저자: Kevin · 외국인
요약 보기
AI 기업 Anthropic의 'API 중심 수익' 및 학습비용 제외 이익 지표가 과도하다고 비판. 학습비용은 매출원가로 봐야 하며 경쟁 심화로 API 성장 지속은 불확실. 엔비디아는 PEG 기준 저평가로 보이나 성장률 둔화 시 밸류에이션 왜곡 위험, AMD도 유사 리스크. 학습비용 증가는 반도체(커스텀칩·ASIC: 브로드컴, 마벨, 엔비디아, AMD) 강세 요인으로 봄. 대형 장비주 비중 축소, 소형 장비주 선호, 기업용 소프트웨어(Salesforce, UiPath 등) 가격결정력 강조.
다듬은 글
In this video, we'll look at the dangers of Anthropic's earnings forecasts, why there's risk in those forecasts, and what it means for other stocks, including one very well-known but incredibly cheap company we talked about this morning during a live stream for course participants. We'll touch on that here to understand why this stock is so incredibly cheap, and it's one of the biggest names in artificial intelligence, but there's also a baked-in red flag.
Let's get through all of this as quickly and efficiently as possible. First, there are Anthropic profit and loss forecasts that people make up. And all this is related to profit before its appearance. This is a new metric that people are making up, and they're essentially saying, " Hey, uh, instead of assuming that the cost of training our AI models is part of the cost of goods sold, let's just adjust that and then say, "Hey, look, we're actually a profitable company if we just ignore what it took to get here." It's a little crazy because you need to keep learning to stay on the cutting edge.
Otherwise, you will no longer be a cutting-edge laboratory. So, in my opinion, your AI training costs are absolutely part of your current cost of goods sold, because without it, you'll just fall behind and other companies will take you over. This creates some really big problems, in my opinion, for anthropogenic indicators.
But first, let's prove why this training is so important. Let's go back to this from Mark Zuckerberg. Mark Zuckerberg from Meta told us that we now have our next set of more advanced models and training. And this is where he is asked about capital expenditures. And this work, I think, will simply continue.
I mean, it's a cycle. I don't think we'll be done with this anytime soon . You are simply constantly focused on training smarter and more efficient models . This is a really important lesson. And let's try to break it down simply. So if Mark Zuckerberg says, "Hey, we're going to keep training the model because in the case of Meta, we want to get ahead." Then that means that in the case of Anthropic, to stay ahead, you need to keep training the model, right?
So actually, as part of these big AI companies, we really expect the learning to continue. So they can stay on the front lines because there is something they want. OK? So training, uh, like cost of goods sold. We'll just put the gears here, right? Cost of goods sold is like the gears in a machine. Training as part of the cost of goods sold, in my opinion, continues.
Even when you dig into the potential situation, I'm not saying it's going to happen, but even if you go into the hardware market or into some AI market, it's very likely that companies around the world will continue to learn and learn and learn because they'll want to use this market downturn as an opportunity to gain market share.
Remember that not every AI company is public. If public stocks are falling and AI is entering the market without AI, there will be many startups that will see this as an opportunity for growth, and they will continue to learn. I also think that companies like Meta, which are huge, and it's not the big company that we'll talk about in a moment, will continue to spend money on trading because they, again, even though they're public and they'll see their stock value probably go down in the market without AI, will want to get a share because what do people want?
What do these companies want? They want APIs, uh, habits. So I'll call it the API habit. I'm talking about API habit because what you're really trying to do with an AI model like LLM these days is have people use your model by default. Whether it's through Palunteer, or through Perplexity, or through Microsoft, you know, Copilot or something else, it doesn't really matter what platform you use.
When you choose your model, Enthropic wants it to be Enthropic, Meta wants it to be Spark, uh, you know, OpenAI wants it to be GPT 5.6, which we have some updates that we're learning about now. But this is where the money is. How do we know that's where the money is hidden? Well, because take a look at those predictions right here .
If you look on the left, you can see consumer subscriptions, business subscriptions, enterprise subscriptions, and API revenue. Essentially, this is where developers integrate models, right? You can see that according to the first quarter of 2026 data, all other sectors combined–business, consumer, enterprise–accounted for about $742 million in revenue, but API was worth over $4 billion, meaning that about 80% of revenue comes from API.
