Latest / Deep Values: Value Investing Deep Dives / Why Nvidia‘s AI factories terrify Wall Street
Transcript
- 0:00To access more advanced researchers, applications and
- 0:03our latest published personal intelligent multi agent deep
- 0:06value intelligence Welcome to subscribe on deep values dot AI
- 0:11to track total market temperature.
- 0:13Welcome to search and download the iOS app Buffett Indicator
- 0:16from the Apple App Store. Disclaimer All analysis is based
- 0:20on publicly available sources and may be subject to revisions
- 0:23or differing interpretations. All contents from this channel
- 0:26is for educational and research purposes only and does not
- 0:29constitute investment advice. Conduct your own due diligence
- 0:32before making investment decisions.
- 0:34If you like to access wider topics, including life and
- 0:37investing philosophy, please subscribe to our sister channel
- 0:40at Deep Values. Don't forget to like this video.
- 0:42Much appreciated. OK, so I want you to imagine
- 0:46being like the smartest kid in your entire school.
- 0:48You take the hardest test of the year, you score a 100%.
- 0:51You even get all the extra credit, right?
- 0:53And you somehow organize the teacher's lesson plans while
- 0:56you're at it. You're definitely expecting a
- 0:57parade for that. Right.
- 0:59You are naturally expecting a parade, but instead the
- 1:02principal calls you into the office, looks you dead in the
- 1:04eye, and says you did great, but we're pretty sure you're going
- 1:08to fail next semester and then they just suspend you.
- 1:11It's, I mean, it sounds completely absurd when you frame
- 1:14it in human terms like that, but honestly, that is the exact
- 1:17paradox playing out at the absolute highest levels of the
- 1:20tech industry. Right now, yeah.
- 1:22And that is exactly why we are doing a deep dive into NVIDIA
- 1:26today because they just posted the kind of financial report
- 1:31card that companies literally dream about.
- 1:33And the market's reaction was, well, was absolutely brutal.
- 1:36Brutal. Is an understatement.
- 1:37Yeah, so. To figure out what is actually
- 1:38going on under the surface, we're pulling from 2 incredible
- 1:41sources. Today we have an exclusive, very
- 1:44candid interview with Nvidia's CEO Jensen Wong on the All In
- 1:48podcast and repairing that with a really rigorous financial and
- 1:51strategic analysis from the channel Old Stone Talks Chips.
- 1:56Both of which are essential if you want to understand the full
- 1:58picture here. Exactly.
- 1:59OK, let's untack this. We are going to figure out
- 2:02exactly what an AI factory is, why you might soon have a
- 2:06literal robot coworker, and the hidden physical vulnerabilities
- 2:12that are just terrifying Wall Street right now.
- 2:14So let's start with that staggering surface level report
- 2:17card. In their latest quarter, NVIDIA
- 2:19posted a 73.2% revenue growth. Which is astronomical for a
- 2:24company of that size. Right.
- 2:25And it gets crazier because their gross margin was 75%.
- 2:28To put that in perspective for you, Apples gross margin is
- 2:31usually under 50% and Tesla's is hovering somewhere around 18%.
- 2:36Nvidia's basically operating a legal money printer.
- 2:39Yeah, they are, which makes the subsequent market punishment so
- 2:42incredibly jarring. Following those flawless
- 2:45numbers, the market wiped out $260 billion of Nvidia's value
- 2:49in a single day. Wow.
- 2:51Yeah, I mean, that is the equivalent of entirely erasing 5
- 2:53Intel corporations off the map just overnight.
- 2:56Gone. OK, so if the engine's running
- 2:58that perfectly, why did everyone panic and hit the brakes?
- 3:01Well, what's fascinating here is that the panic stems from a
- 3:04really fundamental misunderstanding of what NVIDIA
- 3:07actually is. Right now.
- 3:09The broader market is still, you know, valuing them as a
- 3:12traditional hardware company that just sells computer chips,
- 3:15GPU's. Or your graphics.
- 3:16Cards. Exactly.
- 3:18But Jensen Hong is declaring very loudly that they are no
- 3:22longer just selling chips, they are selling entire AI factories.
- 3:26That pivot is, I mean, it's massive.
- 3:30It completely rewires their entire business model.
- 3:33Let's think about it like the auto industry, right?
