저자: Kevin · 외국인
요약 보기
엔비디아 실적 발표 앞두고 기대치 상회 가능성 제시, 마진 압박 우려를 언급하면서도 PEG 0.69로 저평가라고 주장. 브로드컴은 ASIC 수혜 긍정적이나 부채 우려, 마벨 선호. 추론용 ASIC 확산과 기업 AI 소형칩 수요 전환을 강조. 단기 조정 시 매수 기회 제시.
다듬은 글
The Wall Street Journal just reported that Anthropic expects potential revenue of $30 trillion in the future. This is like a complete reimagining of the total market, as it is effectively 100% of the entire GDP of the United States in 2025. Which sounds optimistic, perhaps, for Nvidia. And that's exactly what we'll be talking about in this segment: Nvidia's earnings expectations and Open AI's new "jalapeno" chip.
We have to discuss Nvidia robotics, Nvidia open scales, enterprise AI, GPU rental rates, what are the current trends, what Edgewater tells us about GPU and server demand, where is demand falling and where is it growing? Let's break all this down and get started. So, the first thing we need to realize is that Nvidia's report is tomorrow.
As always, after the stock exchange closes. We expect Nvidia to beat forecasts . Historically, they have always outperformed them. Current figures so you have them at your fingertips. You can write down the unadjusted earnings per share (EPS). So when the news comes in, you can run it through the Meet Kevin Alpha app if you want, or I'll come on air and we'll talk about that as well .
So, write this down. Expected adjusted earnings per share—$ 2.09. Nvidia has beaten expectations seven out of the last eight times. Expected GAAP earnings per share are $2.15. Nvidia has beaten expectations eight out of eight times. Expected revenue is 92.29 billion, with an exceedance of eight out of the last eight.
Bank of America believes it will be closer to $94 billion, and that's a big deal for Nvidia because even though Nvidia is outperforming revenue eight times out of eight , the last four reports have sent the stock down, which is an insult to this business that makes so much money and is so incredibly cheap right now.
This is madness . But, in the end, 94 billion is Bank of America's expectation. You really have to exceed expectations that are at the upper limit. So, the current average consensus forecast is 92.29. Typically, for a stock to go up, you have to exceed higher expectations. This is such a strange psychological game.
In any case , the expected net profit is $50.98 billion, which, if you do the math, means they are taking about 55.2% of the net profit from revenue. Of every dollar earned, 55.2% is after- tax profit. This is very good. So, the focus of this earnings report will be on margins, especially since we just saw a 15% price increase, which was actually less than Edgewater Research's projected 20-30% for the third quarter.
We have increased the price by 15% to offset some of the memory costs. We are also seeing some changes in the specifications and settings of some of Nvidia's advanced chips, where they now have less video memory (VRAM). So, essentially, the smaller impact of this memory problem, which makes you wonder: why are we creating " shrinkflation" in chips?
Why do we reduce memory? But maybe we just don't need that much memory in these chips. We'll talk about that in a moment when we discuss corporate AI, which I think is a really important focus, and it's really inspiring to me about the future and investing in stocks, because I feel like we as a startup are on the cutting edge of what corporations are doing with AI, which I think is fundamentally different from what leading labs are doing.
Anyway, let's talk about that in a moment. So, the " Jalapeno" inference chip was just announced by , well, OpenAI via Bloomberg. Bloomberg released an article stating that OpenAI claims that their new chips can outperform Nvidia processors in tests. This chip is called "Jalapeno". And it outperforms the GB300 in their benchmarking .
Obviously, this is biased, and I mention it because people perceive it as a risk factor. This "Jalapeno" chip was developed in partnership with Broadcom. Broadcom is a pretty interesting asset because they have a lot of debt. Their balance sheet is not great, to say the least, but the valuation is incredibly low right now.
They trade with a PEG ratio below one. We just conducted a detailed analysis of them always recommend, if you haven't already , that you consider joining a Meet Kevin membership where you can receive these analyses on a regular basis. We have a significant price increase at the end of the week, and we are changing the pricing structure, so existing members will receive a kind of bonus package that those who join later will not have .
