원문 보기
Quick reminder, the coupon on meetkevin.com expires today. In this segment, we will look at the most pessimistic graph in the field of artificial intelligence. And this comes against the backdrop of Anthropic's desire to significantly increase its credit lines. $15 billion pre-IPO credit line. They're essentially telling the banks, "Hey, if you want to participate in this IPO, you need to provide a loan." They had a $2.5 billion credit line .
This is now 7 times larger and three times the size of the pre-IPO credit line that SpaceX received. Therefore, the credit line capacity is much greater. Typically, you take out a line of credit as a contractor when you have work for clients, but you have to pay employees and then repay the debt after getting paid by clients.
It is hoped that Anthropic will be able to generate more revenue to cover its bills, especially after signing a $ 45 billion deal with Lambda . This is a 6-year contract. It will cost them $7.5 billion a year for computing power. They have a computing contract with SpaceX. They have one from Lambda.
They have a lot of contracts for computing power. So bills—that's a really big problem for Anthropic. But this is probably not the scariest schedule. The scariest chart is actually about what we have here. This is an article from Zero Hedge discussing what Goldman Sachs calls its most pessimistic material.
I'll get to the most pessimistic part right now, and we'll delve into the basic concepts of this issue. The most important thing to know today: the J Hole coupon on meetkevin.com is expiring. Be sure to take advantage of it before tonight, as we expect a significant price increase. Additionally, we are closing the opportunity to subscribe to lifetime access to Reinvest Homes AI on meetreinvest.com .
So if you don't make it today, you'll never be able to get lifetime access again. So consider visiting meetkevin.com and meetreinvest.com. Now that that's done, look at this graph. This is the pricing chart for Large Language Models ( LLM). And this is based on usage. So it's, you know, a weighted average based on usage.
How much money are people paying for 1 million initial tokens? And in fact, we see that the indicators have not changed since last year. Last year—this is somewhere around here where I started covering this. That's when this index started, this graph appeared. So, we don't have data for a longer period.
But here we see two spikes. These two surges, I think, came about because of optimism about collaboration and coding. That is, the efforts of agent systems for non-developers, as well as Claude code and the important moments we had at the beginning of the year. You can see peaks in how much people were willing to pay for 1 million initial tokens in January, and another peak in May.
In fact, we collectively consider this an Anthropic moment. The problem is that it's falling now, and the real question is: if this figure is falling, what does that mean for the future of advanced computing and spending? Many people believe that we will simply get the next level of results, which is where the enthusiasm for coding and AI will come from.
Maybe it's not coding anymore, because we've already passed the coding point. Maybe it will be healthcare or finance. But there is a slight problem with this assumption, and we will return to the Goldman material. You see, The Economist made a great argument for why this coding moment for Anthropic is potentially risky and not as long-lasting as it seems.
If you analyze, they are actually arguing that coding is structurally different from many other sectors of the economy. In AI programming, you have a lot of data availability. Remember GitHub. It is very easy to get a lot of data for training. Think about how easy it is to validate data for coding. The code runs without errors or it doesn't work.
There is vulnerability or there is not. It's a slightly more binary situation. People who use artificial intelligence in coding become clearly and quickly more efficient. So, people have an incentive to program using artificial intelligence. Whereas some benefits in finance or medicine may be more long-term and less obvious.
Therefore, you potentially have a lower level of implementation in other sectors or a slower rate of adoption. And then there's the idea that engineers are the very first users of technology anyway. It's just a little different in that programmers can benefit the most right away, but they're also usually the early adopters, the software engineers.
So The Economist makes an argument that essentially says: be careful thinking that other categories will pick up on this AI explosion that we saw with Anthropic, when everyone was excited about tokens and was willing to spend an average of $2 per million initial tokens. And now it's down to less than $1.
