Nvidia Stock Crashed — Understand It Better Than 99% of Investors — Transcript
Full transcript
- 0:00Nvidia is a story of a company that has
- 0:03always looked far into the future. In
- 0:062016, they were talking about how their
- 0:09GPUs would uniquely benefit from end
- 0:13markets such as AI, autonomous vehicles,
- 0:16and robotics. Fast forward 10 years, and
- 0:20looks like they were pretty correct
- 0:22about that AI thing. In fact, their
- 0:24revenues when they first made that
- 0:26prediction were just $5 billion in total
- 0:30with less than a billion dollars of
- 0:32profit. Fast forward 10 years till today
- 0:36and they have over $300 billion in
- 0:39revenues and $200 billion in profits.
- 0:43The stock over that period was up 100x.
- 0:48And in fact, if we rewind all the way to
- 0:50the 1991 IPO of Nvidia, we'll see that
- 0:55this is one of the strongest stock
- 0:57performances of all time. If you
- 1:00invested $1,000 at IPO, congratulations.
- 1:04You're now a centmillionaire. That is
- 1:06someone worth $10 million, almost as
- 1:09much as many Nvidia employees. And while
- 1:12no one can question their past success,
- 1:15questions about what the future brings
- 1:17for Nvidia are very open-ended and have
- 1:20gotten much harder to answer, especially
- 1:23this week with the announcement that
- 1:26Nvidia CEO and founder Jensen Hong would
- 1:30resign and step down after not one but
- 1:33two scandals in the same week. one
- 1:36involving an affair with a potential
- 1:38Chinese spy, another one involving an
- 1:41affair of Nvidia executive. And so now
- 1:45that Jensen Hang, who has led Nvidia for
- 1:48the past three plus decades and has seen
- 1:51around the corner of all sorts of
- 1:53potential risks to Nvidia, many of which
- 1:55we'll cover in this video. The question
- 1:58now is who will be the new leader of
- 2:00Nvidia? And I don't have an answer for
- 2:02you. And this leadership vacuum comes
- 2:05all at the same time that the risks are
- 2:07piling up. Some of Nvidia's biggest
- 2:10customers, Microsoft, Amazon, Meta,
- 2:15Google, not only spend hundreds of
- 2:18billion dollars together on their chips.
- 2:21They also are actively trying to cut
- 2:23them out of the data centers because
- 2:26they want to create their own chips and
- 2:28keep that margin for themselves. On top
- 2:31of that, the CUDA programming language,
- 2:34which was thought to be Nvidia's moat
- 2:37that would be impossible to surmount,
- 2:39has started to see some cracks with real
- 2:42competition coming up just in the past
- 2:45couple years and getting real adoption.
- 2:47And lastly, what about all of these
- 2:50circular financing arrangements and the
- 2:52fact that Nvidia is becoming a sort of
- 2:54credit back stop for over a hundred
- 2:57billion dollar in various guarantees? Is
- 3:00that concerning? We will cover all of
- 3:03this in much more in this deep dive
- 3:06video on Nvidia. If this is your first
- 3:08time here, what have you been doing? But
- 3:12you should know that I am a financial
- 3:14adviser at Davidson Con Management where
- 3:17we manage over half a billion dollars in
- 3:19AUM. I used to work at Goldman Sachs and
- 3:22Capital Group. On this channel, you'll
- 3:25find deep dives on a lot of different
- 3:27businesses as well as personal finance
- 3:30content that will help you make better
- 3:32investment and better money decisions.
- 3:34And so, if that is your thing and you
- 3:35want to get better with investing, go
- 3:36ahead and hit subscribe. Now, let us get
- 3:39into it. So, uh the way I structured
- 3:42this video is a little different than
- 3:43past ones because I feel like there was
- 3:46just so much to talk about and I wasn't
- 3:48sure how to fit it all in. So we are
- 3:51going to still start with the business
- 3:52section. I'll still touch on you know
- 3:55the overviews of the business, get into
- 3:57some of the products, a little bit of
- 3:58history and and context will kickstart
- 4:01us there. Then the next section I have
- 4:04I'm just calling six considerations.
- 4:07There's these sort of six different
- 4:08things that I came up with that are are
- 4:10kind of like key debate points uh things
- 4:12going on with Nvidia, things to
- 4:14consider. Some of them are risk uh key
- 4:17risk. Others are just kind of things
- 4:19that are kind of interesting to think
- 4:20about dynamics that are going on. Not
- 4:22clear whether or not it's going to be a
- 4:24beneficiary or you potential headwind or
- 4:27nothing for them. So that's going to be
- 4:29the next section. Then we're going to
- 4:31get into a bunch of other risk and then
- 4:33finally conclude with valuation where
- 4:36spoiler alert I'm going to disappoint
- 4:38you a little bit because as we're going
- 4:40to talk about uh valuation of a cyclical
- 4:43business is very very tricky. So I'm
- 4:46going to give you a little bit of a
- 4:47framework on how to do that. But there
- 4:49is a reason why a lot of investors have
- 4:52a stewed uh buying cyclical businesses
- 4:55and that includes of course Warren
- 4:57Buffett himself. uh if he does, he's
- 4:59just going to go ahead and buy him at a
- 5:00trough multiple and call it a day. Not
- 5:03going to go about with trying to
- 5:04estimate out uh all sorts of different
- 5:06earnings, but we will take a stab at
- 5:08doing that too. Aman, by the way, is
- 5:10another one who also just avoids
- 5:12cyclical businesses. A lot of investors
- 5:14do. So, one of the key questions though
- 5:15with Nvidia is going to be whether or
- 5:18not this is actually a cyclical business
- 5:20because there is a lot of people that
- 5:22have the opinion that the AI demand is
- 5:25so real and so high that this is going
- 5:28to be they might call it a Super Bowl
- 5:29cycle, but it's equivalent to saying
- 5:31that there's going to be sufficient
- 5:33demand for decades or longer that you
- 5:36don't need to think about this as being
- 5:38a cyclical business anymore. Going to be
- 5:40something we're going to discuss in this
- 5:42video. So, let us get started with what
- 5:45is the business. Now, the first thing
- 5:47you should know about Nvidia is that
- 5:49they're on a really weird fiscal year.
- 5:51Okay, that's definitely not the first
- 5:52thing you need to know about them, but
- 5:54it's worth mentioning because right now,
- 5:57you know, coming up in the next week,
- 5:59they're going to have 3Q earnings.
- 6:00They're calling that 3Q 2027 earnings.
- 6:04Obviously, the year is still 2026, but
- 6:07for some reason, I honestly don't even
- 6:09actually have a great answer for this.
- 6:10Some businesses, they have very weird
- 6:12fiscal years. And so, uh, their fiscal
- 6:14year, uh, the last one ended in January
- 6:172026, which really is basically 2025.
- 6:20It's 11 months of 2025, one month of
- 6:232026. Okay, if that was confusing at
- 6:26all, I'm sorry for taking up any of your
- 6:28bandwidth already because there's so
- 6:30much more we need to talk about. So, uh,
- 6:33I'm probably gonna flip between the two
- 6:35throughout the video. I'm going to try
- 6:36to clarify and just do calendar years
- 6:38cuz I feel like that's simpler, but I
- 6:40feel like someone's going to call me out
- 6:41and make a mistake on that at some
- 6:43point. And so, the first thing though is
- 6:44going to be LTM. They generated $32
- 6:48billion in revenue. So, that's the last
- 6:5012 months. Uh, and then in the last
- 6:52quarter alone, they did $105 billion in
- 6:55revenue with 106% year-over-year growth.
- 6:58And so that growth has actually been
- 7:01accelerating over the past couple
- 7:03quarters, which is just crazy when
- 7:04you're talking about a business that's
- 7:06doing hundred billion, growing 100%. Um,
- 7:09and this is kind of getting ahead of
- 7:11ourselves, but Jensen Hang talked about
- 7:13on the last couple earnings calls, how
- 7:16uh for the following year, which is
- 7:18going to be 2027,
- 7:20uh, he believes that there's going to be
- 7:22at least 70% capacity growth, but
- 7:25there's actually 100% demand growth as
- 7:27well there. That's kind of the
- 7:28difference between what they believe
- 7:29they could deliver versus the demand
- 7:31that's there. The point being it's still
- 7:33growing very quickly, thought to be
- 7:35sustained for at least another year or
- 7:37so. We'll talk more about that when we
- 7:39get to valuation. Now, gross margins
- 7:4175%. Those are up 12 points in the past
- 7:455 years. So, as they are getting bigger,
- 7:47you are seeing that they are getting
- 7:49some gross margin leverage there as
- 7:51well. Uh some pricing power too. some of
- 7:54that is coming through and so gross
- 7:56margins expanded to 75%. What's more
- 7:59impressive though is what happened on
- 8:00the operating income line. Uh on the
- 8:03operating income line we actually see
- 8:05they now in the last 12 months are doing
- 8:0765% margins and that's up 35 points in
- 8:12about the past 5 years. So margins went
- 8:14from you know 30-ish% range to now 65%.
- 8:18And if you're looking at where that
- 8:20operating leverage actually came from,
- 8:21it came pretty evenly from uh SGNA as
- 8:24well as R&D. And so if you're looking at
- 8:26just expenses as a percent of revenue,
- 8:29we see that uh SNM has gone from around
- 8:3110% to 2%. So very low. And uh R&D has
- 8:35gone from 20% to 8%. You know, it's
- 8:38still a high number in absolute dollars
- 8:40because they're such a massive business,
- 8:41but you're seeing a lot of that
- 8:43operating leverage uh flowing through.
- 8:45And that's really what has allowed them
- 8:47to now have, you know, that $200 billion
- 8:48in LTM profit. So, uh, before we get
- 8:51into what their products are, what the
- 8:54business is, you know, the advantage of
- 8:56the products, all of that, let's give a
- 8:58little bit of context and history about
- 9:00this business. I think it's very
- 9:02important to have some idea of where
- 9:04they started, some of the things they've
- 9:06been through to get a better
- 9:07understanding of a business. So, we're
- 9:09going to start all the way in 1993.
- 9:11Don't worry, the history part will be
- 9:14quick. But uh Nvidia was founded that
- 9:16year uh by Jensen Hang and two other
- 9:19co-founders. In 1999, they rolled out
- 9:22the first uh GPU that they had, the
- 9:25GeForce. They didn't invent the GPU
- 9:27exactly, but they did popularize it. And
- 9:29so that was a gaming focused GPU because
- 9:32that's where they started. That's where
- 9:34their roots were. They were in gaming.
- 9:36Now, one thing that's interesting about
- 9:38a GPU, if you're wondering what's the
- 9:40big deal about, you know, the GPU, why
- 9:42are these all the rage now instead of
- 9:44CPUs? Uh, right. So, the thing that is
- 9:47different is if you're thinking of a
- 9:49traditional CPU, the way it computes is
- 9:52it goes kind of line by line in order,
- 9:55whereas a GPU is able to do processing
- 9:58in parallel. So, if you think about a
- 10:00graphics card, uh, a GPU by the way, it
- 10:03just stands for graphics processing
- 10:04unit. If you think about this and you
- 10:07let's say are playing a video game, the
- 10:10pixels you're looking at on the video
- 10:11game, let's say you're thinking about
- 10:13the pixels up here versus the ones over
- 10:15there, they could process at the same
- 10:17time. They don't have any uh computation
- 10:19that relies on one another, right? So
- 10:22the graphics card could simultaneously
- 10:24or in parallel as it's usually said uh
- 10:27produce you know the pixel over here and
- 10:29the mathematical equation that produces
- 10:31the pixel over there. So it could run
- 10:33both of these equations simultaneously.