But, for Broadcom, according to our Talk AI, the target price is $905. The problem is that we have a red flag on the company's balance sheet because they have too much debt. So, your balance is a red flag. The cost estimate is a clear green flag. They fell about 26% from their peak after the last earnings report due to ASIC chip delivery delays .
Okay, so how does this relate to OpenAI? Well, they are developing ASIC chips for OpenAI. That's it. And they are also dealing with Google TPU. So OpenAI is now saying, "Hey, we're getting ready to release these products." That's probably optimistic for Broadcom. This is generally optimistic for the entire sector.
But these ASIC manufacturers— think Marvell and Broadcom. Broadcom has better earnings than Marvell, but Marvell has a better balance sheet than Broadcom. Marvell is also a company with a market capitalization of $211 billion. It grew by approximately 5.6%. I have a stake in Marvell. I don't own Broadcom, but I really like this company.
I'm very inclined to invest in it now. A company with a capitalization of 1.7 trillion dollars. But these are your main ASIC manufacturers, essentially as inference tools. And that's exactly what OpenAI is saying here. OpenAI is telling you, “ Look, we’re not building this chip to compete with the cutting-edge learning technologies for the next GPT-5 , -6, or, say, the next Anthropic Fable, we’re trying to build a chip that will reduce our cost structure in providing data inference solutions, and we’re trying to get them wherever we can.” If we can get them through ASIC—great.
If we can get them from SRAM chips like Nvidia Groq, which was recently announced for ultra-fast token generation, with ultra-low latency, imperceptible to voice AI or customer service. This is an Nvidia product, they acquired virtually all the intellectual property from Groq, G-R-O-Q . This is another competitor to Cerebras inference chips, which are very good.
I hated Cerebras during the IPO. I thought $370 was too expensive, but now that the price has dropped by half, I actually think it's a pretty good deal. These are just short notes about what I think about all this. You see, there are a lot of benefits now from leading labs saying, " Hey, you know, we're no longer just focused on the best cutting-edge chips from Nvidia, like the Vera Rubin chip." We're also really focused on how we can fill our data output stack.
And Nvidia realizes this. And that will be one of the risk factors for Nvidia, and that's probably the reason why Nvidia is selling at such a low valuation. Nvidia's PEG ratio is currently around 0.69, and it's just an insult to Nvidia how cheap it is. That's why our AI for stocks indicates that the upside potential is about 4x.
And our target price for Nvidia is $886, uh, as a fair value , simply because they have such high earnings. So why are they sold so much cheaper? Like, what's going on? Well, let's simplify it and understand why Nvidia could potentially be selling at such a deep discount. OK. So, first , we have Nvidia.
Unfortunately, they have this " big company deficit." You know, it's $5.7 trillion. It takes a lot of effort to move $5.7 trillion in stocks. So that makes the situation complicated. That's one of the things that works against them. Another thing is that they are not the only competitor in the field of ultra-fast token generation, because there is Cerebras, which is actually ahead of them.
This just went into production with Nvidia, it's an Nvidia product now , so it should be successful, but Cerebras is ahead of them. Cerebras simply needs to ramp up production, which they expect to achieve in 2027-2028, probably around the same time as Groq. But the problem is that you have almost two equal competitors, I would even say Cerebras is ahead, while Nvidia has this big "moat" in advanced technologies, right?
Think about Vera Rubin space computing, which, to be fair, Elon Musk is talking about: Vera Rubin's space-optimized chips will be launched into space in the fourth quarter of 2027, and then scaled up in 2028. That's great, but it's a little unlikely. It's kind of your, uh, Vera Rubin for space, I would call it, uh, hope at this point.
It's still a little early, so it comes at a discount. Groq is discounted. Uh, their size comes with a discount . Like, can they really double in size from here, right? That's the problem. You know, 4x growth would require this to be a $ 20 trillion company. Based on how much money they make, this can be justified.
