Better Offline - Monologue: Blame AI For Making Everything More Expensive

Starting point is 00:00:00 This is an IHeart podcast. Guaranteed human. Run a business and not thinking about podcasting, think again. More Americans listen to podcasts than adds supported streaming music from Spotify and Pandora. And as the number one podcaster, IHeart's twice as large as the next two combined. Learn how podcasting can help your business. Call 844-844-I-Hart. All-Zone Media.

Starting point is 00:00:25 Hello and welcome to this week's Better Offline monologue. I'm your host, Ed Zitrom. Better Offline. And today I'm talking about the great parasite of AI and how it's making everything more expensive. And to be clear, before we go any further, what you're seeing in these remarkable sales from Nvidia and all the RAM companies and Broadcom and all them isn't a measure of industry demand, but of speculative purchasing based on the vaguest of projections, because the actual demand for AI, the actual revenue generated by this industry, is kind of pathetic.

Starting point is 00:00:59 And honestly, we don't really have any reason to build these data centers. Nobody has trouble accessing these services. I mean, it's just being done because they've got nothing else to do, and I know that sounds kind of flippant, but really is what's happening. To make that's worse, the more the bubble inflates, the more expensive everything gets. The more data soon as they get funded, the more chips that are ordered, the more resources this takes, and the more the RAM companies can crank up the prices and fuck over the customer. You see, AIGPUs use a special kind of memory, high bandwidth memory, that takes up more of the limited amount of fab space where SK-Hinix, Micron, and Samsung build their

Starting point is 00:01:36 RAM. These are also the only three companies that really can build HBM. It's a triopoly, I guess you'd call it. And because Nvidia, Broadcom and other chip manufacturers quite literally can't build the GPUs that hypers crave without them, their memory companies can charge effectively whatever they want. AI GPUs have become a massive amount of guaranteed revenue as the AI bubble inflates, leaving less fab space for the regular RAM, creating a supply shortage or, put another way, a reason for the memory companies to bump up the prices aggressively. And yeah, they're allowed to do that. There's nobody to stop them. It's not like we have regulatory authorities it would bother. I should also add that Nvidia is one of the largest, if not the largest, HBM buyer in the world.

Starting point is 00:02:17 55 to 65% of all high bandwidth memory sales go to Nvidia. And this makes it actually a victim of this same inflation, because memory is getting more expensive across the board, even for high bandwidth memory. Goldman Sachs reports that high bandwidth memory is on course to increase by as much as 90% year over year in 2020. A linear increase in costs for every single Nvidia GPU and every single new one that one of the four hyperscalers Oafs insists on pre-ordering only serves to inflate a bubble

Starting point is 00:02:46 for effectively no reason. They're not making a profit. Shit, they won't even talk about revenues. Maybe they're not really making that much outside of Open AI and Anthropics compute, like I've been saying. But good Lord, good Lord, good golly, Miss Molly. Do these data centers take up a shit ton of RAM?

Starting point is 00:03:03 Analysts estimate that by the time that Invidia ships its first next generation Vera Rubin GPUs, it'll cost $18.40 per gigabyte of high bandwidth memory sold 288 gigabytes at a time. In fact, let me give you some scales to how much RAM and a data center actually needs, because it's an astonishing amount. The current generation GB300 racks, the vast 72 GPU servers that fill out modern data centers and they have a CPU for every two GPUs, have about 20.7 terabytes of high bandwidth memory and 17 terabytes of 7 terabytes of LPDR5X RAM. And by the way, that's the RAM that you find in smartphones and mobile devices, which is also

Starting point is 00:03:43 contributing to the supply chain shortages and driving up the price. Anyway, a gigawatt data center is about 4,933 of these racks working out to about $1.94 billion in high bandwidth memory alone for a single data center, a number that's certain to increase across the board, even for current generation gear because the price of high bandwidth memory is going up. I haven't even included the many separate computers and different machines inside the data center that require RAM too. Entire racks of power management gear that themselves require RAM or HVAC systems alone, there are effectively servers unto themselves.

