Latest / Deep Values: Value Investing Deep Dives / NVIDIA's Q2 Earnings Explosed Monopoly Beyond The Shovel Analogy
Transcript
- 0:00So when people talk about the gold rush, you know, the cliche
- 0:02you always hear is don't mind the gold, sell the shovels.
- 0:07It's clean, it makes sense, and it's comforting.
- 0:09But when you look at the current computing landscape, that shovel
- 0:13analogy just falls wildly short. We aren't just looking at a
- 0:17company selling shovels. I mean, we are looking at an
- 0:19entity that somehow owns the only shovel factory on earth,
- 0:23and they dictate the exact technique every miner must use
- 0:27to dig. And on top of that, they on the
- 0:30roads the miners take to get to the mountain.
- 0:32We are looking at NVIDIA and trying to calculate the actual
- 0:36structural reality of a monopoly like this requires looking way
- 0:40past the daily hype cycle. But before we get into the
- 0:44weeds, I have some required reading here to access more
- 0:48advanced researches, follow Guru's latest movement out and
- 0:51leverage our latest Deep Values agents.
- 0:53Welcome to download Deep Values APP from App Store to track
- 0:56stock market temperature. Welcome to download iOS app
- 0:59Buffett Indicator from Apple App Store or Bubble Watch from
- 1:02Google Play. Links in the description.
- 1:03Disclaimer All analysis is based on publicly available sources
- 1:07and may be subject to revisions or differing interpretations.
- 1:10All contents from this channel is for research and
- 1:12entertainment purposes only and does not constitute any
- 1:15investment or financial advice or service.
- 1:18Conduct your own due diligence before making investment
- 1:20financial decisions. OK, let's unpack this.
- 1:23Yeah, a good place to start the unpacking is with a fundamental
- 1:26reset of what this business actually is today.
- 1:28Right, Because public perception is severely lagging the
- 1:31corporate reality. Like a large portion of the
- 1:34market still views NVIDIA strictly through the lens of a
- 1:36chip supplier. You mean like the graphics card
- 1:39you slot into APC? Exactly the ones used to render
- 1:41hyper realistic video games. Which?
- 1:44I mean, that used to be their core identity.
- 1:45It did, but looking at the historical data sets across the
- 1:49last decade, gaming is now hovering around just 9 percent
- 1:52of their revenue. Wow, just 9.
- 1:54Percent, yeah. The other 85 to 90% is entirely
- 1:57driven by the data center. OK, but here's what I want to
- 2:00push on with that selling chips to a data center instead of a
- 2:04teenager in their basement playing games.
- 2:07It doesn't seem like enough to explain evaluation in the
- 2:10trillions. Well, it wouldn't be if they
- 2:13were only selling the chips. The pivotal shift.
- 2:16What's fascinating here is that they have transformed into an
- 2:19end to end data center systems architecture provider.
- 2:22OK, architecture provider. Right, They are no longer just
- 2:26handing a tech giant a tray of advanced GPU's like the Hopper
- 2:29or the new Blackwell chips and just wishing them luck.
- 2:32They are selling the entire nervous system.
- 2:34But wait, selling the network cables along with the chips
- 2:38doesn't sound like a revolutionary pivot.
- 2:41Like why does controlling the networking infrastructure matter
- 2:43so much? It comes down to the actual
- 2:46physics of how artificial intelligence is trained today.
- 2:49OK, when a company is training a massive AI model, they aren't
- 2:53using one chip, they are linking 10s of thousands of GPU's
- 2:57together. Wow, 10s of thousands.
- 3:00Yeah, and these chips have to constantly talk to each other.
- 3:03They are sharing pieces of the neural network back and forth in
- 3:06fractions of a second. I see.
- 3:08So if the network connect, connecting those chips is slow.
- 3:10It literally doesn't matter how fast the individual GPU is.
- 3:13Because it's just sitting there waiting.
- 3:14Exactly. A $40,000 processor will
- 3:17literally sit idle waiting for data to arrive.
- 3:21Okay, so NVIDIA realized that third party network bottlenecks
- 3:25were basically making their expensive chips look slow.
- 3:28Precisely. So.
