Latest / Deep Values: Value Investing Deep Dives / NVIDIA’s Software Moat and Valuation Paradox
Transcript
- 0:00Imagine a company manufacturing like heavy physical metal boxes,
- 0:05right? And they're shipping them across
- 0:07the world on cargo planes, managing this massive, complex
- 0:11global supply chain. Which is incredibly expensive to
- 0:14do. Exactly.
- 0:15But somehow at the end of the quarter, they pull in the profit
- 0:18margins of like a pure software application.
- 0:21Yeah, which historically, I mean, that's impossible.
- 0:23It is. Hardware has a massive physical
- 0:26overhead and software has infinite scalability.
- 0:29Never the twain shall meet. Usually, yeah.
- 0:32But today, we're looking at a company that hasn't just crossed
- 0:34that barrier, they've entirely shattered it.
- 0:37Welcome to today's deep Dive. We are unpacking A staggering
- 0:41stack of financial research, 10K filings and discounted cash flow
- 0:45models, all centered on Nvidia's Q1 fiscal year 2027 earnings and
- 0:51their long term evaluation. It's honestly it's arguably the
- 0:54defining financial puzzle of the decade.
- 0:56It really is. The goal here is to shortcut all
- 0:58the market noise and figure out if this is it's just a classic
- 1:01bubble ready to pop or, you know, a modern value investing
- 1:04masterpiece. Because we're looking at a
- 1:06scenario where the sheer velocity of capital formation is
- 1:09just testing the limits of traditional valuation model.
- 1:12Oh, absolutely. Now before we dive into the
- 1:14mechanics of those models, I need to share this quick message
- 1:17with you to access more advanced researches, applications and our
- 1:20latest published personal intelligent multi agent
- 1:23DeepValues intelligence. Welcome to explore and join Deep
- 1:26values dot AI community. Track total market temperature.
- 1:30Welcome to search and download iOS app Buffett indicator from
- 1:34Apple App Store. Disclaimer All analysis is based
- 1:37on publicly available sources and may be subject to revisions
- 1:40or differing interpretations. All contents from this channel
- 1:43is for information and research purposes only and does not
- 1:46constitute investment advice. Conduct your own due diligence
- 1:49before making investment decisions.
- 1:51If you like to access wider topics, including life and
- 1:54investing philosophy, please subscribe to our sister channel
- 1:57Deep Values. Don't forget to like this video.
- 1:59Much appreciated. OK, so where do we start with
- 2:02this? Well, to set the baseline for
- 2:04our analysis today, all the sources we are looking at cover
- 2:07financial data ending April 26th, 2026.
- 2:10Got it. Which is super recent.
- 2:12Yeah. And that's crucial because we
- 2:14aren't dealing with optimistic projections here or like 5 year
- 2:17forecast. We are looking at the hard data
- 2:21of what has already materialized in the first quarter of fiscal
- 2:24year 2027. Right, it's real money in the
- 2:26bank. Exactly.
- 2:28And it perfectly captures the absolute scale of the global
- 2:30transition to, you know, generative AI and accelerated
- 2:34computing. OK, let's unpack this because I
- 2:36want to anchor this with the top line number.
- 2:38Oh yeah, the Q1 FY20 27 revenue. Right, it hit 81.6 on $5
- 2:45billion. Which is just wild.
- 2:47It's an 85% year over year growth.
- 2:50Like these numbers look like a typo.
- 2:52It's like looking at a legacy hardware manufacturer but seeing
- 2:55the profit margins of an elite software company.
- 2:57It really is. And to put that magnitude in
- 3:00perspective, their data center revenue alone is nearing a $200
- 3:04billion annual run rate. 200 billion just for the data
- 3:08center. Just the data centers, that
- 3:09segment by itself is larger than the entire GDP of some mid sized
- 3:13countries and you know it represents A fundamental
- 3:16structural shift in their identity.
- 3:19We can see this detailed extensively in their Fiscal Year
- 3:21202610K summary. Right, the 10K breaks all of
- 3:25this down. Yeah.
- 3:26Out of their total annual revenue for FY20 26, which was
- 3:31nearly 216 billion, the data center segment made-up 89.7%.
- 3:36OK wait, so traditional gaming like the graphics cards that
- 3:39built their cultural brand in the 1st place, that's basically
- 3:42just a rounding error now? Pretty much it's down to just
- 3:457.4% of total revenue. That's crazy.