Starting point is 00:04:19 And every data center is basically a giant warehouse of different RAM-dependent devices and machines, each controlled by other RAM-dependent machines, all to do one thing. AI that's yet to prove it even has a meaningful business model, or one that, I don't know, makes more money than it loses. And that high bandwidth memory, by the way, is also not particularly useful for anything else, at least not at the scale it's being built, meaning that once there's a noted capex pull bag, somebody is going to be left with a bag,

Starting point is 00:04:44 by which I mean somebody is going to either have too much high bandwidth memory or gear that they can't sell with it put on the side that won't be useful, or really sellable to someone else. And I must be clear how much more of this they are selling than usual. Back in 2022, they sold about 600 million gigabytes of high bandwidth memory. So it sounds like a lot until you realize they're expecting to sell 7 trillion or more of it in 2026. That's a problem. That's a problem because the more of it they sell, the more expensive it gets.

Starting point is 00:05:16 But they're going to have years worth of this crap just either sitting around or being attached to Nvidia GPUs that can't sell to anyone. And the longer it takes for Kappex to retract, the more expensive everything becomes as a result. sold. The more GPUs that gets sold, the more capacity that gets put towards high bandwidth memory, and the more that Micron, SK Hinex, and Samsung can charge for it, which makes it more expensive to buy AI GPUs, which increases the amount that hyperscalers are spending on AI CAPEX for effectively the same amount of gear. The longer that hyperscale has sustained this pace, the larger the return needs to be, and at this point, none of them have disclosed their actual AI revenues, which heavily suggests that there's yet to be even a single dollar of profit, and the actual

Starting point is 00:05:57 quarterly revenues are pretty pissed poor. Yet the more they commit, the more committed they have to be. Pulling back at this point will prove to the markets that all of the hyperscalis are committed far too much to too much capacity. Yet not pulling back means that hyperscalers will continue to turn their free cash flows negative in pursuit of some indeterminate goal. It's a vicious cycle made worse by the fact that every spin of the Kappex wheel increases the price of just about every consumer electronic in the world, creating a market-wide inflation for what amounts to a speculative asset bubble, and will leave them sitting with a bunch of GPUs they don't really need now and certainly won't in the future. What great works are these

Starting point is 00:06:36 data centers in Searchhold? What possible achievement can you point to at this time that remotely justifies the current expenditures, let alone the trillion plus more plan for the future? This is the vicious cycle of the AI bubble. The more data centers they build, the more expensive it becomes to build them because of the cartel of memory manufacturers that adjust the prices up at basically whatever rate they want to, and hyperscalers that have no idea what to do other than spending more money. Ram is in effectively everything, and thus everything is going to get more expensive from here. Samsung, SK-Hinex and Micron have committed more than $2 trillion to building out more capacity, even though this bubble has been inflated based on building high bandwidth

Starting point is 00:07:15 memory that is being built at scale beyond any reasonable level of industry demand. While that capacity, because you can build other things in the same RAM capacity, could theoretically be used to for something else. The question is how they intend to pay for it and how many of these actual commitments they've made with their own money and indeed how many of the commitments made to them by hyperscalers by Invidia can be cancelled. And I expect we see a larger more grotesque version of the massive write-offs and losses across memory manufacturers that we saw when the supply chain crisis ended in 2023 after the supply chain crisis that started in 2021. And when the time comes know who to blame. Sam Altman, Dario Amadei, Sundar Peshai, Andy Jesse and Mark Zuckerberg. The men who created this

Starting point is 00:07:59 entirely avoidable and horribly destructive situation. They all fucking suck. They could have done literally anything else. They could have built housing. They could have fixed their own products. But they chose to do this because they have no vision. These men are losers. I despise them for what they've done. And I hate them for what they've done to the computer. Thanks for listening. We'll be back next week with a special memory focused episode with Steve Burke of Gamer's Nexus. And I'll be back, of course, with another monologue. Cheers, Medea! This is an I-Heart podcast.

Starting point is 00:08:34 Guaranteed Human.

Better Offline - Monologue: Blame AI For Making Everything More Expensive

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.