse, and Mark Zuckerbo the 149 00:07:38,520 --> 00:07:42,040 Speaker 1: men who created this entirely avoidable and horribly destructive situation. 150 00:07:42,440 --> 00:07:45,000 Speaker 1: They all fucking suck. They could have done literally anything else. 151 00:07:45,240 --> 00:07:46,880 Speaker 1: They could have built housing, they could have fixed their 152 00:07:46,920 --> 00:07:49,080 Speaker 1: own products, but they chose to do this because they 153 00:07:49,160 --> 00:07:52,880 Speaker 1: have no vision, These met a losers. I despise them 154 00:07:52,880 --> 00:07:54,640 Speaker 1: for what they've done, and I hate them for what 155 00:07:54,640 --> 00:07:57,360 Speaker 1: they've done to the computer. Thanks for listening. We'll be 156 00:07:57,400 --> 00:07:59,920 Speaker 1: back next week with a special memory focused episode. Would 157 00:08:00,120 --> 00:08:03,120 Speaker 1: the Burg of Gamers Nexus And I'll be back, of course, 158 00:08:03,120 --> 00:08:05,240 Speaker 1: with another monologue. Cheers, madeirs