- 3:30To solve this data starvation problem, they just developed
- 3:33their own networking infrastructure, things like
- 3:35Quantum InfiniBand and Spectrum X Ethernet.
- 3:38Oh, I've heard of InfiniBand. Yeah, they engineer the
- 3:41communication pathways to be optimized specifically for their
- 3:44own silicon and that networking component alone is now pulling
- 3:48in multi billion dollar annualized run rates.
- 3:52That's incredible. Yeah, they control the brain
- 3:55computing the data and the nervous system transmitting it.
- 3:58Right, but controlling the physical hardware is 1
- 4:00something. The layer that seems to make
- 4:02them practically untouchable, at least from what I'm reading, is
- 4:05the software. Oh, absolutely.
- 4:07Because. People throw the word mode
- 4:08around constantly in finance, you know, to describe a
- 4:12competitive advantage. But Nvidias's software platform
- 4:15CDA, It just seems to operate on a completely different level.
- 4:19It does. CDA is arguably the most
- 4:21powerful asset they possess. I mean even more than the
- 4:24physical silicon designs, it is the invisible glue holding their
- 4:28entire monopoly together. The way I was visualizing this
- 4:31while going through the research today is that CDA is essentially
- 4:35the English language of the AI world.
- 4:37That's a great way to put it. Like imagine if every single
- 4:41developer, researcher, and data scientist on earth has spent the
- 4:45last two decades writing their code and communicating
- 4:48exclusively in English, and now a competitor comes along and
- 4:52says hey, I have this brand new, slightly cheaper hardware you
- 4:56can run your programs on, but the only catch is you have to
- 4:59translate your entire life's work into a brand new language
- 5:02you've never spoken before. Yeah, and very few developers
- 5:05are going to take that deal. You are describing the exact
- 5:08mechanics of switching costs, right?
- 5:10Over the past 20 years, millions of developers have built highly
- 5:14optimized globally shared software library specifically
- 5:18for the CUDA platform. Because this would be the only
- 5:20game in town. Exactly.
- 5:21So when you try to take code written for CUDA and run it on a
- 5:25rival chip, say from AMD or Intel, it doesn't just work
- 5:28natively. You have to run through a
- 5:29translator. Right.
- 5:30And translation introduces friction and in the computing
- 5:33world that friction is known as overhead.
- 5:36It slows the execution time down.
- 5:38So if you are a massive tech company spending billions of
- 5:41dollars on hardware, and your goal is to be the fastest to
- 5:44train a new AI model, you really cannot afford software
- 5:48translation overhead. Because time is money.
- 5:50Huge amounts of money. The slight discount you might
- 5:53get on a competitors chip is entirely wiped out by the drop
- 5:57in performance plus the labor cost of rewriting your code.
- 6:01Wow. OK.
- 6:02So they own the language of AI, they own the computing brains
- 6:05and they own the nervous system connecting them.
- 6:06Yes. And that monopoly like dominance
- 6:10is, I mean it is aggressively reflected in the financial data
- 6:13we have here. Very aggressively.
- 6:15Yeah. Like in fiscal year 2016, their
- 6:17revenue was sitting around $5 billion.
- 6:20Fast forward to fiscal year 2025, and that revenue has
- 6:23exploded past $125 billion. It's hard to even wrap your head
- 6:27around that scale of growth. It really is, but here is the
- 6:30mechanism. I don't fully grasp if they are
- 6:32constantly inventing these complex multi die architectures
- 6:36and engineering advanced data center systems, shouldn't their
- 6:39capital expenditures and operating costs be skyrocketing
- 6:43right alongside their revenue? You would think so, yeah.
- 6:46How are they extracting so much actual profit out of this
- 6:49physical hardware? Well, the secret lies in the
- 6:52structure of the fabulous business model.
- 6:55Fabulous. Right.
- 6:56NVIDIA designs the chips, they architect the systems, and they
- 7:00write the software, but they do not physically manufacture the
- 7:03silicon wafers themselves. Oh, I see.
- 7:05Yeah, they outsource the actual physical fabrication primarily
- 7:08to the Taiwan Semiconductor Manufacturing Company or TSMC.
- 7:12So they basically get to skip building the giant billion
- 7:15dollar fabrication plants. Exactly.
