원문 보기
A quick announcement today: the Jhole coupon for Meet Kevin membership is expiring. It gives you access to all nine courses, every Alpha report, every one of our private live streams, deal alerts, what we're buying long- term—whatever it is —on meetkevin.com. Take advantage of this Jhole coupon that expires today.
Tax experts say this could be a tax deduction. And if you want to purchase a package, today is the last day to get lifetime access to a Reinvest Homes membership. We are moving to a monthly and annual subscription model. If you activate lifetime access, you will have it forever. Okay, fine. By the way, if you are already a course participant and would like to add Homes AI, email us at [email protected] and we will provide you with a special package code so you can get lifetime access to Homes AI.
The end of the AI semiconductor era in 2026: are computing power calculations still converging? We have a viral six-page analysis of whether it makes sense to invest in computing. Does it still make sense to invest in equipment? Is this a " bubble"? When will this end? This article raises many questions and provides the author's perspective on many of them.
For example, after programming, what will be the next step in regular income? Does this economic model work? Why might this be different from the era of the Internet? Yes, the famous words: "this time everything is different." I processed, summarized all of this, and commented on each point. So we'll highlight the most important points and explain their argument, because, frankly, a six-page text is a bit difficult to read .
So, we will move on to the main conclusions. So let's begin. The first argument they make is related to the idea that we are observing another form of exponential growth. There is a risk to this argument, but I understand what they are saying. They're essentially saying that back in the days of the internet, you just multiplied the number of people by a fixed rate, and that's how you could make money.
So your S-curve was based on how many people you could serve, multiplied by a flat fee of, you know, $49 a month or something. Here, as they claim, you have something different. You have people using AI multiplied by tokens, and there are actually two S-curves at play here. So you have an exponential growth in the number of people using AI, as well as the volume of use of that AI.
And then you could multiply these two S- curves and assume that the hardware rally should last significantly longer than, you know, the Internet boom or the dot-com bust era, because you have this additional vertical growth, essentially, or this additional multiplier effect. I understand this argument.
The problem, for me, is that I don't really see user usage as exponential. Mainly because even if I go from using 10 tokens to 100, 1000, or 10,000 tokens—it's an exponential curve, but it's held back by the rapid depreciation of the token value. And there is a limit to how many tokens I will actually use.
For example, how much do I actually need to communicate with artificial intelligence? There is a functional and practical limit. My time, productivity, profits, budgets, whatever. But, in addition to my limitations on how many tokens I actually need, it's worth considering— which they don't mention, by the way—that tokens are an asset that depreciates quickly.
And people forget about it. Remember this. As for tokens, I can use a Chinese model from Alibaba called Qwen and spend 41 cents for 1 million initial tokens. If I go to Frontier Labs, who spend money on Nvidia like drunken sailors, I'll spend $50 for a million initial tokens in Claude 3.5. This is madness.
That's a huge difference. And we are only at the beginning of this depreciation phase. So I don't know if I agree that this time it's different , because the tokens create a multiplier effect. I think the factor of their rapid depreciation somewhat negates this argument. So, this was my, let's say, refutation of this thesis.
Then they say this. They say that token spending per person, the median AI spending in US businesses is now $12 per person per month. In the long term, this figure could well reach 10 % of the average office worker's salary, approximately $12,000 per year or $1,000 per month. So my reaction to this is: you'll be laughed at in any company other than IT if you come to them and say, "You need to pay $1,000 a month for ChatGPT." Dude, 90% of American businesses will just kick you out the door if you say that.
Because a dental hygienist only needs a chatbot or a large language model to a certain extent, and it's definitely not going to be a $1,000 subscription, okay? It seems to me that they are looking at the development of AI somewhat through rose-colored glasses. This is indeed an extremely optimistic forecast, which we will now turn to.
Every employee is associated with... Mainly because I don't think that's where the value chain accumulates. I believe that in AI there is an infrastructure level that is depreciating and a token level that is also depreciating. The software level wins because it is the true source of ultimate value.
Let's take the example of dentistry: if I have an X-ray machine that automatically determines the probability that a certain spot on your teeth in the image is caries. Well, great. As a dental office, I will pay a certain amount for this software, but I highly doubt it will be $1,000 per month for each employee, right?
So, we will pay for the value of the result obtained. Sure, it might be worth something in a monthly calculation for the entire clinic, but I don't think it's $1,000 a month for each employee. In any case, this is just an example. They rightly argue that any exponential growth eventually hits a ceiling.