Part of the reason it has dropped below $1 is likely due to the release of open-scale models that are becoming increasingly powerful. This is actually positive for certain sectors of the economy, and it's one of the reasons why we've seen Dell stock, which we cover on the channel—be sure to subscribe—have skyrocketed.
And I think this will potentially help create a " hardware 2.0" rally before the big bubble collapse of doom. We'll explain that in a moment. But focus on this for a minute. This graph is not a usage graph. This is a graph of how much companies are paying for a million initial tokens. When it falls, it is positive for use.
This is not a usage graph, it is a positive for usage. This is positive for anyone implementing corporate AI. This is positive for anyone running local models. This is positive for OpenAI, which is acquiring Max Studios to run inference to get faster results for its clients. This is likely a positive for companies like Cerebras, which are releasing ultra-fast tokens through their hardware, as well as Nvidia's release of the Grok model, mainly because it is now possible to distinguish between tokens.
A token is no longer just a token. There is a regular token, and there is a superfast token. In any case, in my opinion, this is also positive for value providers. We keep talking about the stack that tokens are hard to monetize in the long term because they become a commodity, infrastructure becomes a commodity, value delivery wins.
We've seen this surge recently in software development companies. We talked about this, and that's why we had a thesis about the Software Rally 2.0, or sorry, 1.0. The software 1.0 rally was that in Q3 and Q4 we would see software stocks bottoming out , and so an extended software ETF like IGB , which lost a little bit today , but that's nominally relative to the trend, we really see that trend and the bottoming process is complete.
And now we are still in an uptrend, so the bottoming process is over, the uptrend for software continues. These are the value providers. In my opinion, these value providers are similar to what we do with our meat reinvest product. You don't have to buy it if you don't want to. Just look at him. Look at the product we have at meat reinvest.
And imagine you want to buy a house and you are considering a certain county. Here, I'll put it on the screen so you can see. And if you don't want to, you don't have to go there. You look at a specific county and you say, "Hey, you know what? I want to go to the AI tab for homes. Hey, I want to find a good deal.
Show me the deal." Okay, what do we have here? We have Ventura County. Okay, here are a few options. What is it? Here is an offer for 522. I know this deal. It turns out that this is confiscated property. This is a good deal. Of course? The AI instantly tells me, " Kevin, if you're looking for a great deal in Ventura County today, take a look at this.
I wouldn't be surprised if it gets a few offers." And honestly, I think the app is even underrated. I think the cost after repair here is closer to 675, but you might have to spend a little more. So, you will have to spend somewhere around 60. So in fact, the estimate is quite close to reality. But people can gain capital from this deal.
you can use...we run this on our own GPUs and CPUs. We own the equipment. This is an example of how cheaper tokens that we can use with payback on our own machines can provide real value to people, agents, and consumers through this application. And if we sell this app, we make money, the consumer and the customer make money—it's a win-win .
It's like what Palantir offers. So cheaper tokens are in our favor. I welcome this schedule. But that doesn't mean there isn't a big problem. And that's exactly what we're going to talk about now. Quick note: I want to make it clear that we are not raising funds or anything like that right now. People ask, "Kevin, this is crazy.
I thought your startup was a real estate company." That's right. I want this to be clear. We are a real estate company. That's why we changed the name to Reinvest, because we simply take profits or excess money and reinvest almost all of it into real estate. So yes, we are a real estate company, but a hidden AI proptech firm.
That's pretty cool. Okay, what do we have here? That's the big problem. The AI industry has just faced the problem of the unit economy. And this is from the Goldman study, which claims that there is significant revenue concentration as Broadcom advertises how big an Anthropic customer it will be, but that could be a problem because Anthropic and OpenAI are significant customers.
And if there is a downward revaluation of tokens, then you may see the price of work fall faster than expected from the development of computing power, so to speak. So, I call it the V-shaped problem. Let me first explain what I consider a V-shape, and then I'll do a comparison with Coinbase in a moment.