But whether you can actually achieve that is another question, right? Next, robotics is still very early. Nvidia just announced, uh, another, uh, kind of release in the field of robotics, if you will. This is the Jetson Orin Nano, a low- power edge computing robotics computer. This is the future, but robotics is, again, hope for the future.
It could be decades before we truly scale GPU-intensive vision-based robotics training . You're more focused on labs now, on edge computing, right? I should be careful with the word " lab" because advanced labs use a bunch of GPUs, right? Such as Anthropic and OpenAI. This should not be confused with laboratory robotics, which is in the very early stages of development.
Mainly because there are still issues with the actuators and mechanics to make these devices work longer than the 30- minute demonstration. So, whatever , you have these problems. Large size. Perhaps Cerebras is ahead of them here. I mean, they're already on the market. Cerebras has a customer concentration risk due to OpenAI.
Some of their IPO problems have been alleviated, especially now that the stock has fallen in price. So I like them more now than I did then. Vera Rubin is more of a future option. Robots are also a future option. So now a lot of it comes down to margins. And that's going to be the main focus for people: if you raise prices by 15%, and memory prices have gone up significantly more, how stable will your margins be before the earnings reports?
So there is some concern that we could see a decline in profits, especially as corporate attention shifts to inference chips rather than these cutting-edge chips. Think for a moment about inference chips. It sounds crazy, but a gaming PC running on a 5090 GPU is perfectly suited to powering many enterprise AI systems.
You don't even need to buy what we bought—the Blackwell 6000 . We bought, you know, a few of these Blackwell 6000 chips. We bought them when they were $7,000. We doubled the money invested in this damn chip . I don't even think the stock has doubled in the time it took the chip to double in price, right?
It's pretty crazy, but many enterprise AIs don't need that level of memory. You don't need 96GB of DDR7 stock. You can perform many enterprise inference tasks with the 5090, which runs at a third of the capacity of this device, with 32 GB of video memory. So what does this mean? This means that companies may prefer ASICs from Broadcom, TPUs from Google, which we expect to scale next year, ASICs from Marvell, or Nvidia 5090 or 6000, over the latest Vera Rubin chips.
I think that's why NVIDIA is starting to say, "Hey, we're going to reduce the memory on these huge, cutting-edge chips, because enterprise applications don't really need that much memory in these big stacks." So, there's a lot of rumors going around on social media: " Oh, they're reducing the amount of memory they put in the chips." This is shrinkflation, you know, they're trying to save money.
I think it's to keep the price low and make them useful for business, because the chips are getting too big. We don't need these huge stacks, because so much work can be done with these small chips. You can literally pick one up for $4,900, that seems to be the retail price. That's not even the wholesale price , is it?
You can grab an NVIDIA 5090 RTX chip right now. Oh, they went up another $200. Are you kidding? They literally cost $4,900. There's only one left in stock, at least in this store. But anyway, you know, just last week it was $4,900. What? And the shipping, you won't even get it until September? This is absurd .
Well, anyway, yes, I mean they are being snapped up. One of the reasons they are being snapped up is because you can take a company like Moderna, which is doing fantastic work using artificial intelligence to accelerate the creation of therapies for its customers. In my opinion, you could do what they do with AI on a 5090 GPU and you wouldn't need those crazy Vera Rubin processors.
I'm not...GPU. GPU. What I mean, and I want you to visualize this because I find it very helpful. Once upon a time, back in the days of Moderna, let's say , in 2022, well, 2022, let's call it that. It took about 60 days to sequence a cancer patient's tumor and then choose which mutations the therapy should target.
And then in 60 days they could affect maybe one to five mutations in these cancer cells. It wasn't exactly ideal. You want to affect significantly more mutations to get a longer-lasting result. I'll be making a separate video about healthcare and explaining that in more detail, so subscribe if you haven't already and you'll see it.
And now, with AI in 2026, you could do it in a matter of hours, you could identify up to 32 mutations that could be targeted as part of a personalized cancer treatment, and you could literally run it on a 5090 gaming GPU. The problem is, these gaming GPUs have lower margins for Nvidia than the big $4 million racks.