- 10:35That's what a GPU is. And so this
- 10:38started in graphics, but then we fast
- 10:40forward to around 2006 and Jensen Hong
- 10:44started having this idea that this sort
- 10:46of compute could actually be useful for
- 10:48a lot of other things and that is why he
- 10:52created the CUDA platform. So this was
- 10:54for programmers to help program GPUs to
- 10:57do a variety of different things to be
- 11:00more general purpose than just graphics
- 11:02and take advantage of this unique
- 11:04functionality of the GPU. Now we're
- 11:06going to fast forward a little bit more
- 11:07to 2012 and there was this AI
- 11:10breakthrough uh something called the
- 11:12AlexNet which was a neural network that
- 11:14dramatically improved uh image
- 11:17recognition and it was trained on Nvidia
- 11:20GPUs. So, this was uh one of the first
- 11:23times there was this real idea that hey,
- 11:26wait a minute. These GPUs are kind of
- 11:28uniquely good for this thing called AI.
- 11:31Maybe that'll be interesting one day.
- 11:33And you fast forward a couple years to
- 11:35the 2016 investor day uh that Nvidia
- 11:38had. And Jensen Hong is already kind of
- 11:40going allin on AI. They're talking about
- 11:42in their investor day in 2016, they're
- 11:45talking about how big AI is going to be.
- 11:47They're talking about autonomous
- 11:48vehicles. They're talking about
- 11:50robotics. These are big use cases or end
- 11:53markets that they really believe are
- 11:55going to create a lot of demand in the
- 11:57future and it's kind of just admirable
- 11:59if you will. Imagine seeing something
- 12:01coming and getting ready for it for like
- 12:04a decade. But that's that's kind of what
- 12:06Nvidia did. So that was back in 2016.
- 12:09But uh it should be noted that their
- 12:11their entire business at this time it's
- 12:13basically just gaming graphics cards,
- 12:16right? you know, selling them to desktop
- 12:18gamers, you know, they they'll sell them
- 12:20sometimes to console makers, to laptops,
- 12:22whatever. It's just these PC end markets
- 12:25is really what their business is. And,
- 12:27you know, they're thought to make good
- 12:29graphics card, but as you can imagine,
- 12:31it's a very cyclical business because,
- 12:33uh, consumer demand for computers, well,
- 12:35it fluctuates with the economy a lot.
- 12:38And so, in in 2008, revenues actually
- 12:40peaked at the time, hit $4.1 billion.
- 12:44and they wouldn't surpass that level
- 12:46again for 5 years until 2013. So the
- 12:49revenues they made in 2008 were not
- 12:51surpassed until 5 years later. Uh and
- 12:55profit margins over this time, they're
- 12:56bobbing around the mid- teens to 20%.
- 12:59Then the financial crisis, they're
- 13:00actually negative. There was another
- 13:02year in there, they also were pretty
- 13:03weak. Um so you could really see all of
- 13:06kind of the cyclicality in this business
- 13:08at the time. And then the crypto boom
- 13:09happens, you know, starting in 2017,
- 13:11really starts to take off in 2018. uh by
- 13:142019 they're now doing 11 billion in
- 13:16revenues, 32% margins and this just
- 13:19extends even through co there was a
- 13:21little bit of a lull in 2020 but uh then
- 13:24you know co it comes back with a
- 13:25vengeance but then there's the aftermath
- 13:27of this and by 2022 revenues are flat
- 13:31but they have all of these uh GPUs that
- 13:33they have to get rid of. They have to
- 13:34discount them basically to get rid of
- 13:36them. Can you imagine that? Just 4 years
- 13:38ago, this is in 2022, they were
- 13:41discounting GPUs to get rid of them. And
- 13:43uh their operating margins fell from 37%
- 13:46to 16%. And so this is just such a a
- 13:49crazy fact. If we rewind just about four
- 13:52years ago, their revenues were 26
- 13:54billion, their operating margins were
- 13:5816%. I'm doing dramatic pauses so it can
- 14:01sink in. Um, four years ago, 16%
- 14:04margins, 26 billion uh, revenue company.
- 14:07Today, 300 billion, 65%
- 14:12margins. That was a 4-year transition. I
- 14:14know we're all looking at like
- 14:15Enthropics ARR and how much that's
- 14:17growing. And, you know, it's fast. So,
- 14:19now it's like made us blasze about every
- 14:22other company that's ever grown before,
- 14:23but that is a crazy uh, level of growth
- 14:26for this business to have. And so uh
- 14:28this is of course the AI boom you know
- 14:30kick off in 2023 about with uh chat GPT
- 14:34and so their their decade long wait of
- 14:36waiting for AI to finally come through
- 14:39paid off for them and it paid off big
- 14:41time. So I whenever I'm looking at a
- 14:43business I love looking at a bit of
- 14:45history to understand where they're
- 14:46coming from. gives you a better sense of
- 14:48where they're going too because I think
- 14:50a lot of times if you're just popping in
- 14:52at the last quarter, the last year and
- 14:55then you're just projecting what
- 14:56happened in the last year out forward,
- 14:58you're going to miss a lot in the
- 15:00history, right? And so understanding
- 15:02that this was a pretty cyclical business
- 15:04and their success, you know, this level
- 15:07of success certainly is like a very new
- 15:09phenomena, I think is kind of helpful
- 15:11knowledge when you're going to try to
- 15:12predict out, you know, the next decade
- 15:14and you're real realizing, wait a
- 15:16minute, it's only really been a great
- 15:17business for the past four years. Um, so
- 15:20I think that's something that that's
- 15:21kind of worth taking into account. And
- 15:23when I'm doing this, I like to look at a
- 15:25lot of financial data and I do that on
- 15:27Fiscal AI. They are a sponsor of the
- 15:29show, but I used them before they
- 15:32sponsored the show. And it's nice
- 15:33because I can just pull up all of, you
- 15:35know, the operating margins going back,
- 15:37you know, two decades, two and a half
- 15:39decades, and can just see all of this
- 15:41data very clearly. And it makes it very
- 15:43easy to then pick out something I if
- 15:45something looks like it it doesn't
- 15:47really belong. And then you could hone
- 15:48in on that and ask more questions. So,
- 15:50if you want to use fiscal AI to do the
- 15:53same sort of analysis I do, uh there's a
- 15:55discount code in the show notes below.
- 15:57Now, that is enough of a plug. Let's
- 15:59move back to an overview of the
- 16:01business. So, uh, a year ago, their
- 16:04disclosures were a little bit different
- 16:06than what they offer now. Uh, I actually
- 16:09like some of the new disclosures they
- 16:11added. We'll talk about them in a
- 16:12second, but uh, on the old disclosures,
- 16:15they did have these different market
- 16:16segments they would break down, which
- 16:18was data center, gaming, professional
- 16:21visualization, and automotive. And so,
- 16:24we could see the market share stats of
- 16:25all of this. And when I tell you these
- 16:27numbers, it's going to be clear why they
- 16:29kind of didn't bother with these
- 16:30disclosures anymore. So, gaming 7% of
- 16:33revenues. Professional visualization. Uh
- 16:37this is going to be uh like 3D graphics
- 16:40and uh doing simulations and models uh
- 16:44stuff for professional design. So, this
- 16:47this was 1 and a.5% of revenues and then
- 16:50automotive was 1% too and data center
- 16:53was 90%. So, And since then it looks
- 16:57like data center actually grew even
- 16:58more. So uh it makes sense why they they
- 17:00don't bother doing this breakout
- 17:01anymore. Uh when they did do the
- 17:04breakout though they also broke down
- 17:05data center in terms of compute versus
- 17:08networking. So compute that's going to
- 17:10be more like the GPUs. Uh they also do
- 17:13have CPUs now but u mostly GPUs. And
- 17:16then networking. Networking is going to
- 17:18be kind of the things involved with
- 17:20helping the chips talk to each other.
- 17:22And then also the racks of systems, you
- 17:25know, the racks and servers talk to each
- 17:27other as well. So compute was 83%,
- 17:29networking 17%. Networking is actually
- 17:32getting more and more important and it's
- 17:34actually becoming more their kind of
- 17:37moat, if you will, and we're going to
- 17:39talk more about that later. So um, in
- 17:42terms of kind of their major products,
- 17:43like what do they actually sell? So they
- 17:45have a bunch of different architects of
- 17:47GPUs. you go back uh you know uh to a
- 17:50couple architects ago and it was the
- 17:52Hopper. So the H100, the H200, then they
- 17:55replace that with Blackwell, the B100,
- 17:58the B200. Up next is going to be the
- 18:00Reuben. Uh they also do have CPUs. Those
- 18:03go by the name Grace and Vera. Uh
- 18:06there's something called HGX that is
- 18:08going to be a server platform. So this
- 18:11is going to be basically multiple GPUs
- 18:14and then it's connected with their
- 18:16networking technology like an NVL link
- 18:18and so they'll they'll sell a whole
- 18:20server. Then on the next level to that
- 18:23though is going to be getting an entire
- 18:25rack or like AI factory as Jensen Hang
- 18:28likes to talk about that. And so uh DGX
- 18:30that's the name of of kind of the the
- 18:32leading product here right now from
- 18:33them. This is going to be their
- 18:35integrated AI system. It's going to give
- 18:37you eight of the B200 GPUs, uh, two
- 18:41Intel Xeon CPUs, two NV switch chips,
- 18:44and then also NVLink all together in the
- 18:47same server. And so, you'll note I said
- 18:49Intel, so they do still partner with
- 18:51Intel for certain CPU chips. looks like
- 18:54they're looking to replace those in a
- 18:56lot of architectures in the next
- 18:57generation, but uh Intel still has the
- 19:00x86
- 19:01uh CPU platform that they use and
- 19:04sometimes that's kind of preferred for
- 19:06certain servers. And so generally
- 19:08speaking, Nvidia is pretty integrated in
- 19:11terms of everything they're doing, but
- 19:13they do still become flexible. If it
- 19:16looks like there's an advantage uh or a
- 19:18reason why what one customer wants to
- 19:20like stick with an x86 architect, that's
- 19:22fine. They'll accommodate that. You'll
- 19:25even see later on this NV link, which is
- 19:27their um designed way to have chips talk
- 19:30to each other. They even now created
- 19:32something that allows chips that are not
- 19:34even Nvidia chips to kind of get in on
- 19:36that. And so they are you know flexible
- 19:39I would say in terms of yes they design
- 19:42everything and it works better when it's
- 19:43all integrated but if there is something
- 19:45that they can sell to uh you know a
- 19:48customer they'll they'll still do it and
- 19:49try to make sure they still have a
- 19:51position in the data center uh or
- 19:54conversely if there's a benefit from
- 19:56bringing in you know an outside um
- 19:59players uh technology into them like
- 20:01Intel they'll accommodate that as well.
- 20:03Um then coming up next is going to be
- 20:05the rack scale AI uh supercomputers. So
- 20:09this is like their their biggest thing
- 20:10now. One of the names you maybe would
- 20:12have heard is the Vera Rubin. That's
- 20:15officially going to be called the Vera
- 20:16Rubin uh NVL72.
