저자: Kevin · 외국인
요약 보기
Nvidia가 Groq의 IP와 핵심 인력을 약 20억 달러 규모 라이선스 딜로 확보한 것을 대형 호재로 해석. Groq의 SRAM 기반 저지연 추론 기술을 CUDA에 통합해 추론 시장을 장악하려는 전략이라고 주장. LLM은 commoditize되고 AGI는 실현되지 않을 것이라는 전제하에, GPU를 쌓아둔 Oracle·CoreWeave·Nebius·xAI가 '봉지 홀더'가 되고 Nvidia는 마진을 지킨다는 시나리오. 목표가 $300 제시, 다만 Sam Altman·OpenAI 자금 리스크를 주요 우려로 언급.
다듬은 글
This changes absolutely everything for Nvidia. This is remarkable, especially with Nvidia trading for just a $1.67 peg, suggesting that Nvidia could be worth over $300 per share. Now, in this video, I'm going to break down and reconcile why would it make sense not to hold Nvidia, what are the big downsides, especially after this massive $20 billion acquisition.
But why did Nvidia actually make this acquisition? What's the play here? What are they getting at? And importantly, what does it mean for companies like Coreweave, Nbby, XAI? What does it mean for these other companies that have been hoarding Nvidia chips? And better yet, did Nvidia hint that this was exactly what was going to happen by themselves getting out of the data center rental or leasing business.
Ah, a lot to break down in this video. Let's get into it. A quick note, obviously, I'm not in my studio. Uh, I saw the snow in Tahoe and I'm like, Jack, let's go skiing. I asked Max, too, but he didn't want to come. But anyway, we're gonna be skiing this weekend again. All right, so let's get into it.
So, obviously, Grock, some people pronounce it grog, some people say Grock. I'm just going to pronounce it Grock. That's not to confuse you with X AI's Gro or Elon Musk's Grock, the chatbot. This is a separate company also called Grock spelled GR OQ that is a chip designer and more importantly a software designer for those chips.
Keep that in mind. So this is a $20 billion deal. I wrote down some notes to keep this smooth for us. They were valued at $6.9 billion in September. Now they have a $20 billion essential buyout deal going on. It's not a buyout. It's a licensing deal. They're doing that to skirt and like regul play regulatory arbitrage and get around uh merger and acquisition uh antirust uh laws and blockades and two-year approval processes.
So Nvidia is just basically like, nah, you guys keep doing your data centers. We're just going to take your top talent and your intellectual property and we'll give the company $20 billion. So that's good for the shareholders of obviously Grock, but it gives Nvidia exactly what they want. And you'll see in just a moment why this is so important.
I'm I'm personally very excited about this. I I when we had our course member live stream this morning, I'm like, this is bullish Nvidia. I don't know how long it's going to take for the market to sort of digest it, but this is overall bullish for Nvidia. That was part of our alpha report this morning.
If you're not a member of that yet, go to mekevin.com. You can see that, see that analysis. and of course get my top 10 stocks to buy for the next 10 years. Now, Grock chips. Now, here's the key. A lot of people are going to say this is a play around memory that Nvidia's GPUs utilize high bandwidth memory, which is really just stacked up DRAM memory, which that's kind of redundant because the M and D RAM is memory, but anyway, just to make it simple to understand.
Um, Grock chips use a different style of memory. It's more expensive and you actually need substantially less megabytes compared to gigabytes of memory to process what Grock does well. Now, before I really get into the technical weeds, I think it's worth understanding what Jensen said they're planning on doing and what the CEO of Grock, the founder of the company, said their intentions are.
So, Jonathan Ross, the founder of Grock, who worked on TPUs at Google, was a contributor to the invention of TPUs at Google in 2020 in 2016 and then moved on. He argues uh that these chips are really designed to be the low margin inference computers for the artificial intelligence wave. And he made this pitch and this argument possibly in a way not to piss off Nvidia to say, "Oh, we actually help Nvidia." And there are many interviews of him saying we help Nvidia because Nvidia can sell their high margin GPUs and we'll just take the inference market while they take training and inference is the low margin business.
Now I'm going to get to why this was a perfect setup right now. See the CEO of Grock never said anything bad about Nvidia. instead sold his product as something that could actually help Nvidia. And I think that is exactly what set up the acquisition. Now, why is this acquisition so brilliant by Nvidia?