One of the reasons why this is so close to my heart, this example with Moderna, I don't want to sound like a hackneyed person , but that's exactly what we're doing with our Reinvest AI. This is our, you know, AI for real estate, where you go into a district and say, "I want the best deal in this area," and it sorts the deals by how much money you can make.
This is proprietary AI. Uh, our stock AI is proprietary AI. Uh, we don't create chatbots. We don't want to compete in this niche. We are trying to compete in the field of practical AI, just like Moderna. They're not trying to make you a chatbot, they're trying to give you practical artificial intelligence.
And their practical AI can run on 5090s. And you don't even need to stack them together. You know, there's this idea that you can combine four 5090s together and use them together. You don't even need to do this. Because you can simply serve four patients with a GPU 4090, instead of stacking four cards together to solve a single problem.
You can solve four problems. The bottom line is that corporate AI is moving towards smaller chips, not bigger ones. You know, Gavin just published a post about his take on this Pareto frontier. I think he's trying to sound fancy about how AI is distributed. He called this the Pareto optimality level.
You know, he was just on the All-in podcast not too long ago. But anyway, I think it was about 10 days ago when he was there. He says we should expect a “ Pareto optimal” balance of computationally efficient people, cheaper open source tokens, and cutting-edge tokens. Let me explain this very simply.
I personally believe, and this is again my opinion, that you can break down his words into a pie chart with three points of where the expenses go. So, you have labor, that is, people, human labor, human intelligence. Then you will have advanced tokens. And then you will have corporate tokens. I personally think that this, advanced tokens, will make up about 10% of the share.
The workforce will account for about 45% of the share. Corporate tokens—45% share. You know, probably somehow . And that would only be about 10% of the share. This is my view on the corporate sector . I may be wrong , but that's what I think, and that's what my investment thesis on ASICs is based on, not on trying to sell advanced chips.
My opinion. We'll see. We see many companies doing this. And the payback period for many of these chips, to be honest , is from 1 to 2 years. We calculated some of this receive our trading alerts, see the long-term portfolio, how we invest for the next decade, and how we allocate assets. But join us at meetkevin.com until Friday .
We have a big expiration there. But yeah, you know, we were calculating some breakeven periods and we thought, "Yeah, you know, our chips have actually paid for themselves." So, the idea of short payback periods makes some sense. The question is, how long will these payback periods be ? That remains to be seen, right?
These are spot prices. We are watching how spot prices are converted into long-term contracts. We'll see eventually. But Thomas Reuters does it too. They just announced that they are using Qwen, Alibaba's open model, and believe they are only 3-4 months behind the leaders for significantly less money, while retaining their own data.
So they can process data within the company. Moderna can process its own data independently. Re-invest may process its own data independently. We don't need to send all of this to the lab. The problem is that this is potentially bad for Nvidia's profitability in the future if they sell fewer high-end chips and more enterprise ones .
Margin compression occurs . These are some of the questions that people are concerned about regarding Nvidia. I'm not too worried about this for the next earnings report because I think Nvidia is very cheap, and I think the stock has already factored in a lot of this. Again, if the PEG ratio is 0.69, it's incredibly cheap.
Let's take a look at the financial figures so you can see for yourself. So, these are Nvidia's financials, and these are the numbers we have for Nvidia. The current growth forecast for August 25th is that they will have $9 in earnings per share at the end of the year, with growth rates expected to be 44, 24, 87 and then minus 22%.
So you see some slowdown in growth rates at the end of the decade. This averages out to a growth rate of around 33%, meaning they are currently trading at a PEG ratio of around 0.69, which is very cheap. This may be partly due to the projected negative growth in 4 years. This may be because advanced chips are becoming less critical, and inference chips like OpenAI's Jalapeño , built in partnership with Broadcom, are taking over.