- 20:19Very important you remember that. That
- 20:21is going to have 72 uh Reuben GPUs. So
- 20:24latest GPU architecture there. 36 Vera
- 20:28CPUs plus um NVLink plus a bunch of high
- 20:31bandwidth memory. All of that together
- 20:33is basically set as like a rack that's
- 20:36just ready to drop uh into a data
- 20:38center. And so uh we're going to talk
- 20:40more about this, but over time they've
- 20:42been moving from just being a chip to
- 20:44becoming an entire AI system. And this
- 20:47has actually allowed for them to have
- 20:49more of an advantage, a performance
- 20:51advantage by taking everything together
- 20:53and integrating it and optimizing all of
- 20:56it even when they're perchip level sort
- 20:59of performance drops. If you're just
- 21:00doing comparisons at the chip level, the
- 21:03performance level across the entire
- 21:05system tends to be maintained. And so
- 21:08this has been actually a pretty smart
- 21:09thing that uh Nvidia has been doing for
- 21:11for the past several years is moving to
- 21:13integrate everything together. They
- 21:15bought you know networking technology
- 21:17six years ago and this has been kind of
- 21:19a longcoming effort to really integrate
- 21:21an entire platform and Jensen Hong has
- 21:25been talking about you know AI factories
- 21:27for a long time. This is kind of the
- 21:29idea is that this is not about selling,
- 21:31you know, just a chip. Now, it's an
- 21:32entire AI system. Now, just going to run
- 21:34through a few more products quickly and
- 21:36then we're going to get into kind of the
- 21:38fun stuff, the meat of of this video and
- 21:40in the analysis. So, uh networking,
- 21:42there's going to be NVLink there and NV
- 21:44switch. Uh NVL link basically allows the
- 21:47GPUs to talk to each other and then the
- 21:49NV switch uh coordinates uh that. So,
- 21:52there's a lot of GPUs now that are
- 21:54running distinct workloads. And so the
- 21:56new thing, especially as models get more
- 21:59complex, is figuring out where to send
- 22:02these workloads to and orchestrating all
- 22:04of them. So having technology to help do
- 22:07that, Nvidia also has software for this.
- 22:09Uh that's very helpful to actually
- 22:11optimize um the the chips to actually
- 22:14get the most out of um the performance
- 22:17of these. So uh they also have something
- 22:19called Infiniband and Spectrum X
- 22:21Ethernet. This connects servers and
- 22:23racks across the data center. Uh there
- 22:26is still, you know, software to all of
- 22:27this. CUDA is part of it. There's other
- 22:30software though too that we could name
- 22:31here, but uh it's going to be just too
- 22:33much for right now. So, we'll keep
- 22:35going. Uh in gaming, it's still big. You
- 22:37know, they still have the GE Force uh
- 22:39GPU. Yeah, there's competitor GPUs now
- 22:42and all that. Still a good business for
- 22:44them, though. Um much more
- 22:46inconsequential relative to data center.
- 22:48Uh professional visualization and
- 22:50automotive, these are also going to be
- 22:52smaller businesses for them. Now,
- 22:54Automotive does have something called
- 22:56Nvidia Drive, which is an autonomous
- 22:58vehicle platform to help enable uh all
- 23:01these different car makers to have their
- 23:02own autonomous vehicle uh uh programs
- 23:05and really kind of give them a jump
- 23:07start with building their own EV
- 23:08technology. And then there's also
- 23:10something called Jetson, which is a
- 23:12compact AI computer for robots. Not
- 23:15going to talk much about that right now,
- 23:16but it's worth pointing out that a lot
- 23:18of the AI that we've seen in the world
- 23:20so far, it it's still, you know, LLMs,
- 23:23right? maybe uh agentic AI, but the real
- 23:26next lay layer to this afterwards is
- 23:29going to be what happens when we combine
- 23:31AI with robots and that's going to
- 23:33unlock, you know, a whole new wave of
- 23:36demand. So, still, you know, a little
- 23:38bit far out, at least as of now, but
- 23:40kind of just worth mentioning. Um, but
- 23:42not kind of in any of the numbers we're
- 23:44going to talk about for Nvidia here.
- 23:46Now, I mentioned that they added some
- 23:48new disclosures, and I like these
- 23:51disclosures because I feel like it helps
- 23:54allude to a bunch of the conversations
- 23:56we're going to be having now. Uh, but it
- 23:59also kind of shows maybe a little bit
- 24:00their paranoia. So, uh, if you're
- 24:03looking at these disclosures, they talk
- 24:05about hypers scale revenue versus
- 24:08revenue from AI clouds, industrial, and
- 24:10enterprise. So hypers scale revenue is
- 24:14169 billion in the last 12 months versus
- 24:17AI clouds industrial enterprise uh at
- 24:21109 billion. So the hyperscalers are
- 24:24growing 106% the AI clouds are growing
- 24:2768%. So the hyperscalers still very
- 24:31material portion of revenue and growing
- 24:34quicker right now. And then just before
- 24:36we get into kind of the the fun
- 24:37conversation all of that I should just
- 24:39note they basically have two segments
- 24:41now it's data center and then edge
- 24:43compute. Edge compute is basically just
- 24:45a catch-all bucket of all the you know
- 24:48uh compute that's not really data
- 24:50center. So it's gaming professional
- 24:52visualization automotive. uh that's
- 24:54going to be about 25 billion in revenue
- 24:56growing 35%. Whereas data centers 277
- 25:00billion uh growing 89% and that's going
- 25:03to be 92% of the total. Okay, we got a
- 25:07lot of background, a lot of numbers out
- 25:08of the way. I'm sure you remembered
- 25:10everything I already said, so that's
- 25:11great. Now, let's get into the fun part.
- 25:13So, these are going to be the six
- 25:15different considerations. The first one
- 25:18is going to be about uh the hyperscalers
- 25:21versus the neoclouds. And it kind of uh
- 25:25is alluded to in that disclosure change
- 25:28that I just mentioned where they're now
- 25:30talking about the revenue from the
- 25:32hyperscalers versus revenue from other
- 25:34sources from the AI clouds. And the
- 25:38reason why I think this is very
- 25:40important is because and I I do think by
- 25:43the way Jensen Wong is very very smart
- 25:46regardless of you know what's transpired
- 25:49um whatever sort of lapses in judgment
- 25:51he may have there's several things we're
- 25:54going to talk about here where it's like
- 25:56very early on maybe people were doubting
- 25:58you know why are they doing these
- 25:59arrangements where they're giving money
- 26:00to these different NeoClouds and like it
- 26:03you you'll kind of start to understand
- 26:04why through this conversation. So the
- 26:06first thing we should note is that there
- 26:08are the hyperscalers right who by and
- 26:10large are in the business of selling
- 26:13compute now because they buy a lot of
- 26:16chips and their business is renting them
- 26:18out and they have a lot of scale. They
- 26:21have an incentive to optimize their
- 26:24workloads to get not just the best
- 26:26performance but the p best performance
- 26:29on cost. Plus the fact that they have a
- 26:31lot of scale and they spend a lot of
- 26:32money on chips means that it usually
- 26:35makes sense for them to get into
- 26:37designing their own chips in order to
- 26:39save money. This decision to design your
- 26:42own chip in order to save money. It only
- 26:45makes sense if you're very very big and
- 26:47very massive. So right now we know of
- 26:50course Google has the TPU that's the
- 26:52name of their chip that they've
- 26:54designed. Um Microsoft has uh Maya uh no
- 26:59one talks about it as much but they do
- 27:01have their own chip they've been
- 27:03designing. Amazon uh they have a few but
- 27:05the one we'll talk about that's
- 27:06competitive to the GPU is Tranium. Uh
- 27:09that's their kind of version there. Uh
- 27:11and then Meta also has uh their own chip
- 27:13too. It's MTIA, however you're supposed
- 27:16to say that. And then Oracle is one of
- 27:19the only hyperscalers that does not
- 27:21actually have their chip. And I know
- 27:22Meta is not a hyperscaler, but they're
- 27:24in a weird place right now with all the
- 27:26money on capex they've been spending.
- 27:27They want to be a hyperscaler. Um, and
- 27:30so either way though, the point being
- 27:31that these are all players that have big
- 27:35cloud businesses, Meta the exception
- 27:37there, but they buy a lot of Nvidia
- 27:38GPUs, right? So they buy a lot of Nvidia
- 27:41GPUs and they use them a lot in their
- 27:43business. And they realized, hey, why
- 27:46are we paying this massive margin to
- 27:48Nvidia? Why don't we design our own
- 27:49chips and we can basically cut them out?
- 27:52We don't need them. Now, that's not the
- 27:53easiest thing to do and we're going to
- 27:55talk throughout this video why that is,
- 27:56why it is hard to actually compete
- 27:58against Nvidia. But the thing I want to
- 28:00note now is that their scale does enable
- 28:03them to make that decision to design
- 28:05their own chip. So, if they were small,
- 28:09they wouldn't have the resources to be
- 28:11able to do that. And so what Nvidia did
- 28:14was they started funding a lot of these
- 28:17small cloud competitors. And they didn't
- 28:20just fund one or two. They funded, you
- 28:22know, like half a dozen or more of them
- 28:24cuz they just wanted a bunch of them
- 28:26because they knew that the more uh small
- 28:29cloud competitors there were, they could
- 28:32then freely sell their GPU chips to them
- 28:35without the fear that they would get big
- 28:37enough and get scale to then make the
- 28:39decision to design their own chips and
- 28:40try to cut Nvidia out. So Nvidia
- 28:43invested in a lot of these companies
- 28:45including Coreweave, Nebus, Nscale, uh,
- 28:48Lambda, Cruso, a bunch of other ones.
- 28:51And so they put money into all of these.
- 28:54They gave them GPUs instead of giving
- 28:56them to Google and Amazon because they
- 28:58were trying to basically create their
- 29:00own um, competitor chip. So instead,
- 29:02Nvidia said, "I'm going to give my chips
- 29:04to all these competitors. I'm going to
- 29:05help fund them, too, and then this is
- 29:08what is going to enable me to be more
- 29:10protected because it's very unlikely
- 29:13Nebius or Coreweee now is going to go
- 29:15and design their own chip." And so a lot
- 29:17of their architecture too is now
- 29:18dependent on Nvidia. So, uh, it just
- 29:21puts them in a much better position. you
- 29:24know, yeah, it's possible Google sells
- 29:25like TPUs to to Coreweave or something
- 29:27like that in the future, but uh the
- 29:29Neoclouds are also very scared of
- 29:32getting on Nvidia's bad side because
- 29:34their entire business is basically we
- 29:36were able to get Nvidia GPUs when other
- 29:39people weren't. None of these NeoClouds
- 29:41would exist if Nvidia didn't want them
- 29:43to exist, but they do want them to exist
- 29:46uh because it's in their competitive
- 29:48interest, too. So, this is the the first
- 29:50consideration here. You have this
- 29:52dynamic where you have a lot of your
- 29:53revenue, majority of revenue coming from
- 29:55your customers who are actively trying
- 29:58to cut you out while you're
- 29:59simultaneously
- 30:01funding the competition against them as
- 30:04a kind of way to balance out the dynamic
- 30:07but also ensure that there are all these
- 30:09players buying your chips in case they
- 30:11do kind of fully cut you out. Now, I
- 30:13don't think that Nvidia is going to ever
- 30:15get fully cut out from, you know,
- 30:16Google, Amazon, any of them. I think it
- 30:19it's more likely that you have some base
- 30:22level of compute that is going to be
- 30:24Nvidia. There's still reasons to have
- 30:26it. It tends to be more flexible.
- 30:27There's certain workloads uh things are
- 30:29still the best uh performance per dollar
- 30:31on. But what's slowly happening is that
- 30:33certain workloads are being carved out
- 30:36and are going to these custom chips that
- 30:38they're making. Um Google in specific
- 30:41with the TPU is particularly competitive
- 30:43because um it's kind of you know a
- 30:46little bit more general purpose uh
- 30:49versus some of the other customized
- 30:50chips that the hyperscalers are making.
- 30:52And so the the one other thing worth
- 30:54mentioning here in the same bucket is
- 30:55sovereign AI. So this is something that
- 30:58Jensen Hong's been talking about for a
- 31:00few years. And I was kind of curious.