And then we can get into some more of the technical weeds. Here's a simple way to compare the two chips. This is the simple part. We're going to leave some of the detailed weed stuff for a little bit later in the segment. So the detailed weeds separate right now. Overview GPUs great at training models.
Think uh all like memorizing the entire internet and trying to generate some kind of world view that we can then ask questions about and generate new answers or new ideas or new images from. Genai. TPUs from Google have been an attempt to do something similar at a cheaper cost. Basically cutting out Nvidia's margins.
That's why Google likes their own TPUs. That's why Amazon is making their own chips. There's an incentive for companies to make their own chips to really cut out those 75% gross margins that Nvidia has. Now Gro is really or Grock is really interesting because Grock doesn't prop to say hey we're going to help you train this worldview model instead Grock comes in and says hey we're going to help you where people actually utilize artificial intelligence and that's in communicating with AI bots customer service agents live translations through like metagasses and what you'll find today is people meme AI chat bots with how slow they are.
If you've ever gone on TikTok and you've seen people joke with the the sort of like uh talking to GPT in real life, people be like, "Oh, uh, hey GPT. Uh, can you let me know where the closest pizza restaurant is?" Ah, cool. You want to know where the closest pizza restaurant is? Are you looking for Chicago style or more of an Italian style?
You know, it's it's you know, the latency is the awkward part and people joke about that. Pilots have made jokes about having AI as a co-pilot and you'll get something like GPT gear downoop. Ah, I see. Uh, so you're requesting that I put the gear down. Uh, is there a reason you want to put the gear down now compared to later?
That just put the damn gear down. You know, the captain then freaks out, right? That's this sort of joke. But that joke really helps us understand where the problem in AI sits today. Now, I'm going to inject my opinion here for a moment. I've been providing in my opinion facts and data and kind of trying to build this story up because it's a lot.
It's a lot to digest and so we're going to take this elephant in small bites together. That's how I like to teach. Uh so I want you to think for a moment of uh overall when you ask a chatbot a question uh you're looking for a rapid response but you don't really mind waiting for a moment because it's a chatbot it's in your app.
But what happens when you're on the phone and you're selling customer service that seems humanlike, but it has a massive delay. And that's where that human effect from that humor really highlights the problem with delays. And that gets to my opinion and it has to do with the commoditization of AI. See, watch this.
I don't believe that generative uh AI or any of the AI we have today is going to lead to artificial general intelligence. So to understand my conclusion here, you have to understand that my bias is that AGI is bull crap. It is this moonshot idea that helps people raise money for their startups, but artificial general intelligence is not where we're going.
That's not where the money is going to be made. So you have to understand that first. Two, not only do I not believe we're going to have artificial general intelligence, but I also believe, and I've been saying this for years, that language learning models are going to be commodities. There's nothing going to be in the future.
Nothing will be unique about using Gemini or using uh uh Twitter's Grock, you know, Elon's Grock or using Llama uh or Anthropics Claude, whatever. None of that will matter. They'll all give you probably within one percentage point the same answer. In which case, if they commoditize, then we already have a risk that we've overbuilt our GPU capacity.
Right? This is why I have the opinion that in the long term Oracle, Coreweave, NBIS, these companies are in the long term untouchable because they're bag holders of GPUs that Nvidia wants to sell and not hold. This is partly why I think Nvidia got out of the chip holding business. Now, they've made some creative deals with Coree to lease some of those back and that creates some circular questions, but ignore that for a moment.
So, no artificial general intelligence, commoditized chat bots. Got it? So, then really the companies that win are companies that utilize the chat bots along with their own custom MLS the best. Those could be aggregators. Those could be aggregators like UIP path for healthcare and finance. They could be aggregators like uh you've got the ontological setup where you kind of structure data over at Palunteer and then aggregate which LLMs were going to use to create or decipher conclusions from this big data helping basically big corporations get into big data.
Right? Palunteer where UiPath might appeal to sort of everyone. Palanteer sort of applies or appeals I should say to the richest companies and the rich government. Okay. So they've really niched into the expensive customer cohort whereas UiPath is maybe more of the discount version player, right? The we're here for everybody to try to set up these AI agents for everybody if you will.
Salesforce is in that boat as well. Okay. Uh so so then you have aggregators. Okay. So artificial general intelligence don't see it happen. Uh LLMs are just commoditized encyclopedias. Think about it like Wikipedia in your eyeballs, okay? You've got Wiki right in front of your face all the time. Uh and then on the flip side, we have aggregators that can make money or people who figure out how to actually make the AI technology we have today productive for mass industry.