- 7:17Which completely insulates them from heavy depreciation, right?
- 7:21Manufacturing semiconductors requires buying lithography
- 7:24machines that cost hundreds of millions of dollars each.
- 7:27By letting TSMC carry the burden of buying those machines and
- 7:30building those factories, NVIDIA keeps their balance sheet
- 7:33incredibly light. That is wild.
- 7:36Depreciation expenses currently sit at less than 5% of their
- 7:39gross profit. Oh wow, so it's basically like
- 7:42they write the recipe and they pay someone else to buy the
- 7:45evidence. That's exactly it.
- 7:46And because the market is absolutely desperate for their
- 7:49specific recipe, NVIDIA has total pricing power, right?
- 7:53When you combine that high pricing power with an outsourced
- 7:56asset light manufacturing base, your profit margins naturally
- 8:01expand to staggering levels when demand spikes.
- 8:04And the demand spike is historic, right?
- 8:06Oh, definitely. Looking at the sheet, net income
- 8:09growth hit 581% in fiscal year 2024, followed by another 128%
- 8:15growth in fiscal year 2025. And those margins you mentioned,
- 8:19they have stabilized around 74 to 75%.
- 8:22Which is just crazy. Retaining $0.75 of gross profit
- 8:26for every dollar of physical hardware you sell is virtually
- 8:29unheard of in the tech hardware space.
- 8:31It is. It has generated over $60
- 8:33billion in free cash flow, creating a fortress balance
- 8:37sheet where their current assets cover their liabilities by more
- 8:41than four times. So near term solvency is
- 8:44practically a non issue with that much cash generation.
- 8:47Not an issue at all. They're deploying it into
- 8:50aggressive share buybacks and tactical dividends.
- 8:53Financially, the engine is running Florida State.
- 8:55But retaining $0.75 on every dollar is spectacular for
- 8:59NVIDIA, while in capitalism a 75% margin is basically a giant
- 9:04target on your back. Oh.
- 9:05For sure, everyone wants a piece of that.
- 9:07It practically begs your customers and competitors to
- 9:10find a cheaper way to operate because no Moat is entirely
- 9:13invincible. So if we look at the structural
- 9:16vulnerabilities in the research, the headwinds seem to split into
- 9:19two distinct realities, things that challenge them today and
- 9:23things that could permanently fracture the business over the
- 9:25next decade. Let's start with the immediate
- 9:29physical bottlenecks. OK, so the most pressing short
- 9:32term physical bottleneck is memory supply.
- 9:34Memory supply. Yeah, we discussed how standard
- 9:37networking cables starve the GPU's of data, right, Right.
- 9:41Well, the same physics apply to the memory on the chip itself.
- 9:45To feed the GPU fast enough, NVIDIA relies on High Bandwidth
- 9:49Memory, or HBM. OK so standard memory is like a
- 9:52single lane highway. So they need a multi lane
- 9:54highway. Better yet, imagine stacking
- 9:56standard memory chips vertically, like floors in a
- 9:59skyscraper. Oh good, and placing that
- 10:02skyscraper mere millimeters away from the computing core.
- 10:06By doing that you drastically shorten the physical distance
- 10:09the electrical signals have to travel.
- 10:11I see. So it's all about distance.
- 10:12Yes, it allows massive amounts of data to flow simultaneously,
- 10:16but manufacturing that stack HBM is incredibly difficult I
- 10:19imagine, and NVIDIA relies entirely on a very small pool of
- 10:23suppliers like SK Hynix and Micron to build it.
- 10:26If those suppliers hit a production snag, NVIDIA
- 10:29literally cannot ship their completed GPU.
- 10:31Wow, so they are totally dependent on someone elses
- 10:34factory for that too? Exactly.
- 10:35Then there is the issue of just keeping the hardware from
- 10:38melting. Right.
- 10:38Oh yeah, thermals. The thermal dynamics of these
- 10:41new Blackwell chips are pushing the limits of data center
- 10:44infrastructure. They are because the computing
- 10:47density is so incredibly high, just blowing cold air across the
- 10:51servers no longer works. Right.
- 10:53It's like trying to cool down an oven with a desk fan.