- 3:48Yeah, they've completely morphed into an enterprise AI
- 3:51infrastructure provider. But the true anomaly, the thing
- 3:55that is just breaking the historical precedent of the tech
- 3:57sector, is their gross margin. Oh, the Jap gross margin, right,
- 4:02sitting at roughly 75%, yes. 75%.
- 4:05Which brings me back to that hardware versus software paradox
- 4:10generating a 75% gross margin while scaling revenue by 85%
- 4:14year over year. While physically manufacturing
- 4:16hardware systems. Exactly.
- 4:18It's unheard of, and it generates so much cash that in a
- 4:22single quarter they recorded $50.3 billion in operating cash
- 4:28flow. In 1/4. 1/4 and they returned 20
- 4:31billion to shareholders through buybacks and dividends hikes,
- 4:35their dividend 25 fold and authorized another 80 billion
- 4:39share repurchase with no expiration date.
- 4:42Because the top line expansion is directly translating to the
- 4:46bottom line without getting, you know, eaten up by proportion
- 4:50cost, I mean gap net income for the quarter reached over 58
- 4:54billion. They are operating with a
- 4:56capital efficiency of a lean startup, but at the scale of a
- 5:00global hedge on. Here's where it gets really
- 5:02interesting, though. Those margins are incredible,
- 5:05obviously, but they immediately make me wonder how they survive.
- 5:08Competitor undercutting, Yeah. That's the classic question.
- 5:11Because in any normal market, if a hardware company has 75%
- 5:14margins, that is blood in the water, right?
- 5:16Competitors rush in, they drop their prices by 30%, steal the
- 5:20market share and the margins collapse.
- 5:22Which is exactly what economic theory says should happen.
- 5:24So why doesn't a competitor like AMD just slash their chip prices
- 5:28and take the business so they aren't just selling the pickaxes
- 5:31for the gold rush? They invented a specialized
- 5:34language that you have to speak if you want to mine the gold at
- 5:37all. Right.
- 5:37And what's fascinating here is that assumption most people make
- 5:41that code is just code fundamentally misunderstands the
- 5:45product, right? The true economic mode isn't
- 5:48actually the silicon itself, it's the software ecosystem
- 5:51built on top of. It the CDA platform.
- 5:54Exactly the Compute Unified Device Architecture, or CUDA.
- 5:58They launched CDA way back in 2000. 1006 is a proprietary
- 6:01parallel computing platform. Wow, 2006.
- 6:04So they've been building this for a while.
- 6:05Decades and over the last two decades, it has become the
- 6:08absolute default programming language for deep learning.
- 6:11Yeah, the sources mentioned that more than 98% of machine
- 6:14learning developers optimize their models directly for CEDA.
- 6:1798% it's a total monopoly. But wait, open source
- 6:21communities are relentless right?
- 6:23Hasn't anyone tried to build like a universal translator?
- 6:27If AMD offers a chip that's significantly keeper, won't
- 6:31developers just take the time to rewrite their code or use a
- 6:34translation layer? Well here's the thing.
- 6:36In AI training time is literally money.
- 6:39OK, Translating code on the fly or using imperfect workarounds
- 6:43introduces latency. When you are renting 10s of
- 6:47thousands of GPU's at hourly rates to train a trillion
- 6:50parameter model, a 10% drop in efficiency cost you millions of
- 6:54dollars in compute time. Wow, so you can't afford the
- 6:57translation delay? Not at all.
- 6:59Think of Kita like the Qordy keyboard.
- 7:01It might not theoretically be the only way to arrange letters,
- 7:04but the entire global workforce has deep muscle memory for it.
- 7:07That's. A great analogy.
- 7:08Asking a team of elite AI developers to suddenly switch to
- 7:11Andy's ROCM platform or Intel's One API is like asking a concert
- 7:16pianist to perform on an instrument where all the keys
- 7:18have been rearranged. The switching costs are
- 7:20structural and massive. That makes perfect sense.
- 7:22It's not just having a faster chip, it's that the global
- 7:25workforce of AI developers is already fluent in your
- 7:28proprietary language. Exactly, and they reinforce that
- 7:32software mode with a physical barrier to entry that
- 7:34competitors are really struggling to match.
- 7:36Which is what? System level architectural code
- 7:39design. NVIDIA doesn't really sell a
- 7:42standalone GPU anymore, they architect the entire
- 7:45supercomputing node. OK, let's breakdown the
- 7:47mechanics of that. The sources highlight their
- 7:49custom Grace CPU's and their proprietary NV link memory
- 7:53pools, right? Why does that system level
- 7:56design matter so much to the end user?
