Full transcript of “Die KI-Blase & warum gerade Technik teurer wird...”. Click a timestamp to open the video at that moment. Need another video? Use the YouTube transcript generator.
0:00 Hey guys, lately I've been hearing more and more that AI is just a bubble, and not just any bubble, but a bubble about to burst, that things are about to go downhill fast. And I mean, we’re all clearly feeling the effects. RAM and storage prices have risen drastically. 32 GB of RAM cost less than € 100 not too long ago. Nowadays, you're paying € 350 to € 500 for that. I bought a 1 TB SSD not long ago for under € 50, and now it costs at least € 150. In the server sector, it's even worse, with some price increases of over 470%. Yeah, and then there's the Nvidia RTX
0:46 5090. At launch, it was supposed to cost € 2099. It never cost that little. Now you have to shell out over € 4000 for one, if you can even get one at all, because the drops are often sold out in seconds. Phew, tell me, where is all of this actually leading? And is there an end in sight, a light at the end of the tunnel? I've been making tech videos here on this channel for 13 years now. Actually, I had my 13 -year anniversary this week. Belated happy birthday. Thank you very much. Very nice of you to think of that. And I thought, well, I have a few insights, I have a bit of experience. How do I assess all of this? That's what I want to tell you in today's video. Do you want the good news or the bad news first? I can’t hear you. Let’s
1:34 start with the bad news. Prices will continue to rise and there is actually no end in sight. Even Apple, a company that had long-term contracts for massive amounts of storage and RAM. even they had to raise prices recently, and quite significantly at that, at least € 100, often even € 300. They only spared one product, and that is the iPhone, because it is their most important product. But we already know that the next iPhone will definitely see a decent price increase of € 100 to € 200. All just because of the storage. Normally, the processor is the most expensive component, and memory comes in 4th or 5th place. That has now flipped. Especially in low-budget
2:21 devices, storage is now the most expensive component, costing between $ 80 and $ 120 per device. Nothing CEO Carl Pei says that in the Nothing Phone 4R Pro, it is the most expensive component, and the device costs € 400 to € 500. But even very expensive smartphones, like the newly introduced Galaxy Fold 8 Ultra, now cost € 100 more, totaling € 2200, even though the RAM is still only 12 GB. The crazy thing is, there is more RAM and storage being built than ever before in human history. Never before has so much been built and so much money been earned in this sector. And interestingly, the whole thing is carried by just three
3:07 companies. Samsung, SK Hynix, and Micron. These are the only three relevant memory companies, but they are running completely at their limit. They simply cannot produce any more, and that is where supply and demand come into play; the demand is gigantic, which gives them the opportunity to raise prices, and prices have risen accordingly, and these three memory firms have one major customer above all : Nvidia. They buy practically 90%of all memory in the world, and it is then sent to TSMC, where the processors and graphics chips are built together with this memory. I mean, you have to let that sink in. In the end, one company is buying up all the memory in the world,
3:53 and that is what is driving up the prices we feel so drastically. But it is mainly because these memory companies are no longer producing as much storage and RAM for us customers, but are now focusing primarily on HBM memory. That is memory with extremely high bandwidth, extremely fast, and they have switched their production to this type of memory. So instead of building memory for smartphones and computers, they prefer to build the memory needed for AI servers. One could logically ask the question: why don't they just open more factories? and thereby simply sell more memory, which would stabilize prices while still allowing them to make more money overall. The thing is, not even 3 years ago, all three of these companies were
4:39 losing money. They weren't making any profit from selling memory, and that’s because they had more production capacity than they had buyers. And they certainly don't want to have that happen again. So on one hand, they don't even want to build more factories because, well, that hasn't proven to be a good idea in the past, and on the other hand, it naturally takes an eternity to build such a factory. It is highly complex, you don't do that overnight, and that also means that no new outside companies are emerging because it is far too complex to just spontaneously start a memory company. I keep reading that all these AI companies are buying an RTX 5090 and putting it into their servers, but that is absolutely not the case. They don't buy the graphics cards
5:28 , memory, and chips like we do; instead , they primarily buy the HGX system from Nvidia. This is a server board that consists of 8 B300 AI chips, and the entire server rack has 72 B300 chips and a staggering total of nearly 21 TB of RAM, and god, I really have to read this name out loud. The Oracle OCI Supercluster for AI. Let's just call it the Super Duper Computer for the rest of this video. So, this Super Duper Computer, which is currently being built primarily to be leased to OpenAI, is being built by Oracle. This supercomputer is set to have a total of 800,000 of these B300 chips. That would
6:14 be a total of 230 petabytes of RAM. This entire Project Stargate, which is currently being built in Texas, will cost 500 billion US dollars. And they are only spending all this money, of course, because they think they will A) recoup it and B) even make a profit from it. Yes, incredible amounts are currently being poured into the expansion of AI data centers, into training AI models, and of course into employee salaries, because there are some extremely well-paid people there. My little overview of which company is putting how much money into AI just this year. Amazon 200 billion, Microsoft and Google 190 billion each, Meta 135 billion. And you might be
7:02 thinking now, surely that's a lot of money even for these companies. They must be running out of money soon. But no, they have so much incredible money that they can really afford it. And they also get a lot of money in the form of bank loans, of course, but there are companies that are actually just taking on debt, and that is. For example, Oracle right now: they have an incredible amount of money, but are actually at the point where they are just borrowing money from others. And then, of course, we have OpenAI and Anthropic and many other AI companies that are actually just losing money and living off investor funding. So they simply have the hope that it will all pay off somehow in the future and they will recoup everything they are losing right now. Those are unimaginably large sums of money, but I'll tell you how