- 31:02I'm like why does he keep mentioning
- 31:03this idea? Like who cares that much
- 31:06about whether or not you know a country
- 31:08has their own you know sovereign AI data
- 31:10centers? Well Nvidia cares and do you
- 31:14know why? Because if they are able to
- 31:16convince a country that they need to
- 31:18have sovereign uh their sovereignty over
- 31:21their data centers then very likely what
- 31:23that means is that they're going to have
- 31:25to set it up themselves. And if they set
- 31:28it up themselves, the easiest way to do
- 31:31that is to go directly to Nvidia and buy
- 31:33one of their AI factories. They have
- 31:36whole server racks you could just buy
- 31:38and plop into a data center. And so
- 31:40that's why the AI sovereignty sort of
- 31:42idea or theme is actually really
- 31:44important to Nvidia because it's just
- 31:46it's the same part of the same story
- 31:48where you get a more disperate group of
- 31:51customers buying your product. So you're
- 31:52diversifying out uh where it's going.
- 31:55But the bigger part of that too is that
- 31:56this is going to be another customer
- 31:58who's not going to be a competitor to
- 32:00you. Nvidia is sick of having their
- 32:02customers try to then design their own
- 32:04chips and cut them out. And if you're
- 32:06talking about, you know, a country like
- 32:09Saudi Arabia setting up their own data
- 32:11centers for sovereign AI, they're not
- 32:13going to also get into the chip design
- 32:15business. They're just going to buy
- 32:16Nvidia's chips. So that's why the
- 32:18sovereign AI story is also kind of
- 32:21another consideration here. So, by the
- 32:23way, when I wrap this all up, I I'll
- 32:24kind of conclude with Nvidia's moes and
- 32:27what I think about them, where I think
- 32:28they're strengthening, weakening. But
- 32:30for now, these are just kind of
- 32:31considerations, things going on. So, the
- 32:33second consideration is going to be this
- 32:35idea of an individual chip versus an AI
- 32:39factory. And so, generally speaking,
- 32:41Nvidia's chip is still highly
- 32:44performative on a perost basis. So,
- 32:47there's lots of different ways you can
- 32:49run test. And you know, there's
- 32:51different ways you could kind of fudge
- 32:53it and get different results and and
- 32:55people do that because everyone wants to
- 32:56show they have the best chip. Generally
- 32:58speaking though, Nvidia's chips are are
- 33:00still always up there. However,
- 33:03competitors are getting closer to
- 33:06sometimes passing uh them on a perch
- 33:08chip performance level. So, especially
- 33:11if we're thinking about Google's um TPU
- 33:14and also AMD's GPU in some cases is
- 33:17actually more performant than Nvidia's
- 33:20chip. But there's an asterisk on this.
- 33:22This is going to be usually compared to
- 33:25um their their later generation chip,
- 33:28not the latest one like the one that
- 33:30they have yet to release. And then also
- 33:33uh it's not going to be when you're
- 33:34running the test on a system level. So
- 33:37what I mean by that is let's say for
- 33:39example, this was an analysis uh
- 33:41semi-analysis did was they compared
- 33:43Google's TPU to Nvidia's uh B300 chip.
- 33:48And what they found was the Google TPU
- 33:51uh was cheaper on a per token basis by
- 33:5319%. So that's just comparing one chip
- 33:56to the other chip. However, and this is
- 33:59the important part, if you compare it to
- 34:02the entire Nvidia system, uh which is
- 34:04called, you know, the GB300
- 34:07NVL72. I know you remember that from the
- 34:10product breakdown. What that is, uh,
- 34:12it's one of those AI systems that has,
- 34:14uh, 72, uh, GPU chips, a bunch of CPU
- 34:18chips, NVLink, NV switch, and if you're
- 34:21running it compared to this entire unit,
- 34:24then uh, the performance uh, per price
- 34:27uh, actually is 30% less than uh, than
- 34:30Google's TPU. So, the performance is
- 34:33still better on the systems level. And
- 34:35so that's kind of the consideration of
- 34:38this section is the difference between
- 34:39chip performance versus uh system level
- 34:42performance. And what's been happening
- 34:44is that competitors have caught up by
- 34:47and large pretty close maybe you know a
- 34:49year behind but pretty close on the chip
- 34:51level but not on the systems level and
- 34:55the integration of all of that. And it's
- 34:57going to be harder for them to do that
- 34:58too because philosophically the way
- 35:00Nvidia designed uh their chips, their
- 35:03hardware is they're tying everything
- 35:05together. They're tying together the
- 35:07chip, the linking between the chip uh
- 35:10then also the linking between the racks
- 35:12and then also the software CUDA right on
- 35:15top of that. It's all together working
- 35:17together uh to to get the best
- 35:19performance. And I'm going to talk more
- 35:21about CUDA in a different consideration.
- 35:23And so the other thing I just want to
- 35:24hit on here is that yes, there's lots of
- 35:27uh different companies that'll talk
- 35:28about how their chip is more performant
- 35:31versus, you know, the the regular GPU
- 35:33like Amazon. They came out and said
- 35:35their their tranium chip is 30% better
- 35:38performance per cost versus uh you know
- 35:40a general GPU they're talking about
- 35:42Nvidia, but that doesn't necessarily
- 35:44mean compared to the entire system
- 35:46running the entire system. That could
- 35:48just be them plopping chips in and out.
- 35:50So you have to be careful with that
- 35:52distinction and people can run these
- 35:54tests and kind of fudge the results a
- 35:56little bit if they want to. Um so so
- 35:59that's kind of something worth
- 36:00considering there. And then the last
- 36:01thing I want to say here though is that
- 36:03the more real world deployments you
- 36:05have, the more there's kind of this
- 36:07feedback system of optimizing your
- 36:09hardware and software to a real world
- 36:12workload in order to continue to improve
- 36:14it. That's actually going to be one of
- 36:16uh Nvidia's advantages that I'll I'll
- 36:18summarize with later. Okay. The third
- 36:20consideration is going to be CUDA. So
- 36:22this is their programming platform. It
- 36:24takes a programming language like
- 36:26PyTorch or Jax which themselves
- 36:28sometimes call on different CUDA
- 36:31libraries which basically means instead
- 36:33of you having to like code something out
- 36:35and say like move piece of memory from
- 36:38here to there, pull it out and then like
- 36:40you know save it somewhere else. Instead
- 36:42you could just call on the library and
- 36:44it just does that all without you
- 36:46actually having to code it all out
- 36:47yourself. Um, and then it compiles all
- 36:50that code and then it gives the actual
- 36:53instructions to the GPU. So, the thing
- 36:55to know about KOD because it's not
- 36:57technically a programming language, uh,
- 36:59which I really always want to call it a
- 37:01programming language because it just
- 37:03sounds like it should be, but it's not.
- 37:05Instead, what it is is it's a
- 37:07programming platform. And what it allows
- 37:10is the programming language to basically
- 37:13get translated in a way to give
- 37:14instructions to the chip itself. Because
- 37:17keep in mind these instructions at the
- 37:19chip layer are literally moving
- 37:22electricity around transistors, right?
- 37:24And a chip may have millions of
- 37:26transistors. And so what CUDA is doing
- 37:28is it's taking this program and it's
- 37:31saying here's how you need to move all
- 37:32of the electricity around the chip in
- 37:35order to uh execute this program. So the
- 37:38feedback loop though that exists between
- 37:40design and CUDA is that there's more
- 37:43efficient ways to move electricity
- 37:46around a chip. basically to do these
- 37:48certain functions on a chip. And so over
- 37:51time in every iteration they learn, oh
- 37:54here's a better way that we can optimize
- 37:56the chip to do this function. Let's
- 37:58change that in the software. Let's uh
- 38:00tell it to to change the path it's
- 38:02taking basically on the actual chip
- 38:04level. On top of that, they could say,
- 38:06oh well, what if we design the chip
- 38:08differently? Does that make the software
- 38:10more optimized? And so that's the
- 38:12feedback loop between the two and why
- 38:13this was so so important historically.
- 38:16It still is, but this is what's allowed
- 38:18um uh different programmers to actually
- 38:20get the best performance on the chip.
- 38:23You it does require still that software
- 38:26layer to squeeze it out. So um the thing
- 38:29though that's been happening is you know
- 38:31you got these very big companies like
- 38:34Google, like Amazon, uh AMD and all of
- 38:37them have decided that they cannot be
- 38:39beholden to a single proprietary
- 38:42language. So Google has, you know, their
- 38:45whole TPU stack, but they also have
- 38:47something called XLA that's going to be
- 38:49their own kind of version of this that's
- 38:51their own compiler. So it can take uh
- 38:53PyTorch, you could take Jack's code, it
- 38:56can compile it into instructions that
- 38:58then gets sent to a TPU. So very similar
- 39:02uh to CUDA and it could cut CUDA out of
- 39:04the process entirely. Uh Amazon, they
- 39:06have Neuron, so that's their version of
- 39:08the same thing. Uh AMD has Rockm that's
- 39:12their version too and this is uh open
- 39:14source as well. So uh it's shared
- 39:16amongst anyone who wants to use it. Um
- 39:18OpenAI also has their own programming
- 39:21language platform and compiler called
- 39:24Triton. And so you could write GPU
- 39:27programs very similar to how you could
- 39:29write uh Python code and then it could
- 39:31compile for either CUDA or RockM which
- 39:34is AMD's version. So the idea OpenAI had
- 39:37was um let's help get this open-sourced
- 39:40uh uh programming language very popular.
- 39:43So there's a new ecosystem around it.
- 39:45Then well the decision to use either
- 39:47CUDA for Nvidia chips or Rockm which
- 39:49could be AMD chips since it's open
- 39:51source it could also be other chips too.
- 39:53Um the thing to keep in mind though is
- 39:55that just because something can compile
- 39:59for either CUDA or rockm which just
- 40:01again to be clear what I mean by that is
- 40:03you're taking the pietorch the actual
- 40:05language then when you're compiling it
- 40:07you're turning it into actual
- 40:09instructions for uh the chip. So just
- 40:12because you can compile for CUDA or
- 40:14rockm it doesn't mean you're actually
- 40:16optimizing for the best performance. So
- 40:19very often you kind of have to write for
- 40:21one or the other because the chips are
- 40:23different, right? The the language is
- 40:25different, the architecture is
- 40:26different. Now the thing that is really
- 40:28changed in my opinion and I mentioned
- 40:30this in my AMD video, which by the way
- 40:33that if any of this is kind of
- 40:34confusing, I feel like I kind of had a
- 40:36little more patience and explained a lot
- 40:38of things even simpler in that video.
- 40:40This one I knew would be really a long
- 40:42video, so I didn't do the same thing
- 40:44here, but definitely you should watch
- 40:45that one, too. But in that video I
- 40:47talked about one of the key risks I saw
- 40:49to CUDA being AI. And what I mean by
- 40:52that is AI is very good at translation
- 40:56and it's very good at optimization.
- 40:59Whereas you know before if you're
- 41:02translating something you know it's like
- 41:04a dictionary, right? It's like you you
- 41:05look up this word and you see what it
- 41:07translates to. And that's not a very
- 41:08good way to translate things because
- 41:10sometimes things are lost quote unquote
- 41:12in translation as we say because a
- 41:14direct translation doesn't always work.
- 41:16But AI is smart enough to get the gist
- 41:19of what you're saying and then translate
- 41:22to that which is a similar thing uh with
- 41:24chip architecture and programming for
- 41:27chips. It could catch the gist of what
- 41:29you're saying and then try to write to
- 41:31that. On top of that, it could then run,
- 41:34you know, once it compiles the code, it
- 41:36has this instruction set. It could then
- 41:38run something, you know, a thousand
- 41:39times to see what's the best way to
- 41:41actually run it and optimize it, learn
- 41:44that, and then in the future, it knows
- 41:45how best to actually run that that
- 41:47function. And so AI is something that
- 41:50makes the development of these uh
- 41:53programming platforms a lot lot better.