- 7:59Because AI models have grown so massive that they can't
- 8:02physically fit onto a single chip anymore.
- 8:05A model might be split across thousands of GPU's running
- 8:07simultaneously. So the bottleneck in training an
- 8:10AI isn't usually the processing speed of a single chip, it's how
- 8:13fast those thousands of chips can talk to each other to share
- 8:17data. Because if they can't
- 8:18communicate quickly, the whole system slows down.
- 8:21Yeah, exactly. Standard networking cables
- 8:23create a massive traffic jam. NV Link is essentially A
- 8:28proprietary multi lane superhighway built directly into
- 8:32the server architecture that just removes that traffic jam
- 8:35entirely. And when you combine that
- 8:37throughput advantage with their dominant supply chain scale at
- 8:40TSMC for advanced packaging, you create a fortress.
- 8:44You aren't just trying to beat their chip, you have to beat
- 8:46their software, their networking, and their
- 8:48manufacturing scale simultaneously.
- 8:50It's an incredible setup, but looking at the SEC filings
- 8:53though, there is a glaring vulnerability high in the fine
- 8:57print. Oh, absolutely.
- 8:58Wait, the sources show massive customer concentration?
- 9:01Building an empire on essentially 3 VIP's sounds
- 9:04terrifying. What happens if they just stop
- 9:07buying? This raises an important
- 9:09question because it's a huge risk factor.
- 9:11If you look at the Q2 FY20 six data, they mentioned customer A
- 9:14and Customer B driving 39% of total revenue. 39% from just two
- 9:19customers. Right.
- 9:19And when you isolate the data center segment, nearly 53% of
- 9:23the revenue came from just three unnamed customers.
- 9:26Unnamed. I mean, we all know they're
- 9:27talking about Microsoft, Meta and Google, right?
- 9:29Yeah, it's the worst kept secret on Wall Street.
- 9:32The customer base has migrated from millions of, you know,
- 9:35highly distributed retail gamers buying individual graphics
- 9:39cards, right? The old gaming market.
- 9:40To a highly concentrated pool of hyper scalers and cloud service
- 9:44providers building these massive AI factories.
- 9:48And this concentration is the exact issue keeping risk
- 9:51managers awake at night. For good reason.
- 9:54It highlights the primary long term threat, which is the CapEx
- 9:58trajectory of these hyperscalers.
- 10:00And these hyperscalers? A magnificent 7.
- 10:03They are not just taking Nvidia's premium prices lying
- 10:06down, the research points out. They're actively building their
- 10:09own bespoke silicon to cut NVIDIA out of the loop.
- 10:12Yeah, it's this weird frenemy dynamic at the heart of the
- 10:15industry. Like Microsoft is buying
- 10:17hundreds of thousands of GPU's will secretly design its own
- 10:20custom chip called Maya. Right, and Google has its Tpus
- 10:23and Amazon has Tranium. Exactly.
- 10:25These are custom application specific integrated circuits or
- 10:28ASICS. So wait, if Google and Amazon
- 10:31have the capital and the engineering talent to build
- 10:34custom ASICS, why don't they just dump NVIDIA entirely and
- 10:38run everything in house? Well, building a custom ASIC for
- 10:41a specific internal workload is very different than building a
- 10:45general purpose GPU that the whole world wants to code on.
- 10:48OK, Google can use Tpus for its own internal search algorithms
- 10:51perfectly well. But Google Cloud also rents out
- 10:55server space to external startups.
- 10:57And those external startups like we discussed, they demand CUD.
- 11:00Exactly. They demand NVIDIA.
- 11:02So the hyper scalers are essentially trapped by their own
- 11:05customers preferences. They have to keep buying NVIDIA
- 11:08hardware to remain competitive as cloud service providers, even
- 11:11as they try to incubate their own cheaper alternatives in the
- 11:15background. It's a fascinating standoff
- 11:17because if those in house custom chips ever capture a meaningful
- 11:20share of general cloud inference workloads, or if those giants
- 11:24simply decide they've built enough infrastructure and slow
- 11:27down their CapEx spending. The revenue could contract
- 11:29incredibly sharply. Exactly.
- 11:31There is also a geopolitical layer compounding that risk.
- 11:35Oh, right. Yeah.
- 11:36The forward-looking guidance for Q2 FY20 27 conservatively
- 11:41assumes 0 data center acute revenue from China.