This is an optimistic signal for a company like Broadcom, but keep in mind, and this is also on our stocks tab, you can see this , join us, join the courses. But that's literally from their income statement. We do a lot of research. We try to provide this every day. In the semiconductor sector, we have always said that our ASICs and TPUs have lower margins.
Therefore, as TPU usage continues to grow, there will be pressure on margins. Broadcom's margins are much better than Marvell's, but in my opinion, Marvell is getting closer to Broadcom's margin performance. Now, what are the things that are really bullish for Nvidia right now? Well, $500 billion in funding for data centers and a change in SEC asset securitization rules.
This is just madness. It's like 2008 when the SEC says, " Hey, you don't have to have 5% of your own money anymore if you're going to sell data center-backed securities, because we think this time is different." Oh, we're also going to relax the rules and say that you don't actually have to disclose what these securities are made of or what the usage rates are inside these data centers.
To me, this sounds like a repeat of 2008. I don't like it. I personally still believe you won't see 2008-style peaks until a crash like the one with Anthropic starts. We started this segment with the idea that Anthropic is projecting $30 in potential revenue. This is probably the best total market reach (TAM) trade-off you could come up with.
Well , whatever. These rules are positive in the short term: more debt, but also more spending. And this is where one of the strange ironies arises. A lot of people say, "But Kevin, credit default swaps are skyrocketing at these rates." "Isn't this a risk?" Of course, this is a risk. But there is an irony here.
Hedge funds that had exposure to Meta, Oracle, Google, Nvidia as Nvidia credit default swaps are also rising. Hedge funds that buy Nvidia, Meta, or Google and then hedge with credit default swaps are now effectively winning on both sides . Think about how strange that is. Typically, hedging means: if Nvidia shares rise, the put option should fall in price.
But hedging works differently here. In this case, when Nvidia grows, the credit default swaps that serve as a hedge also grow. So if you own Nvidia stock and credit default swaps, Nvidia goes up because more debt is being injected into the system, stimulating spending, and Nvidia stock goes up. And this additional debt makes people nervous, so credit default swaps also go up, and your hedge actually goes up in price, which would be your insurance against Nvidia defaulting, right?
So, ironically, we are now in a situation where hedge funds can literally hedge and see their hedges and stocks go up at the same time . If Nvidia, Oracle or others go down, these CDS, paradoxically, could still go up. So it's kind of a perfect hedge. Stocks fall—your hedge grows. Stocks rise due to increased debt financing—your hedge goes up again.
This almost forces you to only buy credit default swaps. I mean, maybe it's not a bad idea. Of course, these are derivatives based on trading. That's a slightly different story. But think about the other factors playing in Nvidia's favor. SpaceX is spending $80 billion or spending $ 80 billion raised through an IPO.
They will probably raise another 300-400 billion, Nebius has 8 billion in cash. Google just raised $86 billion in capital spending. Anthropic's IPO is approaching. Yes, everyone wants to grab a piece of the pie to get those chips. So, in my opinion, we will see more of a shift towards production chips and enterprise chips compared to advanced chips.
But so far, Nvidia looks pretty undervalued given all of this. So, let's look at some of our findings. And let's take a look at some research from Edgewater. So, this is a series of server-based studies for Nvidia and CPU . GB200 and 300 were expected to grow by 20-30% in the third quarter. They actually grew by about 15 %, which is slightly below expectations.
Server CPU shipments are expected to grow by 50% in 2027, which is simply fantastic. Here it is . Nvidia or Intel's CPU production capacity is sold out by 2027. Or until 2027, as it says here. Blackwell's product prices are expected to increase further, but this has not happened yet. This just happened.
This was posted a few days before the price increase. The cost increase is mostly memory-related, as GPU prices have remained flat and Nvidia has ensured sufficient supply, so we expect all attention to be focused on margins. Have you raised your prices enough to compensate for the rising cost of memory?
This is what will be the focus of significant attention during Nvidia's reporting. We will cover this tomorrow after the market closes. This idea of shrinking packages, many say, is just memory inflation. I personally also believe that this is simply a way to stay competitive in enterprise AI pricing.