- 41:56On top of that, it could reduce the
- 41:58switching costs for someone who did
- 42:00program entirely in CUDA. Now, maybe
- 42:03it's easier to translate that to a
- 42:05different language. Now, I don't want to
- 42:06overstate this. Translation's still not
- 42:09perfect, and if you're talking about
- 42:10millions of lines of code, it's going to
- 42:12be super annoying trying to figure out
- 42:14where the bugs are, how to optimize it,
- 42:16and all that. But, as AI continues to
- 42:18get better and better and does more and
- 42:20more of this, I I do believe the CUDA
- 42:22moat is shrinking. uh I at least
- 42:25relative to where it was historically, I
- 42:28think that this is a shrinking mode for
- 42:30Nvidia. One other key thing to keep in
- 42:32mind though is that especially when
- 42:34you're dealing with kind of a a chip
- 42:36supply shortage, getting the most
- 42:38performance out of the chip is all the
- 42:41more important, right? To get the best
- 42:43uh cost per performance sort of basis.
- 42:45And so having software optimized for the
- 42:48chip matters even more in these
- 42:50environments. and Nvidia plus CUDA still
- 42:53tends to be the best uh in most
- 42:55circumstances and they also continue to
- 42:57improve every year, right? So they're
- 42:59also not standing still. They also can
- 43:01use AI to optimize this and they're
- 43:03starting from a head start with also a
- 43:05lot of real world AI workloads. So that
- 43:08was the third consideration. The fourth
- 43:10one is going to be training versus
- 43:13inference. So, um, it was generally
- 43:16thought that Nvidia had a better
- 43:19position in training versus inference.
- 43:22And that is because when you train, you
- 43:24need all of your chips to kind of talk
- 43:26to each other. You need to have very low
- 43:28latency because if you're waiting for
- 43:30like one chip to do something before the
- 43:32next chip does it, then your training
- 43:34just takes longer and longer. It becomes
- 43:36more expensive. And so, uh, it was
- 43:38thought that since Nvidia does have very
- 43:40good networking, it could talk to all of
- 43:42its chips, that it was better for
- 43:44training, it had a little bit more of a
- 43:46moat there. Whereas with inference, that
- 43:48would be easier for uh, AS6 um, which is
- 43:52applied um, specialized integrated
- 43:54circuits. It's basically the specialized
- 43:56chips that these companies are making
- 43:59um, would be better for inference. And
- 44:02so you know uh Amazon has tranium that
- 44:05can take some of the inference workloads
- 44:07a lot easier than it would be for them
- 44:08to replace the training. And so with
- 44:10training um Nvidia would be stronger.
- 44:13Now the problem with that is that you
- 44:15know while AI is going to always be
- 44:17training models over time more and more
- 44:19workloads are going to go to inference
- 44:21right and so that would be negative for
- 44:23Nvidia because maybe they keep training
- 44:25but then if inference keeps getting
- 44:27carved out that's a big market that
- 44:28they're losing. Now what's kind of
- 44:31changed more recently and Jensen Huang
- 44:33has been uh clear on on this point is
- 44:36that inference um is actually getting
- 44:38more complex particularly with agentic
- 44:41workflows because with agents maybe
- 44:43you're coordinating a hundred different
- 44:44things all at once and you're waiting
- 44:46for one agent to finish before the next
- 44:48one can step in. That sounds pretty
- 44:50similar to training. And in these cases,
- 44:52again, your um interconnections, your
- 44:55your NVL link, you want the networking
- 44:57uh to really be good so that all of
- 44:59these chips can link and so you're not
- 45:01waiting. And so inference is getting
- 45:02more complex and it's starting to look a
- 45:05little bit more like training in this
- 45:07respect, in which case Nvidia has their
- 45:09advantage there again because of the
- 45:12networking aspect. And so in this
- 45:13consideration, it seemed like this was
- 45:15something that maybe was a risk. I
- 45:17wanted to surface it in case this is an
- 45:19argument that you heard somewhere. It
- 45:21doesn't seem like this is really going
- 45:23to be an issue for Nvidia. So that was
- 45:25the fourth consideration. The fifth
- 45:27consideration is going to be open versus
- 45:29closed source model. So a open- source
- 45:32model is a model that you can download
- 45:34without any restrictions on it generally
- 45:37speaking and you know it could be hosted
- 45:39kind of anywhere. A closed model will be
- 45:42open AI's model or Enthropics model
- 45:44which is going to be hosted by them and
- 45:46they get to decide where it's being run.
- 45:48Now the problem with this is that also
- 45:51means that they get a lot of power uh
- 45:54over deciding where to send uh all of
- 45:57that uh workloads basically. And so they
- 46:01can decide to uh design their own chips
- 46:04for instance and then direct workloads
- 46:07to the chips that they have designed.
- 46:08And this isn't hypothetical. Open AAI is
- 46:11designing or has designed already their
- 46:13own chip and they may be hosting it
- 46:16themselves. Maybe they're still putting
- 46:18it, you know, in partnership with
- 46:19Microsoft. It doesn't really matter.
- 46:21Nvidia is still getting cut out of that.
- 46:23They're also more likely to go to
- 46:24hyperscalers to send a lot of the
- 46:25workload. So uh it's not a huge negative
- 46:29in a world where you know closed models
- 46:31become the most dominant way to access
- 46:33AI but it does present more risk for
- 46:35Nvidia because again they do have this
- 46:37option of cutting them out. Both open
- 46:39AAI and uh Enthropic are designing their
- 46:41own chips. Having said that it doesn't
- 46:44look like the world is really going in
- 46:46that direction where there's just like
- 46:47these couple closed models. Instead it
- 46:49looks like there's a lot of different
- 46:51models. Um, and even within that, even
- 46:54if it is a world with just a lot of
- 46:56closed models, Nvidia still, uh, you
- 46:58know, looks like they'll have a position
- 47:00there. And so, generally thinking, I
- 47:02don't think this is the biggest risk,
- 47:03but the more powerful closed models
- 47:06there are, the more ability they have to
- 47:10basically create their own chips. Um,
- 47:12and also, you know, pressure Nvidia down
- 47:14the line, right? Nvidia doesn't want
- 47:16very concentrated, powerful customers.
- 47:18Instead, they want a bunch of Neoclouds
- 47:20that they can, you know, go ahead and
- 47:22give them chips and they say, "Oh my
- 47:24god, thank you so much. You know, we'll
- 47:25pay whatever." And that's the business
- 47:28that they want. They don't want to sell
- 47:29to customers that then say, "Oh, well,
- 47:31we're going to design our own chip and
- 47:33cut you out." And the result of that is
- 47:34Nvidia has to constantly continue to to
- 47:37move up in performance to justify the
- 47:40premium of their chips. So, that was the
- 47:42fifth consideration. The last
- 47:44consideration has to do with chips
- 47:46obsolescence. So it was generally
- 47:48thought that uh Nvidia's uh GPUs would
- 47:53be more general purpose, more
- 47:55programmable. They could be good for
- 47:57multiple different workloads, training,
- 47:59inference, whatever else. And as a
- 48:01result of that, they would have longer
- 48:03shelf lives and uh better useful lives
- 48:06too as well. And the reason why that's
- 48:08important is because when you're
- 48:09building a data center, you're going to
- 48:11want to put into the data center
- 48:13something that's going to last a long
- 48:14time so you know you're not wasting your
- 48:16money. So for that reason, you may be
- 48:18less likely to buy something like uh you
- 48:20know an Amazon tranium chip if they were
- 48:23selling them or some of these other uh
- 48:25specialized purpose chips that you could
- 48:26get that maybe are cheaper in the short
- 48:28run to run inference but then become
- 48:31more expensive in the long run. So, this
- 48:34was thinking for a while that that uh
- 48:36Nvidia's GPUs would have better useful
- 48:39lives, less obsolescence risk. I think
- 48:41it's generally true, but I think
- 48:43Google's TPU is not that far off, and I
- 48:46think that other chips also won't be
- 48:48that far off in the future, but I do
- 48:50still think this is a little bit of of
- 48:52moderate advantage for Nvidia currently
- 48:54is that their chips are less likely to
- 48:56go obsolete. So people are more
- 48:58comfortable buying them for their data
- 49:00centers knowing that they'll get that
- 49:01full useful life out of them. Um in this
- 49:05environment though with everything so
- 49:07compute constrained it almost doesn't
- 49:09fully matter as much anymore but that is
- 49:12another consideration. Now let us get
- 49:14into Nvidia's moes and so there still is
- 49:17the software the cudas there it's
- 49:20formidable. I did say though I think
- 49:22that moat is starting to erode a little
- 49:24bit but that doesn't mean it's eroded. I
- 49:26just see it not being as strong in
- 49:27several years from now as it is uh you
- 49:30know in the past. But having said that,
- 49:32it still is something that's formidable,
- 49:34hard to overcome because as of now, they
- 49:36still generally do get the best
- 49:38performance out of their chips because
- 49:40of their language and competitors do
- 49:43still have to close that gap. Uh the
- 49:45next thing that's worth noting is going
- 49:46to be the system level. This is kind of
- 49:49the biggest thing worth pointing out is
- 49:51that they still have the best
- 49:53performance on a systems level. that is
- 49:56the cheapest to run. So as long as that
- 49:58continues to be the case, that is the
- 50:00simplest and best selling point for
- 50:02Nvidia. It's very similar to Meta going,
- 50:04you know, we have the highest return on
- 50:06ad spend. Nvidia can say, you know, we
- 50:08have the best performance on a per cost
- 50:11basis, per token basis. And that is
- 50:13enough of a selling point on its own as
- 50:16long as that continues to be the case.
- 50:18It's hard though, right? You have to
- 50:19continue to execute. I'll say more about
- 50:21that in a moment. Um the other advantage
- 50:24is they have a lot of real world AI
- 50:26workloads. This shows up as an advantage
- 50:28for them because that means they're
- 50:30better able to optimize their chips,
- 50:31their software, and then when it
- 50:34actually gets into a data center, the
- 50:35customer actually sees the performance
- 50:38uh actually work the way it's expected
- 50:40to and it continues to improve over
- 50:42time. So that's what allows them to
- 50:44continue to have top performance. Uh the
- 50:46next one's going to be networking and
- 50:47integration. Uh I mentioned a few times
- 50:50this NVL link thing that's pretty
- 50:52important to get all the chips to talk
- 50:53together. NV switch uh helps orchestrate
- 50:56all of them. Uh they also have other um
- 50:59networking technologies so the servers
- 51:01talk to each other. All of this together
- 51:02is what has moved them from just
- 51:04competing on a chip basis to an entire
- 51:06you know AI systems basis and has
- 51:10allowed them to maintain their
- 51:11performance because it's one thing to
- 51:13build a chip. It's another thing to also
- 51:15build networking too on top of that and
- 51:18you know they have to outsource a lot of
- 51:19this to to Broadcom to Arista and then
- 51:22when you're outsourcing it and it's not
- 51:23integrated together it just doesn't
- 51:25quite work as well uh as if it was fully
- 51:28integrated together. So that is going to
- 51:30be another advantage Nvidia has. Uh the
- 51:33next one is that their uh timing of
- 51:35improvements have sped up over time. It
- 51:37used to be every two years they came out
- 51:38with uh new updates on on CUDA and new
- 51:41uh chip architectures. now it's every
- 51:43one year. So, they've sped that up. So,
- 51:45that's been um kind of an operational
- 51:48advantage, just the fact that they've
- 51:49been able to do that um and have an
- 51:52organization that, you know, can have
- 51:54enough R&D spend and and be an R&D
- 51:56organization that's capable of that.