- 11:44Zero. Wow.
- 11:45They are proactively de risking their pipeline from shifting
- 11:50trade constraints. And I mean, while that is a
- 11:53prudent management decision, it effectively walls off a massive
- 11:56piece of the global market. Which brings us to the ultimate
- 11:59puzzle of this deep dive and where the traditional finance
- 12:02world is completely clashing with the tech world.
- 12:05Well. This is the best part.
- 12:06How do discipline value investors?
- 12:08Look at a business printing unprecedented free cash flow,
- 12:11protected by a software monopoly but deeply reliant on the
- 12:15ongoing CapEx of three or four tech giants.
- 12:18Do the old valuation rules still work?
- 12:20Well, if you run this company through the classic Benjamin
- 12:23Graham deep value tests, it fails spectacularly.
- 12:27I mean, deep value investors typically look for APE ratio
- 12:29under 15, right? The trailing PE ratio here is
- 12:3244.8. That is miles outside the
- 12:34comfort zone for a classic value play.
- 12:36Completely outside of it, and the price to book ratio is even
- 12:40more extreme. A traditional value target is AB
- 12:44under 1.5. This sits at a massive 34.5.
- 12:4834.5 That's just astronomical. Right, because the market is
- 12:51pricing in the proprietary software platforms the developer
- 12:55lock in and the intellectual property rather than the
- 12:58tangible book value of their factories and inventory makes
- 13:01sense. But because these metrics are
- 13:03stretched so tight, we're actually seeing legendary value
- 13:06investors take evasive action. Yeah.
- 13:08The research indicates that Stanley Druckenmiller entirely
- 13:11exited his long term position, right And David Tepper disposed
- 13:15of the vast majority of his holdings.
- 13:17Terry Smith exited completely specifically setting concerns
- 13:20about that customer concentration we just discussed.
- 13:23Yeah, they're seeing the risk in pulling back.
- 13:24The source also explicitly notes that Michael Bury actively
- 13:28shorted the stock in late 2025 and early 2026.
- 13:32He isn't just avoiding it, he's betting against it.
- 13:34He is anticipating A structural valuation multiple compression.
- 13:39These investors are looking at the massive sector wide run up
- 13:43and seeing a total violation of the margin of safety.
- 13:46So what does this all mean? Are the billionaires right to
- 13:49run away, or are they trying to navigate a brand new world using
- 13:53a very old map? Yeah.
- 13:54Well, if we connect this to the bigger picture, there are other
- 13:57ways to measure a company's trajectory, and that's where the
- 14:00deep value paradox emerges. OK, tell me about the paradox.
- 14:04While it fails Benjamin Graham's tests, it actually passes Peter
- 14:08Lynch's tests with line colors. Really.
- 14:11How so? Lynch focuses on the PEG ratio,
- 14:14which adjusts the PE multiple for the company's actual growth
- 14:17rate. A PEG under 1.0 is generally
- 14:20considered attractive and. What's Nvidia's?
- 14:22Despite the massive run up in stock price, the PEG ratio
- 14:25remains incredibly low at 0.68. Wow, 0.68 that's surprisingly
- 14:30low. Right, this suggests the market
- 14:32might actually be under pricing the sheer velocity of its long
- 14:35term compounding. Add to that a return on retained
- 14:37earnings of 84.1% over the past decade. 84%.
- 14:42For every dollar management decides to keep rather than give
- 14:45back to shareholders, they're generating an 84 cent return.
- 14:49That is phenomenal capital allocation.
- 14:52And so let's get into the mechanics of the discounted cash
- 14:54flow model, the DCF, because this is where the rubber meets
- 14:57the road on valuation. The sources provide a detailed
- 15:01base case. How do the assumptions actually
- 15:03hold up under scrutiny? The base case DCF is incredibly
- 15:08revealing about what the market expects.
- 15:10It starts with an assumption of $96.6 billion in base free cash
- 15:15flow, which aligns with their audited FY20 26 numbers.
- 15:19All right, almost 100 billion in free cash flow.
- 15:21Then it models a 10 year projection period fading down
- 15:24from 25% growth to 12% growth, a 3% terminal growth rate, and
- 15:29applies a 10% discount rate. Under those specific conditions,
- 15:34the intrinsic value comes out to $193.09 per share.
- 15:37OK, let's unpack those assumptions.
- 15:39A10 year fade from 25% down to 12% growth.
- 15:44Is that realistic in a world where Microsoft and Google are
- 15:47actively trying to bypass them? It's debatable for sure.