This is where obviously I'm biased, but I think it's a great example. I like that we are creating real estate AI that does not rely on artificial general intelligence. It actually relies on our own machine learning training and tuning partly in combination with LLMs which are commoditized. We love that they're commoditized because they reduce the expense of using LLMs.
But our ML is what gives us proprietary data and skill and technology at House Hack and Reinvest, right? you know that already. We have a fundra for that closing at the end of the year. If you want to buy the AI product, which releases next week, you can do that . Yes, we are ending the 5% yield uh starting December 31st.
Now, if you get in before that, you still get the 5% yield through conversion. Obviously, it's a real estate backed company and we're launching our AI based on the skills that we've been building and the MLS that we've been building since 2018. Before the AI boom, we have been working on our MLS and so we're so excited to bring these to mass market for agents and otherwise.
So, you have to understand this baseline. Now, where does Grock come into all of this? Where Grock comes into all of this is speed. They eliminate the latency of live AI translations for language. This is one of the reasons why I think Dualingo has kind of been tanking a little bit because yes, you can learn a language in a fantastic manner with Dualingo.
But why do you need to learn a language if you could just put on AI glasses or use your phone or use AirPods like what Apple is doing now saying these are, you know, simple. Put them in, two people put them in, what do you got? Live translations. any form of latency kills the functionality of that. Gro Gro's product solves that and they solve that by using SRAMM on the actual chips.
Now, that does end up leading to a larger footprint for each chip, which isn't ideal, but it solves that problem by embedding the memory where the compute is occurring. The reason they can do this and the reason they can do this with so much of a smaller set of memory. Yes, SRAMM is much more expensive than DRAM.
It's like replacing, you know, copper with gold. It's very expensive, but you need so much less of it. Why? Software. And this is the 4D chess move from Nvidia. What I've just built out for you are my baseline opinions, the deal acquisition that happened and how Grock's product is faster even though the actual underlying memory SRAMM is more expensive.
This is a faster product. But why is Nvidia doing this? What is the 4D chess move? Well, this is where I wrote down that I believe this is a brilliant move by Jensen. Even though people talk about this being a memory hedge to hedge against rising DRAM prices, to me, this is actually a hedge against artificial general intelligence, as in AGI never happening.
Because if AGI doesn't happen and LLMs get commoditized, then GPUs aren't actually that valuable. And guess who's left holding the bag? Not Nvidia. No, Oracle, Nbby, Iron, and the other uh plays like frankly XAI uh or any GPU hoarder, they're the bag holders. See, AGI would need a lot of GPU and compute.
But commoditized LMS do not. Commoditized LM, in my opinion, need speed. If we could do a deep research and sit around for 45 minutes waiting for the deep research to run or we could get it instantaneously, that's where the money is. Now, one of the reasons practically that Grock is able to pull this off is they utilize a software infrastructure to predetermine how the chips are going to operate.
Now, I'm not going to get into the deep weeds of how the transistors and the capacitors and the refresh rates and all that stuff works inside of the actual chips. We're going to keep this a little bit more basic. But what I am going to say is that this speed is critical and the speed is created by software.
Now, if you remember what has Nvidia's moat always been, it's not necessarily that they have TSMC supply chains on lock. That is very true. It's not necessarily that uh they have uh SKH Highex's supply chains unlocked or Micron or Samsung supply chains unlocked to make sure Nvidia gets what Nvidia needs.
And it's not necessarily to say that Nvidia's GPUs won't be valuable, but it is to say that what has made Nvidia's GPUs even more valuable and has enabled them to get a lock on the industry is their CUDA software. But their CUDA software today doesn't actually integrate the same software that Grock uses to speed up inference output.
That's what this is a play about. If you read between the lines with what Jensen shared in a leaked email, he said, quote, "We plan to integrate Grock's low latency processes into the NVIDIA AI factory architecture, extending the platform to serve an even broader range of AI inference and realtime workloads." Ah, realtime workloads.
in other words haven't actually been able to output realtime workloads for the AI factories the digital twins that Jensen always pitches because they don't have as fast of an inference technology because even their Reuben chip which was supposed to be faster at inference still relies so heavily on GPUs that because the GPUs themselves the way they are architected out have built-in traffic cops they are slower.