The V-shaped problem, in my opinion, is that right now a lab like Anthropic is making a lot of money selling cutting-edge tokens. I think the next growth curve for a company like Anthropic is the S-curve of selling you software. So, this is the ARR (annual recurring revenue) from selling you the software.
The problem is that it will fall. And so, in the end , you get what I expect, right? We have no guarantees. But I expect you'll get an intersection where you actually see what looks like a V. You see this V right here, right? Do you see this? Descent, then ascent. Good? The faster this V is, the better, because the software essentially compensates for the reduced margin on the tokens, right?
Anthropic, in essence, follows the path of Coinbase to some extent. When Coinbase went public , the whole idea was, "Hey, the money we make from the exchange rate difference from people buying Bitcoin and tokens on our platform is going to go down." So, there was an expectation that they were going IPO at the peak of tier one revenues, if you will.
And that they would compensate for this by earning more from second- tier revenues, which were essentially crypto-related financial services. And we saw how long it took for financial services for crypto to really become popular. And now there are many other structural risks, such as FedNow, which is another structural risk.
To be fair , some of these companies, like Coinbase, have been doing really well lately, but that's because Bitcoin has jumped from 60k . up to 80 thousand, and there was a lot of enthusiasm around the crypt again. But that doesn't change the fact that we were promised a V from Coinbase , and we got a very long V that was actually more like an L.
And that's what I wrote here on the side. This is a risk for Anthropic as well—to incur a loss, as it will take them a long time to grow their ARR business and offset the increase in token costs. Because, as they argue, more growth and lower prices do not automatically mean success if prices fall faster than consumption.
This is a standard baseline scenario. Token prices are falling rapidly, and although consumption is increasing, you are selling them with lower margins due to pressure from Chinese models or simply less profitable tokens. So it seems like it's pretty much the same thing . Either the rate of growth in the number of tokens slows down, or the margin falls.
These are actually components A and B , right? Think for a moment about the Google IO event. Google IO: Wow, token usage increased 50x from 2024 to 2025. And if you look at the period from 25 to 26, the use of tokens increased by 6.9 times. So you have a 50x increase and then a 6.9x increase. This is still a huge increase in token volumes.
At Google IO, figures at the level of quadrillion tokens were advertised. But the second derivative is negative. The second derivative is the rate of change of the growth rate. The growth rate changed from a huge positive number to a speed increase that became negative. We went from 50x to 6.9x. This is a decrease in the growth rate, that is, the second derivative.
So, the growth rate may decline and prices may fall due to competition from models with open scales, which we can already see in this graph. That's why this is the most pessimistic graph in the field of artificial intelligence. Yes, right next to that promo code ending in mekevin.com. Many say that it can be written off as investment expenses, education, or however you declare it.
But consult your accountant about this. I am not your accountant. They say there is some growth potential. So we don't want to be complete pessimists. The growth potential, they claim, is that huge demand could increase token usage and margins, perhaps. Maybe software companies will be able to make profits faster, right?
A V- shaped graph can be steeper. Or, frontier models like Astra could encourage people to pay more for better computing power. Personally, I think the hope that better models like the Astra will push prices up is fading. You may only see a brief up-and-down surge. So, there is a certain unevenness.
Like what we saw here, but probably even less in new models in the future. So you really have high hopes that AI development will continue, frankly, with a margin large enough for Anthropic and OpenAI to continue paying their bills. This is the basic baseline scenario. And the more OpenAI expands, the riskier it becomes.
Does this mean it's all over? No. In my opinion, if it's a circle, it's a pretty bad circle. You know, it's such a vicious circle. But I want to make my point clear because people sometimes get confused. And that's okay. We will bring clarity here. I think we're probably in the hardware cycle for AI here right now.
Of course? This means that I really think we're in for another small rally. I call it Hardware Rally 2.0. But during this same rally period, I believe we are still moving through a software rally. This doesn't mean that all software companies will win. During the , I know that many participants ended up selling UiPath because we identified some red flags that weren’t immediately apparent in the earnings reports.