- 51:58That's kind of their their advantage,
- 51:59but it is, you know, execution uh
- 52:02dependent very much so. And then the
- 52:03last one's going to be the developer
- 52:05ecosystem. So this is similar to CUDA
- 52:08but this is the fact that a lot of
- 52:09people are familiar with CUDA know how
- 52:11to use it and will thus prefer to use it
- 52:13and when they use it they'll be you know
- 52:15higher performing. So that that's kind
- 52:18of uh a bunch of the advantages that
- 52:19Nvidia has. The most important one
- 52:21though again is going to be that they
- 52:23have the highest performance on a per
- 52:26cost per token basis. Uh as long as
- 52:29that's the case they're going to be
- 52:30fine. But making sure that's the case
- 52:33far out in the future, that that's a
- 52:34pretty hard thing to do because they
- 52:36have to continue to execute and
- 52:38outperform competition in order to
- 52:41justify their price premium. Whereas a
- 52:43competitor uh like with Google and the
- 52:45TPUs, they could just undercut them on
- 52:47price. And so they could be a little
- 52:49less performant, but since they're
- 52:51cheaper, the performance per cost is the
- 52:54same. So, Nvidia has to continue to do
- 52:57better than everyone else in order to
- 52:58maintain their margins, which is also
- 53:00kind of a key risk as you can imagine.
- 53:03So, uh why don't we actually get into
- 53:05risk? I just gave you one there and some
- 53:07of the other ones we're going to talk
- 53:08about are going to be this circular
- 53:10financing thing that people talk a lot
- 53:12about. Um I'm going to just tell you
- 53:15it's interesting to talk about some of
- 53:18it. It does remind me a little bit of
- 53:20maybe some of the things you hear about
- 53:22during the the you know uh tech uh
- 53:24bubble in the 2000s, but it's nowhere
- 53:27near that level just in terms of dollar
- 53:29amounts. Even though these are large
- 53:30dollars relative to how big Nvidia is
- 53:33and how much cash they generate. I told
- 53:35you they have 200 billion in earnings in
- 53:38just this last year. That is more or
- 53:40basically as much as everything we're
- 53:42talking about here. You're talking about
- 53:43one year of earnings. So circular
- 53:46financing, they've done about hundred
- 53:47billion in equity investments. A lot of
- 53:49these investments that they do do
- 53:51involve them then turning around and
- 53:52buying their chips. The reason why they
- 53:55do that very often is because Nvidia
- 53:57wants to continue to entrench in the
- 53:59ecosystem. So give a couple billion
- 54:01dollars to Coree because they want more
- 54:03of these Neo clouds and then they go and
- 54:05turn around and spend that on chips.
- 54:07Now, Coreweave has a couple billion
- 54:09dollars of Nvidia chips that then get
- 54:11rented out to other um customers that
- 54:14now are stuck using Nvidia chips and and
- 54:17building programs on Nvidia chips. And
- 54:20on top of that, uh Nvidia does still
- 54:22get, you know, an equity stake in this
- 54:24company and then they got a margin when
- 54:26they bought their uh chips on top of
- 54:28that. So, uh it's a little confusing to
- 54:30think about, but I'm generally just not
- 54:33that worried about it because of the
- 54:34size. So, it's been a 100red billion in
- 54:36equity investments. so far. And by the
- 54:38way, a lot of that too is an anthropic
- 54:40and open AAI stake. That's uh those
- 54:42together, my estimates are are probably
- 54:44around like 30 to 40 billion each. Uh so
- 54:47that's like most of it right there. Uh
- 54:49you know, another 25 billion in future
- 54:51commitments. Not the biggest amount. Um
- 54:5336 billion in Neocloud compute um
- 54:56commitments. So this is basically
- 54:57guaranteeing uh compute contracts so
- 55:00that there's going to be demand for it.
- 55:02Nvidia is just so optimistic on how much
- 55:06AI demand there is just generally
- 55:08speaking that this is kind of very
- 55:11little relative to what they expect
- 55:12there to be and especially in this
- 55:14environment where all the compute that
- 55:16exists is basically being rented out and
- 55:18being rented out at very high uh lease
- 55:20rates. I can understand why they'd feel
- 55:23comfortable uh guaranteeing that sort of
- 55:25commitment that they'd be able to lease
- 55:26out that compute. I don't see that as
- 55:28being um very worrisome either. uh there
- 55:31these credit guarantees they're doing.
- 55:33It's an interesting way to kind of
- 55:34finance things and raise money. Those
- 55:36are uh estimated to be around hundred
- 55:38billion. So again, you add all of this
- 55:41up together and you're talking about one
- 55:43year of earnings. So this is a very cash
- 55:47flow generative business. And because of
- 55:49that, it can sustain all of these
- 55:51investments and throwing all of these
- 55:53billions around without it really, you
- 55:55know, being that impactful uh to them.
- 55:58if it gets much bigger, you know, then
- 56:00that can become worrisome as it stands
- 56:02today. I don't personally see it's the
- 56:05biggest risk with Nvidia. Now, I I think
- 56:07one of the biggest risk uh is going to
- 56:10be, as I was kind of alluding to, this
- 56:11is a pretty execution dependent
- 56:13business. They have to continue to stay
- 56:15on top of all of this. And you saw all
- 56:17of these trends that were happening over
- 56:19time that Jensen Wong was pretty good at
- 56:21at catching. you know, the shift from uh
- 56:23just chip to the importance of
- 56:25networking uh to this whole AI factory
- 56:28to AI sovereignty to um funding all of
- 56:31these neo clouds to doing these creative
- 56:34financing arrangements to make sure
- 56:35Nvidia chips continue to get sold. Um
- 56:38all of this has helped really entrench
- 56:41them. uh is, you know, even if some of
- 56:43the financing stuff makes some people
- 56:44nervous, it did get a lot more of their
- 56:47chips sold. And the more chips they get
- 56:48sold now and the more AI kind of
- 56:51continues to get developed on them, the
- 56:53more their ecosystem is kind of
- 56:55entrenched with AI and and the harder it
- 56:58could be to switch off in the future. Uh
- 57:00you want all of these models training on
- 57:02Nvidia chips. you want them running on
- 57:04Nvidia chips because Nvidia then in turn
- 57:06gets better at optimizing its chips for
- 57:09all these AI workloads and it's just
- 57:11harder and harder to catch up to them.
- 57:12So it it makes sense everything he's
- 57:15done. I think he's done an absolute
- 57:16phenomenal job setting this business up
- 57:18um competitively and he's out. He's
- 57:21gone. So the the new CEO is Colette
- 57:23Crest. Uh she's going to be an interim
- 57:25CEO until they find a replacement. She's
- 57:28been the CFO since uh 2012. Uh, but it's
- 57:32going to be very hard to replace Jensen
- 57:34Hong for no less reason other than the
- 57:36fact that he had 60 direct reports. I
- 57:39can't even imagine how confusing uh of a
- 57:42business this will be now. 60 people
- 57:44used to directly report to Jensen Hang
- 57:47and now they're, you know, probably
- 57:50reporting to her. But this is not a
- 57:52structure most people are equipped to
- 57:54handle or would want to handle. And this
- 57:56is not a normal way to run a business.
- 57:58And so I can imagine a ton of confusion
- 58:01and I can also imagine now multiple
- 58:03layers being moved above these 60
- 58:05reports and them resenting that and some
- 58:07of them probably leaving as a result of
- 58:09that. Uh because if you used to directly
- 58:12talk to the CEO to the the founder
- 58:14himself and now you have different
- 58:15people in between you that's
- 58:17demotivating. So that's going to be a
- 58:19problem. That's going to be going to be
- 58:21definitely a lot of you know operational
- 58:24uh issues to figure out within Nvidia.
- 58:26And th this is going to be something
- 58:29that I think is going to eventually lead
- 58:31to the missing things. I don't see how
- 58:33it can't. And you know, you look at all
- 58:35of these very brazen and kind of
- 58:38ambitious things that Nvidia has done.
- 58:40You know, hundred billion dollar equity
- 58:42investments, hundred billion dollar
- 58:44credit financing guarantees. You can't
- 58:46really do these things unless you're
- 58:48like the founder of the company. It's
- 58:50very hard, you know, as an interim uh
- 58:52CEO uh or even whoever comes next to be
- 58:55as brazen as uh Jensen Hang was with all
- 58:59that. So, uh I see him missing things,
- 59:01too. It's also less likely, you know,
- 59:03they're going to be looking at
- 59:04acquisitions, which have been important
- 59:05for them, uh for for the near term. It's
- 59:08kind of like a lot of things just get
- 59:10paused and they need to figure out how
- 59:12to just run this business again without
- 59:14Jensen being a part of it, which is
- 59:17should not be understated how disruptive
- 59:19that can be and how much that could lead
- 59:22to their competitive edge um faltering.
- 59:24And so I think any way you cut it,
- 59:26Jensen Hong not being there does dim the
- 59:29prospects for Nvidia in the future. I I
- 59:32think there's a much higher risk of
- 59:35disruption essentially without someone
- 59:37like him there because again he was
- 59:40really looking over the corners to the
- 59:42future multiple times with the Neoclouds
- 59:45with the financing arrangements moving
- 59:47into networking the AI factories
- 59:49sovereign AI uh the acquisitions with
- 59:52Melanex getting them into networking the
- 59:54Grock licensing getting them a position
- 59:56in some leading uh inference chips uh
- 59:59hugging face getting them a position
- 1:00:00with uh the open models and all of that
- 1:00:02and and kind of seeing where a lot of
- 1:00:04different ways can go and balancing all
- 1:00:06these different uh interest in this
- 1:00:08ecosystem. So definitely a loss uh for
- 1:00:10the business him not being there. also
- 1:00:13just a big risk because now you don't
- 1:00:15know who's running it and execution for
- 1:00:17this business is very very important
- 1:00:19right now because there's a lot of
- 1:00:22competitors who would love to take their
- 1:00:23lunch and you know I talked about how
- 1:00:25now they're trying to move to this
- 1:00:27one-year cycle or they move to a
- 1:00:29one-year cycle of uh new chip design and
- 1:00:32development now this reminds me a little
- 1:00:34bit like Intel who used to have
- 1:00:35something called tick talk where uh tick
- 1:00:39was going to be manufacturing
- 1:00:40improvements then talk was going to be
- 1:00:43architecture uh chip architecture
- 1:00:45improvements and so it went tick tock
- 1:00:48and they did one improvement every two
- 1:00:50years. So there was a full cycle done on
- 1:00:53both of them every two years and they
- 1:00:55did that for 10 years and then it
- 1:00:57stopped working and it's because it's
- 1:00:59very hard to continue to execute at a
- 1:01:02very high level and accomplishing what
- 1:01:04they're trying to accomplish where they
- 1:01:06have to constantly iterate on the last
- 1:01:08design and do better and then do better
- 1:01:11than the competition because they're
- 1:01:12priced at a premium and they need to
- 1:01:15justify that premium by being better
- 1:01:16than the competition. That is the race
- 1:01:19that they are running. And tripping up
- 1:01:21even once one year, they lose that lead
- 1:01:24forever. You never get it back. If a
- 1:01:26competitor catches up to you, it's going
- 1:01:29to be near impossible to ever take that
- 1:01:31lead again. And in fact, it's more
- 1:01:32likely that competitor surpasses you.
- 1:01:34And so that is kind of the key risk here
- 1:01:37because if they do get surpassed and
- 1:01:38they no longer have that best
- 1:01:40performance, they're going to lose their
- 1:01:42pricing power, that that margin that
- 1:01:44they've been able to to maintain, that
- 1:01:4665% operating margin. So that is another
- 1:01:49one of the key risks. And then of course
- 1:01:51I have to just tack on here you know
- 1:01:53general AI potential bubble risk all of
- 1:01:56that. Um it could be data centers stop
- 1:01:59being built because of regulation
- 1:02:00reasons. You know, there's some fears
- 1:02:02that with uh a lot of uh Democratic uh
- 1:02:05senators, governors, uh Congress that
- 1:02:08there could be some legislation that
- 1:02:10either puts a damper on AI development
- 1:02:12or at the individual state level uh more
- 1:02:15of these AI projects uh get stopped uh
- 1:02:18particular fear about that in Texas
- 1:02:20where a lot of these uh builds were
- 1:02:22happening. So, that's just uh something
- 1:02:24else that that could weigh on them
- 1:02:26because if all of a sudden these data
- 1:02:28centers are slowing down, then all of
- 1:02:30this slowly starts to unwind because all
- 1:02:33these uh capex commitments, well, they
- 1:02:35don't come through because they're
- 1:02:37you're not going to build anything and
- 1:02:38so they're not going to buy any chips.