I think they understand that there is a limit, but some of their assumptions are very, very optimistic. Okay, let it be. Then they say the marginal cost of copying blah blah blah, what is that? OK. Another reason why the demand for AI development is higher is because the internet, well, let's see, the penetration parameters have turned into separate tokens.
This is essentially the idea that in the days of the internet, there was no additional cost to generating marginal revenue. You're just letting someone get a copy of the CD, essentially. You sell software, you earn more money, you get a higher margin. There are costs with tokens. And that's true. But the question is, to what extent will the value of the token exceed the costs of its creation?
I think that in reality there will be a very low margin, while in software it remains high. The essence of the software is that you sell subscriptions with a gross margin of around 75–90%. With tokens, I think in the future it will be like mining Bitcoin, where you mine a Bitcoin for $90,000 and hope to sell it for $91,000.
That spread will shrink a lot because anyone can do it. You're just creating a server farm, right? Okay, one more nuance worth taking into account. Let's continue. They then say that in principle everything can be called coding, because, as they say, "Hey, what will be the next coding for AI?" This is coding.
This is their argument. They say, “Oh, AI for science is just coding. AI in finance or business—it’s all coding.” Let it be so. You can rebrand and say that everything related to AI becomes coding, because frankly, behind the scenes it's just a type of code, but that doesn't change how much people need AI.
It's just a kind of rethinking of how you position AI . A great wonder. The difference, and this is the point they don't touch on, is that they are very excited about the growth rate of Codex—it's a new product. You're comparing to low growth rates, so it's easy to say, "Oh, Codex active users grew 108x in legal and 41x in sales." Yes , that's great.
But we are comparing it to a database where six months ago no one used Codex. So, this is another difficult statement to compare. Ultimately, what concerns me is that we had some growth in AI, then the moment with Claude Code and co -working, and the moment with Codex. And my concern is , was that moment just part of a cycle?
And are we here now? You know, I think that's a perfectly logical part to consider. Someone writes: “Excel is just coding.” Yes. That's right. So yes, you can call everything coding, right? That's the idea. They wrote it here somewhere. So that's another problem I see with this material. Then they come over here and say, “Here’s the thing.
Anthropic slowed revenue growth in June and July solely due to a lack of computing power. There was a restriction on distillation that closed many accounts, and Codex competed with them.” OK. This all sounds logical. If you close many accounts, you will lose profit from them. If there is not enough computing power to scale, it is understandable that growth slows down.
And when a big competitor appears, that's also completely understandable. But they missed the most important thing. They missed the most important narrow point with profit margins, which are destroyed by open-weight models. This is the example I gave here on the board. Why should I pay $50 for 5 million Fable token withdrawals when I can pay 41 cents for Quen, which is 95% as good?
Going back to that person's argument, like, oh, you know, we need AI in the biological sciences or, uh, you know, in my spreadsheet or somewhere. Do I really get significantly more value with a good, well- structured query? Am I really getting a hundred times more value from Fable 5 compared to Quen?
Probably not. You might be getting 10% more value, but that only justifies the extra 4 cents of cost, not, you know, $49. So this is a problem they don't mention in their optimistic forecast. Although they quite rightly claim that there are reasons why Anthropic's revenue has slowed. Okay, let it be.
Quite right. So, next they mention here, they talk about the potential growth of Anthropic. And this is where they really show their optimistic scenario. They say they really believe Anthropic can grow revenue here. Listen, let me just read you this line. In 2028, when there is no visibility, one can only guess.
So, if there is no visibility, let's just assume the following. We assume that in 2028 Anthropic will be able to increase its annual recurring revenue by 80 % per year. Which then justifies seven times more computing power than OpenAI and Anthropic have today, as they have about six gigawatts of computing power.
They say, "If they can increase revenue by 80%, we can increase the gigawatts of compute to 45, which is again 7 times more than today." "And then, if annual recurring revenue doubles again in 2029 and 2030, with an average annual growth rate of 40%, then OpenAI and Anthropic's planned 100 gigawatts of computing power by the end of 2030 looks quite reasonable." Okay, wait a minute.
There's so much to sort out. We'll just break it down really simply. Today we have about six, maybe seven gigawatts of computing power. They claim that in four years we can justify 100 gigawatts of compute and still see OpenAI and Anthropic's revenue grow at a 40% compound annual rate, when that's the competition they face.
I don't see this. I don't see this. I think the computing power could reach 100 gigawatts, of course. But I believe that a significant portion of them will ultimately be open-source calculations. And they address the problem of computing with open weights, which disappointed me a little. So, here it is , this is their argument.