And if we look here, let's see how these numbers fit together in the year-end forecast. At the end of the year, 26863, we have about 87.7% of the revenue that is projected to come from API. APIs or the habit of using these models in your software, whatever, that's where the big money lies. And these projections assume that Anthropic will take its explosive growth in the first quarter and continue to operate without competition.
And this is where things get a little funny, because if you actually look at what competitors are doing, they realize that's where the money is. And here you go. Just today, Meta is releasing Muse Spark 1.1 in preview, open to developers. On Thursday, Meta provided long- awaited developer access to its Muse Spark artificial intelligence model, along with an updated version that directly pits it against anthropic and open artificial intelligence (AAI).
This is part of their goal – to produce superintelligence. And Meta says the updated model can write code, debug code, use software, external tools , understand text, images, and perform multi-step tasks with less human interaction, more agentically. Is n't that right? This is literally what Anthropic is doing with Claude, Claude Co-work, Claude Desktop, Chrome extension.
That's pretty good. Claude won because they were first, but that doesn't mean they will continue to win forever. And competition from OpenAI, uh, Google, and, uh, of course, Meta is coming. Open AAI told us the following this morning. Sam Allman was on CNBC, and the notes I took are that the cost of compute and memory, the cost of those who are trained, and the logical conclusion is that it's a headwind to profitability, which will be true for all companies.
Um, but they mentioned that their current and latest model is 54% more efficient in terms of tokens on genetic coding. The more effective these models become, usually because of that good old Javon paradox, the more people want to use them. The question is, how long, or how much at all, can these LLM companies actually make a profit?
And in my opinion, the only way to pretend that you're making a profit on these LLMs is to assume that this API growth here is going to go on forever. And I think that's a mistake to some extent, given the competition. I think this is a red flag. You see, the Wall Street Journal supports an increase in revenue from 4 billion to about 10 billion from the first to the second quarter.
The person who wrote this chart then extrapolated this, essentially, by a factor of two and three . And then they break down the training costs into what is actually needed to remain a leader. And they're projecting that training costs will actually go down from about 50% of the costs they incurred in the first quarter to API costs.
And those costs will go down from 50% to about 40% here. So they think that over time everything will become more efficient, and that these costs will gradually increase. They just added the quarter's bill here, something like your, uh, how is it labeled? other operating expenses. So your SGNA costs here and then your logical-logical costs–that's really the only cost of goods sold that they're listing here, and that they're scaling, uh, roughly in line with the same percentage, uh , uh, API revenue growth.
The point of all of this together is that in order to show that this company is profitable on EBIT, uh , you know, before teaching profitable here or getting breakeven profits, we have to assume that this API leadership at Enthropic will last forever. And I think that's a risk because I think whether it's Meta or Gemini or open AI or whatever, I think the risk is that API revenues won't grow vertically.
In any case, companies will have an appetite to spend less money on AI, and they are going to use competition as a weapon to do so. Companies don't want to endlessly spend money on their developers by spending endless tokens. remember, if the goal is to get profitable API calculations, and in order to get API calculations, you have to be on the front lines.
That's when it becomes really interesting to continue investing in training, especially if the market starts to look somewhat negative towards training. And right here is this little $ 5 trillion company called Nvidia, which has a gross profit margin of 75%, which is fantastic, and a net margin of 71%, which is just crazy.
They print money. This is an example of a money tree. For every $100 they sell, they literally put $71 in their pocket, and they don't even know what to do with it. That's so much money. It seems like they don't even know what to do with them. This is incredible. But there is one interesting problem here.
Some will look at this and say that's a red flag. Others will say "no." You know what? Nvidia could be really interesting compared to AMD, whose stock is up a bit right now . Partly, well, because Nvidia is so insanely big. But look at these predictions for Nvidia. I will write them to you now. Wall Street is now predicting Nvidia through 2030.
This is the average of Wall Street consensus estimates. And keep in mind that Nvidia's valuation is now, based on P/E volatility, the lowest since 2019. But if you actually look at the growth forecast, we're going to take $8.99 EPS at the end of the year, 43.9% growth next year, 26% after that, 81.6% after that.
It fluctuates like crazy because analysts' forecasts are very contradictory. Then you'll have the first average negative year in 2030 , that's their 2030 or 2031 fiscal year. 5266 divided by 4 is O, okay. No problem. We will calculate it like this . This is an average, uh, growth rate. Er, 43.9. Let's do this right.