- 10:55That is exactly what it is, so the industry is being forced to
- 10:58transition to rack scale liquid cooling.
- 11:01Liquid cooling. Water or specialized fluids are
- 11:04pumped directly to the chips to absorb the thermal density.
- 11:07But piping liquid through data centers full of electrical
- 11:10equipment, I mean that that introduces severe engineering
- 11:13challenges. Oh massive, you have leak risk,
- 11:15massive yield tuning issues as NVIDIA tries to ramp up
- 11:19production on these liquid cooled systems.
- 11:21Any engineering hiccups directly delay their revenue.
- 11:25Right. And even if they build them
- 11:26perfectly, there is the lingering threat of a
- 11:29hyperscaler CapEx digestion period.
- 11:32Yes, the digestion period. Because the hyperscalers, you
- 11:35know, Google, Amazon, Microsoft, Meta, they have been front
- 11:37loading their capital expenditure spending 10s of
- 11:41billions on NVIDIA chips out of fear.
- 11:44It is essentially an arms race. They are terrified of missing
- 11:48the AI wave. But eventually all that capital
- 11:51expenditure must be justified by actual consumer demand and
- 11:54software. Revenue, right?
- 11:55They have to actually make money off these models.
- 11:57Exactly. If these T1 cloud providers
- 12:01realize they have purchased more computing power than they can
- 12:04immediately monetize, they will pause their orders to digest the
- 12:08inventory they already have. A sudden 12 to 18 month pause in
- 12:13purchasing from just three or four major companies would cause
- 12:17Nvidia's revenue to drop significantly.
- 12:19That makes total sense. So those are the near term
- 12:21cyclical threats. But let's look at the
- 12:23existential threats, OK, The permanent structural risk, they
- 12:26could cap their growth entirely. The first one is a matter of
- 12:29global policy, right? The US Department of Commerce
- 12:32has implemented stringent export controls on high performance
- 12:35compute silicon, yes, and this directly targets shipments to
- 12:38China and adjacent markets. Looking at this purely from a
- 12:42mathematical and business reality, setting aside any
- 12:46political philosophy, these restrictions represent a
- 12:49permanent curtailment of their total addressable market.
- 12:52Right, because a large chunk of the globe is simply off limits
- 12:56for their highest margin most advanced products.
- 12:59It just creates an artificial ceiling on their customer base.
- 13:01You can't sell to them, period. Right?
- 13:03But the second long term structural risk goes back to the
- 13:06very fabulous model that makes them so profitable today.
- 13:09The foundry concentration. It's like running the most
- 13:12highly sought after, wildly profitable restaurant in the
- 13:15world, but you rely on exactly 1 farm for all your ingredients,
- 13:21which is TSMC in Taiwan, right? If that farm is a bad harvest,
- 13:25an earthquake, or a geopolitical disruption, your restaurant has
- 13:29absolutely nothing to serve, no matter how many people are lined
- 13:32up at the door. That single geography reliance
- 13:34is the glaring vulnerability of the fabulous model.
- 13:38But we need to take that analogy a step further.
- 13:40Not only are you relying on one single farm, but some of your
- 13:44biggest customers, the hyperscalers, eating at your
- 13:47restaurant every night, right? They are actively building their
- 13:50own kitchens in the back. Oh man, the custom ASICS.
- 13:53Yes. Google's TPU.
- 13:55Amazon's Tranium. Microsoft's Maya, the very
- 13:59companies driving Nvidia's $125 billion revenue, are
- 14:04aggressively trying to replace them.
- 14:06And they are doing it because of that 75% gross margin we talked
- 14:09about earlier. They want that margin for them.
- 14:11To be Google and Amazon don't want to pay the NVIDIA premium
- 14:15forever, so they are designing application specific integrated
- 14:19circuits, or ASIC. Yeah, a GPU is a generalist.
- 14:23It is incredibly powerful because it can do almost
- 14:25anything. And ASIC is a specialist.
- 14:27So if Google just needs to run one highly specific search
- 14:30algorithm or recommendation engine, they don't need the
- 14:33broad flexibility of an NVIDIA GPU.