There is underutilization in the GPUs and you're just not going to get answers as quickly. Now, this is where we could finally put the pieces of the puzzle together to understand the 4D chess move here. And I want to be clear, I'm not trying to shill Nvidia. Uh like I think there are big risks with Nvidia.
We're going to talk about that here. But this is the 4D chess move. Nvidia buys Gro's IP. They integrate the software technology of using software to be the traffic cop rather than the chip to organize in a deterministic way how the chips are going to operate. In English, make it damn fast through software.
They're going to integrate that software intellectual property into CUDA. Now CUDA becomes from behemoth to mega behemoth of the future. See, I think we go from a future of AGI, which ain't happening anytime soon, so nowhere, right? I think we go from this world view language model or world model generation through GPUs to just give us answers fast and real-time answers fast.
And that will be useful in piloting, in factories, in machinery, in healthcare, and in surgeries, anywhere you need realtime solutions. Now, Nvidia builds that software into CUDA, making CUDA even more of a behemoth for the next generation of artificial intelligence. And ironically, the next generation of artificial intelligence isn't actually any different than what we use today.
The only difference is faster. It's the same crap delivered to you faster. Nvidia moes it into CUDA. When Nvidia moes it into CUDA, guess who [clears throat] is going to be building the chips for these inference purposes that then the next bag holders can buy from Nvidia at no not low margins but at high Nvidia margins.
Well, the people who are going to buy it are going to be the next generation of data centers. Whether they're data centers in space, probably not, or just NBIS 2.0, Corewave 2.0, Oracle 2.0, whatever. Those new bag holders are then going to buy these SRAMM centric chips that speed up our artificial intelligence inference output.
So, the answers that we get from AI. So, the next generation of bag holders will buy those chips. The first generation again, Oracle, Enbis, Cororeweave, uh, whatever, XAI, they'll suffer. Gen two will then meme up until the next innovation comes and then those become the next bag holders. But guess who wins every time?
Every generation of bag holders continuously the company that wins is the one selling the product at the highest possible margin because they've locked the industry down with CUDA for worldview training and now CUDA for inference output. It's Nvidia. It's brilliant. Let me take a look at some of my other notes because that was all kind of from the top of my head there.
But I've been I've just been absorbing so much of this. I love this. This I think is very exciting. I I I think Jensen is not only the most brilliant salesman in the world uh for selling his stock and his company and his product next to maybe what you see at Palanteer. Uh but but he actually puts out like these margins are incredible.
So let's take a look at some uh some of these other notes here. So uh let's see. We spend okay there's this thesis that you spend money on training and you make money on inference. So I kind of agree with that. I think people don't really care that you're training a model. They want the output. It's kind of like us with my little AI startup disrupting real estate, right?
We're trying to democratize wedge deals across the entire country. Nobody knows wedge deals more than me. We created the wedge deal. Basically, it's like buying a Warren Buffett real estate with a margin of safety. And if we can put that into an app and sell it and help people make tens to hundreds of thousands of dollars, knock on wood, hopefully the value of that app will be massive.
And the point there is not again to pitch house hack. It's to simply say that when you can utilize AI in places that hasn't been used before, like this old school industry like real estate that most tech bros know nothing about today. That is where money, actual money can be made. So I I couldn't be more excited about that.
So um the founder of Grock also talks about how much latency matters. I love this. This was from a 2024 interview. He says if you show a human A or B and say which is faster, they won't know. It's imperceptible these days to tell you which is 100 milliseconds faster A or B. But if let's say B is 100 milliseconds faster, conversion with that 100 millisecond improvement will increase by 8% on desktop and 30% on mobile.
In other words, it matters. Speed matters. That's what he's trying to say. But we we already know this. I mean, do you want a humanoid robot that's walking around theoop? No. Even Jack knows he doesn't want that. He's laughing about it over there because nobody wants a boop robot. We want fast answers, baby.
So, this is very exciting. Now, um, Grock LPUs, language processing units. This is just sort of the fancy name they have for it. Great for voice AI. Instant text. Relies on static. A random access memory. That's SRAMM. It's fast. It's expensive. It's low capacity. The bottleneck is how much capacity you could have, but you could string hundreds of these together if you want, which is exactly what Nvidia loves to hear.
Like, think about that. Nvidia hearing what what somebody wants to string hundreds of these together. We'd love to sell that product. It's kind of brilliant. Uh, okay. uh they uh kill logic control on the actual chip which lets the chip just do what it wants to do which is focus on speed. Uh the analogy that I came up with this uh with on this is think about the GPU kind of like a semi-truck.