Sorry , friends. Okay, here it is . So, this is the growth rate. 43%26%81%and minus 20%. If I add them together, I get 131.5. 131.5 divided by 4 is an average growth rate of 32.8, about 87, right? This is the average growth rate now. Nvidia sells for around $200. At a P/E ratio, they sell for 22.25.
22.25 divided by 32.88 88 puts this company right at the PEG ratio, uh, at the end of the year, the PEG ratio is 0.6767. We mentioned this morning that this is like a meme on the live stream of course participants. It's very, very, very cheap. This means that Nvidia, based on this average growth rate, could actually justify the current price, uh, by the end of the year.
Uh, this is getting a little crazy. Okay, are you ready for this? Uh , but they could potentially justify the $795 price tag . 795 dollars. How crazy is this? Let me just check if I calculated correctly. And multiply the current price by two. Yes, $790. Here it is. So, $795. How did we get to this point?
We arrived at this assuming that Nvidia's growth rate would remain at around 32%. But here's the problem. That's why I think Nvidia is getting a discount. I don't think Nvidia can sustain a 32% growth rate if it really starts to turn negative. Because if the growth rate here goes negative, and then what if the next three or four years are negative, then Wall Street starts, like Micron, which also has a PEG ratio of close to one.
Wall Street is starting to take into account that these numbers will change, and this creates risk factors for the market. However, the market is still supported by the belief that API costs will skyrocket, which, yes, is likely, but it will likely be spread across other models, not just man-made ones.
Now, how does this compare to AMD? And then let's take stock, because we're giving away a lot , and it's time to take stock. And know that this doesn't mean we need to post another coupon for coupon on meetke.com. This means we should say that AMD is worth somewhere around $520. Er, the projected earnings per share (EPS) is around 720, which means we are currently trading at around 72 on a P/E basis.
The projected growth for AMD is better than Nvidia's, which makes sense. 81.5 34.8 14.9 and 27.6 divided by 4 gives me approximately 39.7. 7222 divided by 39.7 equals 1.89. This gives me a preference for AMD of about 42% multiplied by 520 of the stock price. Um, AMD's fair price right now is around $740.
Why, why could this be true? Why might this be wrong? Okay, here you have to understand that the PEG ratio assumes that growth will continue in the future. We're using four years of growth, but if the markets expect, and I think this is important, if the markets expect the next four years of growth to look something like this, right?
Well , maybe let's picture it a little more like you had this initial crazy boom that you're seeing now from memory, and then you have this growth. Okay, let's assume the markets assume this is a 4-year snapshot that we have here. But then the markets think that the growth rate will fall. If this is what the markets are assuming, then the PEG ratio would look very cheap for both of these companies.
And that will make them insanely expensive when you go to, say, the green field and take a snapshot of those four years, because it will just be negative and your P/E or PEG ratios will look infinite and you will be at a loss. Right? This is the problem with the PEG ratio. So why are chip companies so risky?
Chip companies, even if they look cheap today, including memory stocks , are risky because we don't really know what the growth rate will be. My point, my opinion is that I actually believe that recession or not, training costs will continue to grow, whether it's startups or other companies, because people are going to see this as an increase in market share for a very long time to come.
So, let's summarize a little. In my opinion, this supports the bullish sentiment towards custom chips and chips that can be essentially trained. Good. So, the bottom line, uh, is learning, betting on learning, let's say , betting on continuing learning — it's a bullish pick on custom chips, whether it's AS6 from Marll, Broadcom, uh, or Nvidia, AMD, right?
These are your big ASIC manufacturers. You also have a CPO here in Marll. Broadcom prefers copper for in-rack chips. This is really a topic for a separate video. Uh, although Broadcom also has some, uh, fiber optic, uh, work. And then of course you have the traditional, uh, Nvidia, AMD. If the focus on training continues even during the recession, then the growth of these companies could continue.
The question is what these growth rates will look like, because they dictate the valuation. Now my thoughts on how to invest in this. Uh, number one, you want to avoid anything that becomes commercialized. Er , that's just a memory for now . I think there is probably another benefit to having a lot of hardware and memory, but I would avoid the commercialized sector.
Secondly, I would have some experience with hardware, but, uh , on a 10-year commitment. So basically, I would commit , hey, I'm going to, I'm going to have some experience with hardware, but I, you know, if we go through a recession, I know it's going to collapse. I will have enough courage to endure it.