- 3:35For years and video was just building the engine, the GPU.
- 3:38Yeah. And they would sell that engine
- 3:39to someone else like Dell or HP who actually went and built the
- 3:43car around it. But now with their new Ruben
- 3:45platform, NVIDIA is just refusing to sell you the engine.
- 3:48They want to sell you the entire car, the racetrack and the pit
- 3:51crew. That's a perfect analogy.
- 3:53The Ruben platform is the absolute embodiment of that
- 3:56shift. It isn't just a single silicon
- 3:57chip you plug into a motherboard anymore.
- 3:59It is a massively bundled package.
- 4:01Well, it's actually in it. So it includes their next
- 4:03generation GPU's obviously, but also CPU's that utilize their
- 4:07custom Olympus cores. And on top of that, it includes
- 4:10their proprietary networking switches, their own specialized
- 4:14liquid cooling systems, and their entire software stack to
- 4:18make it all talk to each other seamlessly.
- 4:20So they are completely squeezing out the middleman.
- 4:22Oh, entirely. The traditional server builders
- 4:25suddenly find themselves basically relegated to just
- 4:28making the metal outer shell while NVIDIA controls,
- 4:31optimizes, and, you know, monetizes every single piece of
- 4:35the high value interior. Right.
- 4:36And compressing that entire supply chain into one integrated
- 4:39product allows them to make a really specific promise to their
- 4:42enterprise customers. They're essentially saying,
- 4:45look, if you buy our $50 billion factory, the sheer optimize
- 4:48integration of all those parts will ultimately generate the
- 4:51lowest cost output in the industry, OK?
- 4:53Which brings us to the output itself.
- 4:55If you build a massive multi billion dollar AI factory, what
- 5:00exactly is rolling off the assembly line?
- 5:02Tokens, digital tokens, and understanding this is honestly
- 5:06the key to understanding the current shift in the entire tech
- 5:09landscape. We are moving out of the era of
- 5:12merely training AI models and into the era of inference.
- 5:16Let's make sure we define the mechanics of that clearly,
- 5:18because both of our sources highlight this as a really
- 5:21crucial physical distinction. Training in AI is like sending a
- 5:26brain to college, right? Yeah, exactly.
- 5:27You're feeding it massive amounts of data, basically the
- 5:30entire Internet, so it learns how to identify patterns and
- 5:32think, and that requires huge centralized brute force
- 5:36computing power. But inference is entirely
- 5:39different. Right inference is putting that
- 5:41educated brain to work at a desk.
- 5:43It's the AI actually generating tokens in real time to solve a
- 5:47specific problem, answer a question, or execute a task.
- 5:50And mathematically, inference is a highly complex computing
- 5:53problem that requires what Huang calls disaggregated computing.
- 5:57Explain the mechanism behind that for us.
- 5:59Why does inference require a different physical architecture
- 6:02than training? Well, when you train a model,
- 6:04you're blasting massive chunks of data through the system in
- 6:07these predictable batches. But when you're doing inference,
- 6:11the requests are totally unpredictable, and they require
- 6:14really nimble memory recall. Like, you might ask the AI to
- 6:18summarize A50 page legal document document while
- 6:21simultaneously generating an image, right?
- 6:23A single giant brute force chip just isn't efficient.
- 6:26For that you need specialized chips, some handling logic, some
- 6:30handling vast pools of memory, and they all need to be talking
- 6:33to each other at lightspeed, sharing the workload
- 6:35dynamically. And the reason you need this
- 6:38highly responsive, disaggregated computing is because of what
- 6:41these tokens are actually doing. Now we are moving way past
- 6:45simple chat bots to just, you know, spit out a text recipe for
- 6:48banana bread. We're entering the agentic
- 6:50explosion, yes. The jump from chat bots to AI
- 6:53agents is profound. Systems like Open, Devon or
- 6:57Claude are essentially acting as personal computers with Agency.
- 7:00Wait, what? Do you mean by agency in this
- 7:02context? So a chat bot waits for your
- 7:04prompt, right? An agent actually has working
- 7:06memory and the ability to use tools.
- 7:09You can give an agent a high level goal and it can open up
- 7:13your web browser, navigate ACRM, pull data into Excel, write the
- 7:17necessary code to analyze it, and then schedule follow up
- 7:21tasks on your calendar. They're doing actual sequential
- 7:24work. Exactly.