- 14:35They can design an ASIC that only does that one specific
- 14:39task, which ends up being significantly cheaper to produce
- 14:42and consume assumes far less power.
- 14:44Wow, so the hyperscalers successfully migrate their
- 14:48internal workloads to their own custom silicon.
- 14:52NVIDIA loses its biggest buyers. That's the threat.
- 14:55So we have an incredibly wide software mode and flawless
- 14:59current financials, but it's balanced against a vulnerable
- 15:02physical supply chain, capped global markets and customers who
- 15:07are actively plotting to become competitors.
- 15:09It's a lot to weigh. It is, right.
- 15:11So how do you possibly put a price tag on a business in that
- 15:14situation? Well, you apply a strict
- 15:16valuation framework, OK. And if we run NVIDIA through
- 15:19traditional Benjamin Graham value investing criteria, the
- 15:22dashboard flashes bright red immediately.
- 15:25Right, because the metrics that value investors have used for
- 15:28decades just breakdown here. They do.
- 15:30I mean the price to earnings or PE ratio frequently sits above
- 15:3335. Yeah.
- 15:34The price to book ratio is also asked economically high.
- 15:37Traditional value investing looks for a PE under 15,
- 15:40demanding that you pay a low price for current earnings,
- 15:43right. So by those strict historical
- 15:44standards, the stock is wildly overpriced.
- 15:47It is, however, a purely historical lens.
- 15:49Ignores growth. OK.
- 15:52And that is why it passes the Peter Lynch PEG ratio, the PEG
- 15:56ratio, yeah, the PEG ratio takes that high PE multiple and
- 16:00divides it by the earnings growth rate, right?
- 16:02Because Nvidia's earnings have been growing at triple digit
- 16:05percentages, the PEG ratio drops below 1.
- 16:09Oh, I see. So the sheer velocity of their
- 16:12growth mathematically justifies the high multiple in the short
- 16:15term. Furthermore, they are generating
- 16:18a return on invested capital, or ROIC, of over 60.
- 16:23Meaning they are incredibly efficient with the cash they do
- 16:26keep. Exactly proving that every
- 16:28dollar they retain is deployed with brutal efficiency.
- 16:30But the true test of intrinsic value is the discounted cash
- 16:34flow model, the DCF, right, Because we are trying to figure
- 16:37out exactly what one share of this company is actually worth
- 16:40based on the cash it will produce in the future.
- 16:42So the deep value research here runs three different DCF
- 16:45scenarios based on a starting free cash flow of $60 billion
- 16:50and they use a 10 discount rate, which essentially accounts for
- 16:54the oortunity cost and the risk of holding the stock versus a
- 16:58safe asset. Let's walk through what these
- 17:00models actually assume. OK.
- 17:02So the first scenario is the 0 growth case.
- 17:04OK? This models a reality where
- 17:07Nvidia's cash flows flatline today.
- 17:10They never innovate another breakthrough, and they just
- 17:12maintain that 60 billion annually in perpetuity.
- 17:15OK? Under that strict mathematical
- 17:17constraint, the intrinsic value of the stock calculates to
- 17:20$33.36 per share. Which is a baseline floor
- 17:24exactly. No one actually expects 0
- 17:27growth. No.
- 17:28What does the second scenario, the conservative case, assume
- 17:32about the world? So the conservative case models
- 17:35a cyclical mean reversion. Mean reversion.
- 17:37Yeah, this represents a world where the hyperscalers do
- 17:40exactly what we discussed. They realize they overbuilt and
- 17:43they pause their orders. On the digestion period.
- 17:46Right. The model assumes a 15% drop in
- 17:49cash flow next year as the industry digests its inventory,
- 17:53remaining flat in year 2 and then slowly recovering with 10%
- 17:57growth in years 3 through 5. OK, under this highly
- 18:00conservative cyclical assumption, the intrinsic value
- 18:04is $36.02. Wow, OK, so even with a recovery
- 18:08built in, the intrinsic value only moves up slightly because
- 18:11those initial down years destroyed so much compounded
- 18:15value. That's the math.
- 18:16Yeah, give me the absolute bull case.
- 18:18What happens if the AI infrastructure build out
- 18:21continues uninterrupted? That is scenario 3, the nominal
- 18:24case of sustained expansion. OK.