So if you're a participant in the course and you haven't seen that broadcast yet, I highly recommend you watch it, because instead of just looking at the latest financial statements, we looked at some other documents that indicated, "Oh, there are some errors here and a red flag in what's going on ." What will this mean in the long term?
We have expressed our thoughts on this matter. So, not all software will win, but I still think the software rally could remain long-lasting as long as the small hardware 2.0 rally continues. The only way the Hardware 2.0 rally will continue is if Anthropic tells us that despite this drop in token prices, they can still pay their bills and are doing well.
Or that revenue comes from software. Beyond that, it becomes quite bearish to see how close we are to the peak of AI trading potential. If Anthropic can continue to pay its bills, then Broadcom looks very cheap. So you have a sort of binary outcome here, right? If Anthropic issues an upbeat S-1 report in its IPO filings, that should make you optimistic about Marvell and Broadcom.
And AMD and Nvidia, you can add them to the list too. If this turns out to be a pessimistic signal in the S-1 form or Anthropic's quarterly earnings look bad compared to the previous ones , it would indicate a bearish trend for the entire stack and an approach to the peak. So, yes, that means that I simultaneously hold the thesis that we may still have a good six months ahead of us.
You know, I think hardware and stocks in general could hit all-time highs, maybe for the next 6 or even 12 months. But serious risks await us at the end of 2027 and in 2028, when the rate of growth in memory prices will begin to slow down and the supply of iron will increase significantly. This article further mentions—and I find this overly pessimistic—that H100 prices are also falling.
I don't think so. I believe this is normal volatility that you see relatively close to the peak. So I see this as more of a bullish signal. And this reinforces my argument that we still have time. Not forever, of course, but we still have time now. Here they note that JP Morgan data shows a 47% increase in token volumes, while cash spending has only increased by 7%.
In other words, this margin squeeze is bad news for cutting-edge labs. And even if the number of tokens is growing explosively, it is the second derivative of the growth rate that becomes a problem. They claim that yes, maybe Astra will give an extra boost, but let's be honest. Cloud inference, which costs millions per token, only works as long as the model is too large to run locally.
This is a critical issue, and that's why Dell is performing so well right now . Dell is so successful because the more you expand your on-premises computing equipment and create a data center in your office, connecting it to a 240 or 370 volt architecture, the faster you get online, because Dell brings and installs the rack, loads the software, and the switch is already built in.
There's a built-in UPS—this isn't a company burdened with debt and pension obligations—and battery backup . There is a GPU, a CPU. There 's a full stack there. There 's copper wiring, fiber optics, or whatever: Ethernet, InfiniBand. Everything is ready. Here's why Dell wins: companies want to have it on-premises because it really delivers a return on investment (ROI).
Many people ask, "Oh, Kevin, what's the payback on your own computing equipment?" She is. That's just the way it is. We save significantly more money than we spend on the hardware and computing power that would be required to run our Homes AI software in, let's say, our terminal, where we can, of course, change districts and move to other zones.
But what's even more interesting is that you can even see real estate properties that you will lose money on if you buy them. So we are expanding these opportunities. We still have a lot in the works , but, you know, we're releasing this data gradually because the system is still in beta and it's getting better every day.
This is very exciting. But understand: when we look at this, we run into a problem because Mac is now releasing these amazing devices with 256–512 GB of combined storage, which is , frankly, very optimistic for Apple. This could well increase their growth rate from the 9% expected by Wall Street to 20%, which again plays into Apple's hands.
I want Apple shares to be cheaper because I believe in their growth. They just seem a little expensive to me right now. If I were buying Apple at current prices, I think I would be better off investing in cybersecurity stocks at the same value. . But this indicates increased competition for the sale of the same marginal tokens.