- 1:02:40Um and so that is kind of a big risk
- 1:02:42there. And then even aside from that,
- 1:02:44this whole thing could still be
- 1:02:46cyclical, right? uh it could be the case
- 1:02:48that you can't spend hundreds of
- 1:02:50billions of dollars on on capex and
- 1:02:52continue to see a good return on
- 1:02:54investment and as soon as that starts to
- 1:02:56falter they'll pull back on their capex
- 1:02:58and everything will start to go the
- 1:03:00other way. Um and that's a very hard
- 1:03:03thing to predict. You know right now I
- 1:03:05think there's a lot of people that are
- 1:03:06generally optimistic about the return on
- 1:03:09investment. Um there's certainly some
- 1:03:11evidence if I look at Meta uh some of
- 1:03:13the hyperscalers, Amazon, Google that
- 1:03:15they are getting a good return on
- 1:03:17investment, but uh there's still a lot
- 1:03:20of experimental use and and money being
- 1:03:22spent on AI that maybe when people are
- 1:03:25not experimenting and really demand uh a
- 1:03:28full return on that that no longer uh
- 1:03:31really starts flowing through into those
- 1:03:32hyperscalers. And so ultimately what
- 1:03:34really matters is how much uh economic
- 1:03:37activity sustainable economic activity
- 1:03:40AI is really bringing in and there's
- 1:03:42debate around that. Uh these are hard
- 1:03:45things to actually track. Uh generally
- 1:03:47speaking, my opinion is that I I do
- 1:03:49think it's creating a lot of economic
- 1:03:51activity. But whether or not it is fully
- 1:03:55um enough to sustain, you know, the
- 1:03:57hundreds of billions, you know,
- 1:03:59trillions that are going to be spent on
- 1:04:00this in the coming years, that's a
- 1:04:02harder question to ask. Um, if I look at
- 1:04:05again just like Meta for instance, I
- 1:04:06could see the return on investment
- 1:04:08there, but I don't know about, you know,
- 1:04:11an individual business spending on
- 1:04:12enthropic tokens, whether or not they're
- 1:04:14getting a return on that spend or if
- 1:04:16they find out, wow, we wasting a ton of
- 1:04:18tokens, uh, wasn't really worth it.
- 1:04:20We're overpaying. Hard to say. I
- 1:04:23generally am optimistic about AI, but I
- 1:04:25also just know history and historically
- 1:04:29every time there's ever been a capital
- 1:04:31cycle, we've always overshot. We've
- 1:04:34always overbuilt demand. That's always
- 1:04:36happened. So, is this going to be the
- 1:04:38first time in history where that doesn't
- 1:04:39happen? I don't know. Or maybe this is
- 1:04:42just a very, very long cycle. I also
- 1:04:44don't know. So, uh these are just things
- 1:04:46to consider. I talk more about a lot of
- 1:04:48this though in this video I had on uh
- 1:04:51whether or not we're in an AI bubble.
- 1:04:52So, you can check that one out. But now,
- 1:04:54I'm going to move to the valuation. But
- 1:04:56first, I just want to say if you do have
- 1:04:58questions uh on Nvidia, anything else
- 1:05:00you feel like I didn't do uh a full job
- 1:05:02covering it, uh drop a question below.
- 1:05:05I'm going to have a podcast covering
- 1:05:07Nvidia next week and we could cover
- 1:05:09those questions on that podcast. Uh if
- 1:05:12you don't follow our podcast right now
- 1:05:14though, it's called the synopsis. uh you
- 1:05:16could just throw into your search bar
- 1:05:17the synopsis pod. It'll show up. Hit
- 1:05:20subscribe and then you'll uh make sure
- 1:05:22that you won't miss that video when we
- 1:05:23talk about all your Nvidia questions.
- 1:05:25Okay, so moving on to valuation. And I
- 1:05:29did tell you I was going to disappoint
- 1:05:30you a little bit, but I won't do it
- 1:05:31right now. I'll do it in a moment. Uh so
- 1:05:34at $167 stock price, it's a $4 trillion
- 1:05:37market cap. They have $192 billion in
- 1:05:40trailing earnings. That gets you an LTM
- 1:05:43uh earnings multiple of 21 times. If
- 1:05:45you're looking at forward estimates,
- 1:05:47that's going to be about 14 times
- 1:05:49earnings right now. It trades at that's
- 1:05:50going to be the next 12 months. And then
- 1:05:52if you pull up 2027 estimates, that is
- 1:05:5510 times. So trading at 10 times 2027
- 1:05:58estimates. Now what uh Jensen Hong has
- 1:06:01said is that uh they see 70% capacity
- 1:06:04growth for next year, but they see
- 1:06:06demand growth as much as 100%. So what
- 1:06:09he was alluding to there is if they
- 1:06:10could grow capacity more than 70%, they
- 1:06:12could potentially get up to 100%. Now
- 1:06:15the reason why I said I was going to
- 1:06:16disappoint you with the valuation and I
- 1:06:18am going to do this other valuation
- 1:06:20method in a moment but I think this is
- 1:06:22the more important conversation which is
- 1:06:24that before you invest in Nvidia you
- 1:06:28have to decide and consider how much you
- 1:06:31believe that they are still a cyclical
- 1:06:33chips business or whether or not things
- 1:06:36have really changed from them and
- 1:06:37they're going to be a secular grower
- 1:06:39because if it is a cyclical business it
- 1:06:43becomes very hard to value. And I
- 1:06:45mentioned in the intro, a lot of
- 1:06:46investors assue uh cyclical businesses
- 1:06:49because you can't really value them
- 1:06:52easily. You have to basically just
- 1:06:54underpay a lot and have a large margin
- 1:06:56of safety. And if that's the case, then
- 1:07:00it could work out and you kind of hope
- 1:07:02for the best. But it's it's very hard to
- 1:07:04do because imagine you're doing a DCF of
- 1:07:07a company and every year the earnings
- 1:07:10either go up or down and you don't
- 1:07:12always know whether it's up or down. It
- 1:07:14becomes very hard to have confidence in
- 1:07:16it. Whereas if a company's earnings went
- 1:07:18up every year, you could say, "All
- 1:07:21right, I don't know how much they're
- 1:07:22going to go up, but they're going to go
- 1:07:23up a little bit. Maybe I'll be
- 1:07:25conservative, assume a little bit less."
- 1:07:27And you could get comfortable with a
- 1:07:28number around that, right? And so that
- 1:07:32is why it is so much harder to value a
- 1:07:34cyclical business. And to be honest,
- 1:07:37I've never seen good valuations done for
- 1:07:40cyclical businesses because inherent in
- 1:07:42the idea being cyclical is that you
- 1:07:45don't know exactly where you are in the
- 1:07:47cycle and exactly what trough or peak
- 1:07:51earnings are going to be. These are
- 1:07:52estimates people make and then they put
- 1:07:54a multiple on these estimates. So, if
- 1:07:56you're doing it on uh peak earnings, you
- 1:07:58need to put a pretty low multiple on it,
- 1:08:00or you could put a higher multiple on
- 1:08:02trough earnings. And these multiples
- 1:08:04though in and of themselves have
- 1:08:07different earnings growth assumptions.
- 1:08:09And so, it makes it very confusing.
- 1:08:11Honestly, if you did a DCF of a company
- 1:08:13that's cyclical, you would see some
- 1:08:15years, you know, you'd grow revenues,
- 1:08:17other years you'd shrink revenues. And
- 1:08:18it becomes a very hard thing to do
- 1:08:20because what years are you going to
- 1:08:22decide to start shrinking revenues? what
- 1:08:23years are you going to decide to start
- 1:08:25growing them? These are not things you
- 1:08:27really have any idea how to answer. And
- 1:08:29so it just it's a very tricky thing to
- 1:08:31value a business like that. So generally
- 1:08:34speaking, you just want to assume a very
- 1:08:36very low multiple for a cyclical
- 1:08:38business. And chip stocks, they have
- 1:08:41traded at I think Micron at its low
- 1:08:44traded at a three times earnings
- 1:08:46multiple. And maybe that was too low.
- 1:08:48But uh that just goes to show how low of
- 1:08:51a multiple you could put when a business
- 1:08:53is cyclical and it's at peak earnings.
- 1:08:55That three times was at peak earnings.
- 1:08:57If it's at trough earnings, uh maybe
- 1:08:59it's a high singledigit multiple, maybe
- 1:09:01a low double-digit multiple. And these
- 1:09:04are just the way these things have
- 1:09:05typically traded over time. It's hard to
- 1:09:08ever say that's the right or wrong
- 1:09:09multiple though because you can only do
- 1:09:11it after the fact after you know whether
- 1:09:13or not you know earnings did come back.
- 1:09:16uh through that cycle now for at least
- 1:09:18the next few years it looks like Nvidia
- 1:09:21is in a pretty strong position uh there
- 1:09:23of are the risks that we mentioned but
- 1:09:25generally speaking you know if this AI
- 1:09:27build continues to happen they're going
- 1:09:29to be growing uh revenues still pretty
- 1:09:32quickly and a lot so maybe you look out
- 1:09:34a few years and you see what you think
- 1:09:36they could be earning and a few years
- 1:09:38from now you consider that to be peak
- 1:09:40earnings if you're paying a low multiple
- 1:09:42on that you could decide whether or not
- 1:09:44that makes sense for you so Now, to not
- 1:09:46totally leave you with nothing here, I
- 1:09:48did my own revenue build uh where I with
- 1:09:52my best analyst Claude uh started to try
- 1:09:56to figure out how to estimate gigawatt
- 1:09:59capacity out into 2031.
- 1:10:02And so this is going to be total
- 1:10:04gigawatts built. Then from there, you
- 1:10:06translate to how much of that gigawatt
- 1:10:08power is actually going to go run
- 1:10:10through chips. Then from there you make
- 1:10:13an assumption as to what Nvidia's market
- 1:10:16share of that is going to be. And then
- 1:10:18uh from there it's going to be revenue
- 1:10:21per gigawatt. And I should have backed
- 1:10:23up and said that uh if you've never
- 1:10:24listened to Jensen Huang on an earnings
- 1:10:26call he talks sometimes about uh revenue
- 1:10:30per gigawatt. And it used to be 18
- 1:10:32billion then it was 25 billion with
- 1:10:34Blackwell and then lately it's now 40
- 1:10:36billion he mentioned with Vera. And so
- 1:10:39basically as Nvidia continues to have uh
- 1:10:42more not just chips but networking
- 1:10:44technology, they're able to take over
- 1:10:46more and more of that data center,
- 1:10:48provide more of all of the equipment you
- 1:10:50need to run that data center. Their
- 1:10:52revenue per gigawatt has gone up. So
- 1:10:55that's kind of how he thinks about it.