Massive volume, efficiency of volume, not very nimble, unstoppable once it's going like it can move a lot of product across the United States really well. But if you're like, "Bro, I just need this letter from Kevin's house to the post office fast." You take the Tesla Roadster baby, you take the roof off, you put the jetpack on the back, you get airborne, and you go deliver that piece of mail.
That's the difference between Gro's product and a high bandwidth memory GPU semi-truck. Uh so uh you know some people say Nvidia is just hedging against a DRAM uh you know supply chain collapse or some kind of inefficiency coming in the future. Uh maybe maybe I think that's like uh that's like basic level analysis like okay sure uh I think you can also say that you know Gro's had a lot of deals with the Saudis and uh and has raised money from the Saudis and uh has has made deals with the Saudis.
I think there could be an opportunity there for Nvidia also to sort of get deeper into the door of the Middle East because for Nvidia to really keep growing they kind of need to start selling to like rich countries like Singapore or which they already do to some extent but more to the Middle East the oil rich countries right uh Nvidia gets the brains and the AP they're leaving or the the IP and they're leaving the cloud business intact again because Nvidia doesn't want that they want to sell to the cloud business if anything Grock ends up the leftover business ends up being buyer of these Nvidia chips, right?
Uh these now Nvidia Grock chips. It's the ultimate irony. It's like you bought the IP of the software and the technology and the chip from Grock and then turn around and sell to what's left of Grock. Now Nvidia branded Grock chips. It's so brilliant. This to me is offense on commoditization, defense on commoditization.
It's you're basically making a bet that everything's going to commoditize and so you're going on offense on that, but then you're also defending against the decline of revenue that's probably going to happen on GPUs, which is bad for AMD, right? Uh then uh and AMD will catch up. You know, I'm not trying to be bearish AMD.
Uh let's see. Oh, yeah. Okay. And then I did want to say that like, you know, does yes, I I've got a $300 price target for Nvidia, but what what gives me the biggest concern for Nvidia? Okay. biggest concern for Nvidia, Sam Alman. Uh, and this actually reduces some of that concern. I've continuously believe that Sam Alman with his $1.4 trillion of spend is a ridiculous, ludicrous joke and it's going to fail.
It's it's not a bet I want to make on. So, I've been escaping Nvidia because of that. This actually flips that because over the next not necessarily short term but over the next two to four years this actually makes Nvidia a crazy powerhouse. Now that doesn't mean there can't be a correction in the short term especially if you know we actually do end up going into the commoditization direction and Sam Alman falters.
Who knows maybe OpenAI goes bankrupt. They stop getting funding. Private credit creates problems. Labor market stress leads the market to fall. Whatever. If the market falls, then spending amongst the upper 40 to 60% of American spenders falls. That leads the economy into a recession. There'll be a better buying opportunity to buy Nvidia and then that $300 price target will come true in maybe a decade, right?
But uh if if Samman stays up and running, then hey, maybe it'll happen sooner. But those are like that's my concern with Nvidia. So, I can't say like, "Oh my gosh, yolo calls on Nvidia because there are road bumps on the way or in in the path, but if you are a very long-term hodler here, this is a a moat builder.
They just deepen their mode against Nvidia, against Google, uh it doesn't matter. They're they're deepening their mode against anyone, and they're staying away from the bag holders, which are buying the chips from Nvidia at 75% gross margins, 55% net margins. This is absolutely brilliant. 4D chess. Like, man, I think it's so freaking awesome.
Uh, this this is the kind of stuff that like I want to emulate after with with my startup house. Sorry, I always go back to that, but it's my baby, you know? I want to do this kind of stuff over the next 60 years. I think this is this is so cool. Knock on wood, maybe even longer. But, um, yeah, very excited.
So, if you want to learn more, remember, none of this this video can't be a solicitation. Just go to househack.com, reinvest.co, learn more. If your browser for some reason doesn't work, there's one little IP sector somewhere in the country. Uh I think it's Spectrum. Uh use your cell phone and then the website should work.
But like 99.9% of you have access uh with no issues at all. And uh if you want to actually buy the AI product, you can do that. We'll be raising the price on December 29th and we'll have a big coupon expiration because the product will be releasing and the coupon code is release the app. .
Comments 0