We touched on this issue. We have the RTX 6000 here, which I just showed on Amazon. Its suggested retail price has increased from $8,000 to $11,000. So you're raising prices by 37% on enterprise chips because that's where the demand is, which is pretty impressive. I agree with that . Channel demand for servers is falling, and demand for gaming solutions is weakening.
Gaming and PC shipments—it's no surprise that we see weakness there. I find it interesting that Edgewater literally says, “We don’t know how customers continue to find more and more budget for AI servers .” However, standard computing servers are in decline. So, I think when they say here: " Channel demand for servers is starting to fall." I think they mean traditional servers, because AI servers are still growing rapidly.
As they say here, standard calculations—our clients only upgrade when they have to. So this is your classic cloud computing, not AI computing. Other notes they make are that they expect DRAM prices to cool in 2027. This is also not surprising. We may see prices rise. So, this second derivative, the growth rate, becomes sharply negative by 2027, when DRAM prices are expected to grow only 0-5% in 2027, depending on the memory type.
Whereas at the beginning of the first half of 2026, they all grew by 60-120%. Most of this is expected to cool. Edgewater also claims that Nvidia may be starting to feel some pressure from open-source models, which is why they are spending $6 billion to partner with Poolside to launch their own open- source models, but they may also be trying to compete with ambitious projects like what they mention here, the open-source Chinese LLMs.
I mentioned that I think there are some cost savings happening right now. So why should we overload these chips with a bunch of memory if people don't need it? Let's reduce the memory, lower the price of these chips, give people these GPUs, spend less on memory in the overall cost, and let them implement their enterprise AI.
Again, the question is , what will the margin be in the end for Vera Rubin compared to the RTX 6000? And that's why we just saw a 37.5% increase in the recommended price for the RTX 6000, because that's where corporations are going right now. And this is quite logical. We see the same thing. And I always like to share observations like, " Hey, this is what we're seeing in this area." You know, when we were buying a bunch of processors to process our data, we ended up mentioning AMD in our streams.
We said, " Well, this will be good for AMD." And this thing just flew in half. So sometimes we can predict trends a little bit just by being "in the trenches" of AI software development , and that's really cool. So, what else? This is essentially basic information about Nvidia. Again, they are expected to beat forecasts , as they always have, but that is not what will affect the stock.
The stock will be affected by what we hear about margins, the introduction of the latest chips, competition from Open Weight, and the less profitable corporate segment. These are the " red flags" that we will pay attention to. Are we currently concerned with the Space or Groot project for robotics or the corresponding chips?
Not particularly. Will Grok really be better than Cerebras as a logic output chip on SRAM? We don't know. We'll see. It's still a matter of time. All these are just secondary things. The gaming segment is expected to continue to decline. This is also secondary. So I'm not thrilled about worrying about these issues.
It all depends on the cost of advanced technologies. What is truly optimistic is the availability of significant cyclical funding. A lot of money is flowing into Anthropic's promising IPO. I think once the S1 comes out they will raise a lot of money and just spend it on chips. Even on their own chips or on Neo's cloud services , like NVIS or something like that.
We will also monitor Nvidia's investments in other companies. You will not see this as capital expenditure ( CapEx). This is not shown as CapEx , because, you know, capital expenditures are shown here, as part of fixed assets. In fact, you'll see it a little higher—in the line for purchases of non-marketable securities, as well as marketable debt and equity securities.
And you see that there's about $26 billion in investment spending so that those companies they invest in can buy more Nvidia products. So this is where the accusations of " circular" deals come from. So, we will be monitoring this as well . But tomorrow is an important day, and I think many in the market are feeling a little uncertain and nervous.
I actually think this creates an opportunity to buy on the dip today and tomorrow, because I don't think we're going to be on the brink of a recession anytime soon. So, I think there is an opportunity to buy on the dip over the next few days. We'll see. And when we buy on the dip, we will send you a buy/sell alert. Make sure you a lot. .
Comments 0