Another thing that is potentially worth considering is investing, uh, in the software level of a provider of commercialized LLMs. LLM, whether it's Spark, OpenAI, uh Anthropic, or the Cursor product, XAI, whatever, they'll all be commercialized, in my opinion. They are all relatively on the same level, and we will just try to outdo each other by spending more and more money on education.
The company that wins, or the companies that win, are the companies that have users on their platform. These are companies, like them or not. These are companies like Palunteer, Meta, Microsoft, maybe Apple if they really catch up. These are companies that actually have users. So while I strongly support Nvidia's current valuation, I understand that these PEG ratios can be misleading.
And so for me it's more desirable to think about where the money is actually going over the next 10 years. I mean, honestly , companies that have users, whether I like them or not. Now, Salesforce would be another example. It is also worth mentioning UiPath. They have many corporate clients. The coolest thing about Path is that they combine agent-based LLM with robotic process automation, which is algorithmic.
So, you have algorithmic and agent-based automation. And when you combine them together , you get something like Palunteer, right? And not everyone can use Palunteer. So you're stuck on Path. This doesn't have to be a Path plate. I know it sounds like that, but these are companies where the users are.
Uh, it could even be those software stocks that just didn't get anywhere . That's where the users are. This is probably the level where you have pricing power. Pricing power exists where there are users. Users don't like to leave. They don't want to pack things up and change everything. People don't want to switch from Android to iPhone.
Uh, you know, some people want to, but I'm just saying that once you're in it, you don't want to go from iPhone to Android. Once you're in Palunteer, it's very difficult to pack up all your junk and start over . Once you've set up your Salesforce junk, it's very difficult to start over somewhere else.
The same goes for Path. The same goes for your QuickBooks, right? Smaller people who have, for example, one W2, when there's nothing to do, yeah, they're going to go there because you don't need it. But that's okay. In any case, the money is not there. You want to invest where the businesses are. So, we're talking about corporate mobility.
Corporate mobility, or even, you know, mobile phones, because Apple's transition to Android is a transition. But enterprise mobility, uh, provides, uh, lack, provides pricing power, or for consumers we'll call it mobile. Okay, that's what we 'll call it. So I'm trying to reconcile this whole video, because a lot of things were discussed here.
Anthropic numbers, these predictions are a little crazy in my opinion . They may happen, but they don't take into account how strong the competition will be with API tokens. It just draws a straight line and says, "Okay, S- curve here." It will grow forever, and EBIT shows excellent results.” This is a risk.
I don't like LLMs. I also think that LLMs will remain unprofitable because, in my opinion, the cost of training at these companies should continue to rise. The learning will simply continue. As Mark Zuckerberg says, it's a cycle. We will keep pushing, pushing, pushing. Even in a recession, we want market share because a recession won't kill AI.
A recession will kill the stock market. That's all. Recessions kill valuations, not innovation. Oh, that's a phrase . By the way, Max, uh, I would really appreciate it if you would comment on that duck we wanted on the claw machine at Dave and Busters. And he was very happy about it. Uh, but anyway, yeah, Nvidia is making a ton of money, and this company isn't going bankrupt.
They have so much money. They would have to do something stupid to go bankrupt. But, but, its valuation, although it looks cheap today, can very quickly turn expensive, and that's why I personally no longer have any connection with the largest equipment manufacturing companies. I made a lot of money at Nvidia.
I have made money in many equipment stocks, but I am moving away from large equipment stocks to a few smaller companies. We have a few small equipment manufacturing companies that I really like. Uh, we added one recently and one a few months ago, and I really like them. I still think they have advantages.
I think they have advantages over the next decade. I'm very happy with what they're doing. But my experience with hardware is much more limited than that, which is where I think pricing power remains. And this pricing power, even if the market doesn't currently factor it into the price, is where these customers are.
Even if all the hype right now is around the LLM model, it will become commercialized, or around memory, which will become commercialized. And big profitability , that's where Nvidia's momentum should be. And it's not because the market sucks at the hype nonsense . But all this will turn 180 degrees.
Uh, so you need to think about positioning for this now. So I'm saying this clearly so that I don't get comments from people like, "Oh, well, but Kevin, why didn't you tell us about this before it happened?" And then I'm like, "Where have you been?" This is normal. These cabins are only one in a thousand going anywhere.
And many of you, I think, really like this content. I hope you enjoy this video. .
Comments 0