7:49 you can get a little bit of money back from the tax office, namely with the sponsor of today's video, Zasta. On July 31st is the tax deadline, so there really isn't much time left, and you have an extremely good chance of getting a lot of money back from the tax office. The principle is totally simple. Zasta checks for free if you get money back. For that, they only need your name, address, and tax ID. All the rest, meaning the tax return, is then handled by their professional, real tax advisors. And the average refund is an incredible 1240 Euros, and that even goes back up to 4 years retroactively. Registration takes less than 3 minutes. I have never seen a service this easy for getting money back from the tax office. So definitely
8:35 check the whole thing out. And of course, I also have a discount code for you. With Review 20, you get a € 20 discount on the tax consultant fee, which of course only applies if ZASA can get money back for you. You can find a link down in the video description. Check it all out, and back to the video. Okay. Is AI a bubble, and will it burst soon? Well, if we take a look at the points, there is definitely quite a bit that suggests it. When Apple spends € 1, they earn 40 cents from it. When OpenAI spends € 1, they lose 62 cents from it. Yes, they are pouring an incredible amount of money in, and simply not that much money is coming back. And most AI companies are in the same boat. Google, for example, is just lucky that they earn so much
9:21 money from all their other services that they can cover it. Scenario 1: Let's say OpenAI actually runs out of money. What happens then? They can no longer pay anyone—not their employees , not the AI training, and above all, not the servers they rent, especially from Oracle and Microsoft. So those companies then have a problem because their servers are suddenly sitting idle , and they will adjust their prices accordingly. Consequently, that will affect all other data centers. Accordingly, expansion will decline sharply. Consequently, Nvidia then has a problem, and thus TSMC too, and therefore the memory companies as well. Let's say Oracle can no longer afford the whole thing. No one is paying them for their expensive servers. What then? Then, of course, banks are also
10:07 affected because they suddenly don't get their money back. And the whole thing is a massive chain reaction. Banks will fundamentally give out less money to AI companies, perhaps none at all. And other AI companies will be forced to be profitable overnight. But perhaps they can't, and then they go under too. But don't worry, some companies are looking forward to exactly this situation and are well prepared for it. Microsoft, for example , has secured extensive rights to OpenAI. If OpenAI were to go bankrupt, Microsoft would simply take everything over. They definitely have enough money to afford the whole thing, and everything would just go on as before. Yes, and Meta and Google would surely also buy up a lot of what is currently going down the drain. Yes, so OpenAI's plan would have to be, a) to become
10:53 profitable at all costs, and b) they naturally want to go public now, because they don't want to be dependent on investor money anymore, since if investors suddenly lose confidence, they have a major problem. Scenario 2. Two things happen at the same time. For one, cheap, very good, and sometimes even free open-source AI models are coming out of China, and they can suddenly beat the much more expensive models from America or at least keep up very well. That happened, and what happened at the same time? That Anthropic releases Claude Fable. An AI model so good—the best AI model, many say—that the USA suddenly bans it. They say you have to turn it off for
11:40 non-US citizens. But because Anthropic cannot see who is a US citizen and who isn't, they turn it off for everyone. In that moment, governments and companies from other countries realize one thing. We are dependent on whether the USA grants us access or not. And they naturally don't like that, because they don't want to be dependent on them . Accordingly, they grab these cheap models from China or develop their own models and build their own data centers . And that is already happening in Germany right now. Lidl, Kaufland, VW, Siemens—they all prefer to rely on their own data centers rather than the big ones in the USA. One could naturally say that maybe this is just a temporary phenomenon. VW once tried to produce software itself and that didn't work out so well, and in the end, they
12:26 had to buy the software from the USA after all. But who knows. For now, this definitely hurts big growth and accordingly makes the math for these AI models a bit more complicated. Scenario 3: AI models are primarily attractive to the economy because they are cheaper than a human. If an AI model can do the same thing as a human for just a fraction of the price, then a company will say, we'll use an AI model. Only it's turning out now that AI models are getting more and more expensive, and specifically Claude's Fable is so pricey that some firms realize, wait a minute, a human is actually cheaper. And some companies have even gone back to hiring new people and turning off the AI. We have examples like Klarna in
13:13 customer service, McDonald's in the ordering process, CNET, but also Google and Meta. Well, perhaps at the end of the day, we humans are not quite as easy to replace as some investors hastily thought. But even there, I must honestly say, that is probably just a temporary phenomenon. In the long run, AI will obviously keep getting much better, and while we humans will improve too, perhaps not quite as quickly. Scenario 4: The AI bubble doesn't burst at all. Everything works out, everything is profitable, everyone claps their hands and is happy. Autonomous robots take over tedious work for us, handle agriculture, climate change is solved, we live in a utopia, and nobody ever has to work again. We have created a massive superintelligence that takes care of
14:00 everything for us. I mean, when you say it out loud, it really sounds like it wouldn't work, but ultimately that is actually the plan behind all of this, right? So, that is what is being worked towards. And humanity has managed a few crazy things throughout history. Regardless of which of these scenarios occurs, I am certain of two things. First, AI isn't going anywhere. It will definitely stay and get better in the future. And second, there will definitely be some form of crash, because many companies are currently spending completely excessive amounts of money on something that just doesn't pay off yet. There is too much hype behind it, and the whole thing needs to regulate itself, and it certainly will.
14:45 But the bad news is, memory prices will stay high for the next one to maybe even three years. I hope that after this video you understand a little better why we have these high prices now, how it all connects, and what the possible situations are. Let me know your opinion in the comments; I am really very, very curious to hear it. Check out Zasta and take care. See you next time. Ciao.