- 1:10:56So I just jumped on that and I built
- 1:10:58this this revenue model based off of
- 1:11:00that. So I kind of just walked you
- 1:11:01through how that worked. You're going to
- 1:11:03see on the screen now uh kind of the the
- 1:11:05math and the the assumptions how that
- 1:11:07all flows through. I'm going to just
- 1:11:09talk about that base case scenario. So
- 1:11:11uh you could see there it jumps from
- 1:11:14data center gigawatts it converts to how
- 1:11:16much uh is actually going to be on the
- 1:11:18chips. Then it also adds in replacement
- 1:11:20gigawatts. So the idea here is that and
- 1:11:24this is actually very important to
- 1:11:26understand um for Nvidia's potential um
- 1:11:30recurring revenues essentially or how
- 1:11:32durable their revenues will be is how
- 1:11:34much of this gets replaced every four to
- 1:11:37six years and so if chips have a four to
- 1:11:39six year life something like that maybe
- 1:11:41seven whatever it is and they get
- 1:11:43replaced o over that period then you
- 1:11:46have a very different business where you
- 1:11:48do start seeing recurring revenue and it
- 1:11:50reminds me a little bit of when I
- 1:11:51invested in Apple and uh at the time
- 1:11:54people would say oh well people are
- 1:11:55going to buy an iPhone and then you know
- 1:11:58they're not going to want another iPhone
- 1:11:59for many years and I remember thinking
- 1:12:02no like people are going to want you
- 1:12:03know an iPhone every couple years
- 1:12:05they're going to keep wanting the new
- 1:12:06ones and even if maybe that gets a
- 1:12:08little bit delayed over time it's
- 1:12:11generally held true people still just
- 1:12:13want the new the new phones and uh the
- 1:12:15question is whether or not uh data
- 1:12:17center GPUs are going to be similar if
- 1:12:19they could continue to roll out new
- 1:12:21performance uh GPUs that justified the
- 1:12:24data centers replacing them. Um and they
- 1:12:26could continue to get good performance
- 1:12:28uh per uh cost per token on that um per
- 1:12:32gigawatt, whatever metric you want to
- 1:12:34put there on the denominator. Um then
- 1:12:37they'll continue to replace them. And so
- 1:12:39that's what this build has. It has
- 1:12:41replacement gigawatts in there too. This
- 1:12:42is something I've thought a lot about
- 1:12:44though for Nvidia because uh it's kind
- 1:12:46of like we're in this phase where you
- 1:12:47you have this big buildout and
- 1:12:49everything is growing and then there's
- 1:12:51going to be a maintenance phase. But
- 1:12:52when the maintenance phase happens,
- 1:12:55revenues are going to drop a lot, right?
- 1:12:57They're going to drop a lot because
- 1:12:59you're you have a lot less growth and
- 1:13:00then it's just you are replacing what's
- 1:13:02already built. Having said that, if when
- 1:13:05this happens, uh it's already a much
- 1:13:08much bigger base than where they
- 1:13:10currently are at, then even when you get
- 1:13:12to that maintenance level, it could
- 1:13:14potentially potentially uh justify, you
- 1:13:18know, higher revenues than where they
- 1:13:19are today. It seems possible possible
- 1:13:22that they could be at, you know, a
- 1:13:24higher, you know, maintenance uh level
- 1:13:26of revenue than where they are in total
- 1:13:28revenue today. I think that's a pretty
- 1:13:30ambitious assumption. But even if they
- 1:13:33are less, but it's now recurring revenue
- 1:13:36that you have much higher confidence in,
- 1:13:38then maybe it deserves a higher
- 1:13:39multiple. Uh when people see that, okay,
- 1:13:42so this business did, you know, contract
- 1:13:44a little bit, but it didn't contract,
- 1:13:46you know, that much and now it looks
- 1:13:48more stable and now we're coming back
- 1:13:50up. Okay, now we kind of have more of a
- 1:13:52sense of what a cycle looks like. That's
- 1:13:54something that could happen, too. So
- 1:13:56anyway, in this build, we have the
- 1:13:57replacement gigawatts. We assume
- 1:13:59Nvidia's share of it. Um, all these
- 1:14:01assumptions, by the way, this is a
- 1:14:03framework. You could come up with your
- 1:14:04own ideas. You could also literally run
- 1:14:07this math that gets to revenue and then
- 1:14:09you could do it for every year if you
- 1:14:10wanted to and project it out and build
- 1:14:12your own DCF that way. That's that's
- 1:14:14something you could do. Um, I don't know
- 1:14:16if it's going to get you to a better
- 1:14:17result, but if you feel like modeling a
- 1:14:20little bit to make you feel better. Um,
- 1:14:22so that's going to be uh Nvidia share we
- 1:14:25have 55%. This assumes 40 billion
- 1:14:28gawatt. Seems a little high. Um, but
- 1:14:30this is just an assumption. That gets
- 1:14:32you to uh about a trillion in data
- 1:14:35center revenue plus another 50 billion
- 1:14:38in edge revenue. That's going to be a
- 1:14:40$460 billion NOPAT uh figure. So it
- 1:14:43calls this a base case. I think that
- 1:14:45this is still uh seems like a lot to me.
- 1:14:48A billion trillion dollars in revenue.
- 1:14:50Sorry. Um, but I don't know, everyone's
- 1:14:52throwing big numbers out there these
- 1:14:54days. So, if you take this math and then
- 1:14:57you want to put a range of multiples on
- 1:14:59it, let's say you do 15 times multiple,
- 1:15:02that'd get you about a 7 trillion
- 1:15:04valuation. 20 times a multiple would be
- 1:15:06a $9.2 trillion valuation. 25 time
- 1:15:09multiple 11.5 trillion. uh over this
- 1:15:13period from today to uh 2031, this is
- 1:15:16the five years out we're doing, they
- 1:15:18will have collected $1.7 trillion in
- 1:15:21cash. If you assume that that cash is
- 1:15:23being returned to investors as it's
- 1:15:25being generated, a 15 times multiple
- 1:15:27would get you an 18% annualized return.
- 1:15:30A 20 times multiple would get you a 24%
- 1:15:33annualized return and a 25 times
- 1:15:35multiple would get you 29%. Remember
- 1:15:38though, 25 times this is a stock trading
- 1:15:40at you know uh 14 times forward 12 month
- 1:15:44earnings right now. So 15 times though
- 1:15:46that's at 18% right there. The other
- 1:15:48thing I want to note though is that
- 1:15:49these base cases as I was saying they
- 1:15:51are still bullish. It assumes that this
- 1:15:53is a long AI cycle. Uh Nvidia maintains
- 1:15:57their dominance. They maintain their
- 1:15:58margins. They maintain their pricing
- 1:16:00power. Uh the bare case I threw in there
- 1:16:02had earnings uh of $118 billion which is
- 1:16:06you know earnings getting cut almost in
- 1:16:08half. So at that point it would trade
- 1:16:10high30s multiple something like that and
- 1:16:13if it's now proven to see cyclical uh
- 1:16:15probably not going to be able to
- 1:16:16maintain that multiple. So the stock
- 1:16:18would probably get cut um pretty
- 1:16:20severely at that point. Um, the thing is
- 1:16:23that these numbers are very out in the
- 1:16:27air and very hard to predict. And I
- 1:16:29think this is part of the reason why to
- 1:16:31Nvidia trades at the multiple it does,
- 1:16:33right? It's not, you know, a crazy
- 1:16:35multiple to trade at 10 times 2027
- 1:16:38earnings. I think it trades at that
- 1:16:39multiple because a lot of people have a
- 1:16:42lot of questions as to what's going to
- 1:16:43happen in the future, how long this AI
- 1:16:46cycle can really go. And so I think this
- 1:16:49is an opportunity for an investor if
- 1:16:52they do want to take an opinion on that
- 1:16:54to say that I do think that you know
- 1:16:57there's going to be a little bit more
- 1:16:58demand. I think that uh AI is going to
- 1:17:01continue to have very high return on uh
- 1:17:04invested capital for all players and we
- 1:17:06need to build out way more capacity and
- 1:17:08this AI buildout is going to happen for
- 1:17:10many many more years and Nvidia is going
- 1:17:13to continue to be instrumental to that
- 1:17:15buildout. that could be um a situation
- 1:17:18where an investor finds this to to be
- 1:17:20interesting. So, I feel like I kind of
- 1:17:22failed you a little bit with the
- 1:17:23valuation section here, but I think
- 1:17:26that's because you really shouldn't try
- 1:17:29to force too much precision from
- 1:17:33something that is inherently imprecise.
- 1:17:36And that's the issue with cyclical
- 1:17:38businesses because if you try to run a
- 1:17:40DCF on it, it's just going to be
- 1:17:43malarkey, whatever that word Joe Biden
- 1:17:46said was instead of BS. Um
- 1:17:49because whatever your inputs are,
- 1:17:50they're going to be wrong. And you can't
- 1:17:52even make good conservative assumptions
- 1:17:54when trying to do that. That is kind of
- 1:17:56the the issue I have with modeling these
- 1:17:58sorts of things. And so instead what you
- 1:18:01can do is you can look out a couple
- 1:18:04years where you feel comfortable with
- 1:18:06which is maybe you know to 2027 let's
- 1:18:08say and you could see what the earnings
- 1:18:10multiple is on that year that if the
- 1:18:12estimates are correct that's going to be
- 1:18:13a 10 times multiple and then you could
- 1:18:15decide if that makes sense to you or
- 1:18:18not. Now, to give you some context, that
- 1:18:22would be relative to history for chip
- 1:18:25multiples, not as cheap as they get for
- 1:18:28um a trough multiple. So, if you thought
- 1:18:31that those were in earnings peak at
- 1:18:332027, you're going to want to pay a
- 1:18:35lower multiple than that. Uh if you
- 1:18:37thought that there's still going to be
- 1:18:38growth thereafter, then okay, then maybe
- 1:18:41that makes sense. As always, I'm going
- 1:18:43to leave it to you to make that
- 1:18:44decision. But uh a couple things worth
- 1:18:47noting is one, if you're not sure about
- 1:18:49something, there's nothing wrong with
- 1:18:51passing on it. Two, there's also not a
- 1:18:55reason you should need to feel a need to
- 1:18:58force uh certainty from something that
- 1:19:00is uncertain. Uh and three, I think one
- 1:19:04of the biggest issues right now is with
- 1:19:07a company like this, you would really
- 1:19:10want to have comfort in the person
- 1:19:13running it. and knowing that well uh
- 1:19:18even if I can only see out a couple
- 1:19:19years, at least the person in charge can
- 1:19:22see out many more years into the future
- 1:19:24and can run this business and can
- 1:19:26navigate it. And of course, you know
- 1:19:27where I'm going with this. The issue is
- 1:19:29we don't know who that person is. Uh
- 1:19:31because there is not a permanent CEO of
- 1:19:34the business yet. It's up for you though
- 1:19:36to decide whether or not that is more
- 1:19:39than priced in today with the stock down
- 1:19:4130% trading at the multiples it is
- 1:19:44trading at for still a pretty dominant
- 1:19:46franchise in an extremely uh important
- 1:19:50area in the world with a lot of
- 1:19:52tailwinds going for it and growing
- 1:19:54phenomenally quickly. Maybe that can
- 1:19:57make up for them in the short term and
- 1:20:00buy them some time to figure things out
- 1:20:02internally uh and get a new leader. But
- 1:20:05no doubt it makes it much more
- 1:20:07uncomfortable as an investment without
- 1:20:09Jensen Huang there. But maybe that's why
- 1:20:12the opportunity exists. That'll be for
- 1:20:15you to decide. But I will say one last
- 1:20:17thing which is that when I was doing my
- 1:20:20AMD um valuation, I remember at the time
- 1:20:23it was trading at around 10 times 2030
- 1:20:27earnings. And it's possible at the time
- 1:20:29that you know maybe investors really
- 1:20:31thought they were undershooting their
- 1:20:32estimates. So the multiple would have
- 1:20:34been lower at uh for 2030. But that is
- 1:20:37just one data point to give you a little
- 1:20:40bit of context. But if you want much
- 1:20:41more on AMD, go ahead and check this
- 1:20:45video out here. But if you want to know
- 1:20:47what's much more important than picking
- 1:20:49stocks, watch this video here right now.
- 1:20:52Thank you for watching and until next
About this transcript
This page contains the full transcript of Nvidia Stock Crashed — Understand It Better Than 99% of Investors by Drew Cohen, generated from the public captions YouTube serves with the video. The transcript has 15,230 words across 2,137 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.