The more powerful local computers become, the more large-scale models you can run that you would n't otherwise be able to run anywhere else. You know, I even have a little note about it. Where is she? Artificial intelligence. Let's take this note, because I think it's very interesting. Here, let's take a look at some of them.
We already know this is old news, okay? This is kind of my basic thesis. Tokens and infrastructure become commodities. We already know this. But I want you to see it—you can take a screenshot. This is a comparison of a local computer with a 5090 graphics card and a Mac. And while the 5090 may be 50% faster, you'll very quickly run into limitations on what models you can run.
But Mac Studio can run DeepSeek R1, Llama 4 , Qwen 3—a model with 235 billion parameters . The capabilities of the 5090 are limited to the Qwen 3 model with 32 billion parameters. So you really run into some limitations. And it allows you to do a lot more in terms of local agent reasoning, agents for programming, longer research sessions, file reviews for financial companies, better compliance, and reduced hallucinations in legal work.
You can do this locally. This is fantastic because now you are in control of the data and you don't have the hallucinations that come from running out of RAM. This is where they claim that these companies stop being token buyers and become hardware buyers . But they become equipment buyers only once.
That's the whole point. It's like when you bought local hardware, you don't have to buy it again. By the way, this is a bearish argument against Dell. The bearish argument for Dell is what the ultimate growth rate will be . Yes, we're on the rise now, but when people buy this hardware, what will the upgrade cycle look like, and therefore the growth, right?
And this is where Anthropic comes down to what they write about: sustainable commercial models start to resemble subscriptions to places, agents, and results. That's right. What have we been talking about here for a long time? Oh, the winners are value providers like software companies. That's right.
Thank you, Goldman Sachs. You have subscribed to Meet Kevin videos. Hell , they're probably fully subscribed to the Alpha report we send out every day. He's just been tearing up lately. Let's go. And this isn't just another win, but yesterday we predicted the rise of Tesla and software like Palantir.
Today, there is a setback in software , but growth in hardware, particularly with the forecast for a rebound in Marvell and AMD. There have been some really successful predictions recently. And yes, I'm not saying they're always perfect. Don't become overconfident, because mistakes happen. But I'm very happy with how our Alpha report has been performing the last few days.
So, looking at the rest, they further suggest that advertising could be the future. That's right. That's why Kevin believes advertisers are the sleeping giant of investment opportunities right now. There are three of them, and I have investments in these three companies. So I want to be very clear, okay?
I also have investments in Marvell. I want to be very clear . Broadcom is very tempting, but I don't own any shares. But I have positions in Meta, Apple, and Netflix because I think advertising is a sleeping giant right now. I think we can get into them cheaply right now, while, you know, software is already developing rapidly .
We also have a stake in Salesforce. But all these companies will wait. Even ones like ServiceNow. If you read ServiceNow's earnings reports, you 'll see that they complain about hyperscaler costs. They are declining in light of this information and what is actually happening. Also keep in mind that Salesforce just ran a massive advertising campaign and a million dollar giveaway.
I do n't know if Salesforce paid that million dollars, maybe they did. But they did it together with Mr. Beast. I think there will be more such cases from these service providers, because the fight for the market is ongoing. They noted that excessive demand for AI is not automatically a “bullish ” signal, as it is crowding out important AI projects like Anthropic and OpenAI.
And this is perhaps the reason why credit spreads have widened. This is an inverted graph, so it's a little harder to read, but the blue line is the stock prices, and the red line is the credit spreads. Since the red line is going down and the blue line is staying steady, you are effectively saying that the spread here is increasing, as is the risk.
So, the risk increases. This is the argument being made here. And of course, they sum it up like this: tokens are 30% cheaper, but volumes are up 20%, which means a drop in inference revenue. And I write in the note that this obviously destroys Anthropic, and then you face this V- shaped risk factor.
Maybe bots will save us, you know, bot-bot. But these are big risks. So it's worth keeping this in mind for AI-related trading. .
댓글 0