- 15:50Right. And a 10% discount read implies
- 15:53a certain level of safety. If you're a risk manager looking
- 15:57at a company relying on essentially 4 customers for its
- 16:00survival, you might demand a 15% discount rate to account for
- 16:04that concentration risk. Exactly.
- 16:06And if you plug a 15% discount rate into that model, the
- 16:09intrinsic value plummets. That is the exact tension in the
- 16:12model. The 10% discount rate assumes
- 16:15their Moat is impenetrable based on that 193.09 intrinsic value
- 16:22and with the stock trading in this data set around $223.47.
- 16:27Right, it's trading higher than the intrinsic value.
- 16:29Yeah, it currently lacks a traditional 30% margin of
- 16:32safety. To achieve a textbook margin of
- 16:35safety, a value investor would need a buy price closer to 135
- 16:39dollars. Wow so at 2:23 you are
- 16:42definitely paying a premium. You are.
- 16:45However, the analysis concludes that despite the premium, it
- 16:48acts as a Buffett style compounder.
- 16:51It's an extraordinary business trading at a fair, if somewhat
- 16:54stretched, price. But there has to be an
- 16:56underlying condition for that math to work out over the next
- 16:59decade. Oh, absolutely.
- 17:00The underlying condition is macroeconomic.
- 17:03That valuation only holds up provided the global AI factory
- 17:07build out scales to a $1 trillion addressable market by
- 17:1020-30. 1 trillion by 2030. It assumes every major
- 17:14enterprise, government and cloud provider continues to build out
- 17:18infrastructure at an exponential rate and that the company
- 17:21Florida State executes its future architecture.
- 17:23Rd. Maps transitioning smoothly to
- 17:25their next generation platforms without any manufacturing
- 17:28hiccups. That is a lot of things that
- 17:30need to go perfectly right. It is a price to perfection
- 17:32scenario. So to summarize the journey
- 17:35we've been on today, we are looking at a company that has
- 17:37successfully transitioned from selling hardware to a fragmented
- 17:41consumer market right into a full stack software moded AI
- 17:46juggernaut. Perfectly said.
- 17:47They are generating unprecedented free cash flow,
- 17:50over 96 billion in a single year and operating with software like
- 17:5575% margins. It's historic.
- 17:57But the trade off for that growth is severe customer
- 18:00concentration. If you're an investor, you
- 18:02aren't just betting on their engineering.
- 18:04You're betting on the sustained, uninterrupted capital spending
- 18:07of a handful of massive tech giants who are quietly trying to
- 18:12build alternatives. That captures the dynamic
- 18:15perfectly, and you know there is one final lingering variable I'd
- 18:19like to leave you with. Oh, it's.
- 18:20That we spent a lot of time discussing the QWERTY keyboard
- 18:23analogy, right? How Qtuda is the absolute
- 18:26standard for developers today and how that muscle memory
- 18:30protects their margins. Right, it's the language you are
- 18:32forced to speak if you want to operate in this space.
- 18:35But as these generative AI models become exponentially more
- 18:39Cape, fueled by the very clusters of GPUs we've been
- 18:42discussing, their coding abilities are skyrocketing.
- 18:45Oh man, I see where you're going with this.
- 18:47What happens when the AI itself gets so smart that human
- 18:50developers stop writing the code entirely?
- 18:52What happens when a generative AI can ingest a prompt and
- 18:56instantly output perfectly optimized machine code for any
- 19:00hardware architecture, like instantly translating between
- 19:04NVIDIA, AMD or custom Google Silicon without any latency?
- 19:08If an AI can act as a flawless instant universal translator,
- 19:12the swishing costs literally disappear.
- 19:15Exactly. The provocative question is,
- 19:17could the AI eventually bypass human written legacy platforms
- 19:21like CUD entirely if the machines learn to navigate the
- 19:25hardware on their own? Does the software mode evaporate
- 19:28overnight? That is Wow.
- 19:29If the machines build their own map, the old map doesn't just
- 19:32shift, it becomes entirely obsolete.
- 19:34Exactly. That is a brilliant and slightly
- 19:36terrifying thought to chew on. Well, thank you so much for
- 19:39joining us on this deep dive. Keep questioning the assumptions
- 19:42in the models, keep looking beneath the surface of the
- 19:44headlines. And remember, even when the
- 19:45historical rules seem completely broken, the underlying mechanics
- 19:49always have a story to tell. We'll catch you on the next one.