They make an argument called " closed code." They say: closed source vs. open source. I don't remember if it was right here, but wherever they mentioned it, I'll find it. Oh, yes, right here , here, here. So, Open Router . Okay, okay, here it is . They argue that closed- source companies like Frontier Labs will actually be able to maintain their advantage, because if you choose open-source products, you will need engineers to run them, you will have to spend money on computing, and you may not be able to optimize them effectively .
Here it is . If an open source model is poorly optimized, costs can even exceed closed solutions. It 's easier to just give it to a "token factory." So I literally write, "This is crazy." This means that engineers in the value chain are supposedly too incompetent to optimize a model for 41 cents compared to a model for $50 per million initial tokens.
I find this insulting to engineers. I think every engineer watching this will say, " Dude, I can squeeze out as much optimization as possible for this kind of savings." So I thought this was another rose-colored glass argument. And I'm like, "Damn, these are the same people who go viral promoting the hardware cycle." And they are quite right to argue that all these open source models don't actually generate much revenue.
But this is a warning sign. Look at this. They literally write: analysis of China's open source revenue. A billion here, 1.2 billion here, 500 million here, 300 million here. You're barely reaching 10 billion. Okay, that's a problem. If you add up all the open source revenue and barely make it to $10 billion, how do you get 65 billion from Claude?
Well, that's because Claude just had his "Claude co-work" moment. There is a real explosion here. And these models are just beginning their journey. Their thesis or argument is that "well, this is obviously much better, so it will continue to grow at this rate." My opinion: the latter will eventually surpass the former and break this growth rate.
Obviously, these are two different views. You have to decide whose side you're on, but you can see some of their hidden assumptions, like, " well, you know, open source can't be optimized as well ." The fact that you dismiss open source software by saying, "Well, if you don't optimize it well , you 'll spend the same as you would on closed source" is just amazing to me .
Maybe I'm misunderstanding this argument, but it seems to me to be a crazy and blind "bullish" thesis. OK. And I don't want to seem too pessimistic. I just think it's too optimistic. We need to come back to reality a little. Here is another mistake they make. They draw a parallel with MediaTek, which at one time offered cheaper chips compared to Qualcomm's Snapdragon products.
And MediaTek, the Taiwanese player, has actually done a pretty good job of impacting Qualcomm's margins. But they argue that Qualcomm still managed to stay ahead and make good money for years against MediaTek . And that they had Snapdragon products: one for the high-end price segment, the other for the budget segment.
And Qualcomm was able to succeed. So, that's why this more "open" or cheaper MediaTek style, if you will, doesn't matter. Therefore, open source code is also irrelevant for AI. But wait a second. There are significant differences here. MediaTek chips were 30–50% cheaper than Qualcomm. Chinese models with open scales are 100 times cheaper than advanced developments.
That's a huge difference in cost. We are not talking about some insignificant difference of 30–50% in the margin of a cellular chip. We are talking about a hundredfold difference. This is impossible to compare. Second thing: once you buy hardware for a phone, like installing a Snapdragon processor for mobile data, you can't replace it.
I mean, maybe you can , but it's not easy. You won't disassemble the phone to replace the chip with another Snapdragon-style one. You just buy a new phone. Tokens can be replaced. I can just plug in another model like Palantir does. Hey, what model do we want to use? Let's just do the routing, shall we?
And Qualcomm also had patent protection. Therefore, I believe that the author is looking at hardware through rose-colored glasses in every scenario he proposed in this article. I'm not trying to be pessimistic about hardware, but looking at this " bullish" argument, I think: it's ridiculous. Sorry. That's just too optimistic .
And if these are the people who are inflating the prices of these stocks, then it's a bubble. Uh, that's not good. Uh, but anyway, uh, what else? They claim that we are not in a bubble until at least 2027-2028. Okay, maybe. You know, I mean , I'm optimistic about the near term. I think we'll get through Jackson Hole, and uh, we can be optimistic in the near future.
Uh, you know, you can simultaneously think that equipment or stocks are going to do well for the next 6-12 months, and at the same time think that it's all going to come crashing down in an epic bubble burst, uh, you know, in two to five years, right? So, I, that's pretty much what I think. But these arguments, in my opinion, are worth dissecting, because they have many errors, and if someone simply asks an AI to summarize this entire document, it will miss important nuances.
And maybe I'm somewhere, you know, maybe I'm too radical in my counterarguments. I do n't think so. Uh, I think my counterarguments are valid. But there are some fatal flaws in this bullish narrative, and anyone investing based on whether the numbers are still coming together should really understand the arguments on both sides and then build their own case on tha.
Comments 0