- 18:25This assumes 25% annual cash flow growth for the next two
- 18:29years, decelerating slightly to 15% for years 3 and 4 and 10% in
- 18:35Year 5. That is massive growth.
- 18:37It is. And if you run those aggressive
- 18:39uninterrupted growth numbers through the DCF, the intrinsic
- 18:42value jumps to $70.91 per share. And this is where the mass
- 18:46really collides with market psychology.
- 18:47Yeah, it is. Because according to the data,
- 18:50the stock is trading in the public markets well above
- 18:53$120.00. Yep.
- 18:55So even in the absolute rosiest uninterrupted growth scenario we
- 18:59just ran, the intrinsic value only reaches around $71.
- 19:03That's right. And take that a step further, a
- 19:06strict value investor doesn't just buy at intrinsic value,
- 19:09right you? Need a buffer.
- 19:10They demand a margin of safety, usually around 30%, to protect
- 19:14their capital in case their assumptions are wrong.
- 19:17So if you apply a 30% margin of safety to that best case $70.91
- 19:23scenario, your actual buy target is roughly $49.64.
- 19:28Which perfectly illustrates the value investors dilemma here.
- 19:32The underlying business is a structural marvel.
- 19:35The operating metrics are spectacular, but the public
- 19:38market price that 120 plus figure, it demands absolute
- 19:42perfection. Leaves no room for error.
- 19:44Exactly. It prices in not just the hyper
- 19:47growth we've already seen, but a future of continuous,
- 19:50uninterrupted dominance that frankly defies the gravity of
- 19:54traditional valuation models. The market is just stripped
- 19:57away. Any cushion?
- 19:58I mean if HBM memory yields drop, or if liquid cooling pipes
- 20:02leak in a major data center, right?
- 20:04Or if Amazon scales up its custom tranium chips faster than
- 20:07expected. The current stock price has no
- 20:10structural support to absorb that bad news.
- 20:12None. For an investor requiring
- 20:15capital protection relative to structural earnings power, the
- 20:18conclusion from the data is quite stark.
- 20:20What is it? You would basically have to sit
- 20:22on your hands and wait for a severe broad market cyclical
- 20:26pullback to enter the stock safely.
- 20:28You are looking at a fundamentally flawless business
- 20:31carrying an immense amount of expectation risk.
- 20:33So to synthesize this entire architecture for a second,
- 20:36NVIDIA has successfully navigated the most lucrative
- 20:39pivot in tech history, moving from video game graphics to
- 20:42becoming the undisputed architect of the modern AI data
- 20:45center. Their financial metrics display
- 20:47unprecedented scaling, and they are protected by a software
- 20:51language in CEA that the entire world is literally forced to
- 20:55speak. Yet they remain heavily bound by
- 20:58the physical constraints of planet Earth, relying on a
- 21:01single island for fabrication, specific suppliers for stacked
- 21:06memory, and the literal thermodynamics of keeping the
- 21:09silicon from melting. It's quite the paradox.
- 21:12It is, and they are priced at a premium that assumes none of
- 21:15those physical risks will ever materialize.
- 21:18Yeah, they are operating at the absolute bleeding edge of human
- 21:21engineering, but they are ultimately subject to the
- 21:24unforgiving laws of financial mathematics.
- 21:26Which leaves me was one final thought to ponder here.
- 21:28What's that? We've established that Nvidia's
- 21:31deepest Moat, their stickiest asset by far, is the CEA
- 21:34software ecosystem. But their biggest bottlenecks?
- 21:38Taiwan HBM memory, liquid cooling.
- 21:40They are entirely physical, right?
- 21:42So at what point does the immense value of the software
- 21:45completely detach from the hardware?
- 21:47Oh, that's interesting. If the entire globe is, natively
- 21:50speaking, the language of AI, could that language eventually
- 21:53outgrow the very silicon it was built to run on?
- 21:56Wow, that tension between software dominance and physical
- 22:01constraints that is going to define the next decade of
- 22:03computing. Something to think about the
- 22:05next time you hear a new AI announcement.
- 22:07Thanks for joining us on this deep dive.
- 22:09We'll see you next time.