- 7:25And to put the value of that work into perspective, Huang
- 7:27offered a really striking thought experiment during the
- 7:30All In interview. He said, Imagine a company pays
- 7:34a star engineer $500,000 a year. If you look at the books at the
- 7:38end of the year and that engineer hasn't consumed at
- 7:40least $250,000 worth of AI compute tokens to boost their
- 7:44productivity, the company should be deeply alarmed.
- 7:46Which is a wild statement. Here's where it gets really
- 7:49interesting though, because I have to push back on that 1/4 of
- 7:52$1,000,000 a year on compute for one employee.
- 7:54I mean, it sounds like a massive hype cycle.
- 7:57Are enterprise companies actually seeing a real return on
- 7:59investment for those tokens, or are we just throwing wildly
- 8:02expensive silicon at the wall and hoping it makes people type
- 8:05faster? It is the billion dollar
- 8:07question that has Wall Street sweating right now.
- 8:10But Huang's argument requires completely reframing how you
- 8:14view human productivity at the highest levels.
- 8:17He actually compares it to LeBron James.
- 8:18Oh right, spending over $1,000,000 a year on his body.
- 8:21Exactly. LeBron uses hyperbaric chambers,
- 8:25private chefs, specialized trainers, all to stay in peak
- 8:29physical condition. Huang is saying if you have an
- 8:32elite knowledge worker, giving them a swarm of 100 AI agents to
- 8:36act as their personal tireless staff makes them superhuman.
- 8:40It entirely removes the bottlenecks of time and manual
- 8:43data entry. The All In interview actually
- 8:46highlighted a crazy real world example of this.
- 8:48They talked about replacing an entire enterprise software stack
- 8:51on a Sunday night in 90 minutes using a negentic system.
- 8:55Right, That is migration work that used to take teams of human
- 8:58engineers months of painstaking manual labor.
- 9:01So the ROI isn't just about typing faster, it's about sheer
- 9:04speed and capabilities that were previously impossible.
- 9:07And The thing is, those superhuman capabilities aren't
- 9:10just going to be confined to digital spreadsheets.
- 9:13The massive processing power NVIDIA is building is unlocking
- 9:17the physical world. This is where we jump from
- 9:19digital tokens to actual physical movement because
- 9:23according to the sources, NVIDIA essentially envisions 3 distinct
- 9:27computers operating in this new ecosystem.
- 9:30First is the training computer, which we talked about.
- 9:32Second is the simulation computer which they call
- 9:35Omniverse. Which is such a fascinating
- 9:37concept. It really is.
- 9:39It's this high fidelity virtual world where the software
- 9:42strictly obeys the actual laws of physics.
- 9:44Gravity, friction, material tension, everything behaves
- 9:47exactly like it does in reality. Yeah, think of Omniverse as a
- 9:50digital gym for AI. Before you put an artificial
- 9:53brain into a 2000 LB piece of metal moving through a crowded
- 9:56warehouse in the real world, you train it for millions of hours
- 10:00in the Omniverse because a crash in there costs absolutely
- 10:03nothing in terms of hardware or human safety.
- 10:05And that rigorous physical simulation leads directly to the
- 10:08third computer, which is the edge computer.
- 10:11This is the robotics layer. This is the brain inside self
- 10:14driving cars, automated warehouse forklifts, and
- 10:18eventually humanoid robots operating in our homes.
- 10:20And the implications for the physical world here are
- 10:23staggering. Huang predicts 2 massive
- 10:27breakthroughs on the edge. First, digital biology, meaning
- 10:31understanding genes and proteins.
- 10:34He says that's going to have its ChatGPT moment in the next three
- 10:36to five years, completely revolutionizing how we
- 10:39discovered drugs and treat disease.
- 10:42And second, he expects humanoid robots to be highly functional
- 10:45and integrated into our everyday lives within that exact same
- 10:48three to five year window. OK.
- 10:49Wait, let's 'cause right there, three to five years.
- 10:52If the robotics layer masters autonomy that quickly, and we
- 10:55eventually have a robot for every human on earth, what
- 10:58happens to the human work? It's a really critical societal
- 11:00concern. Huang addresses this by arguing
- 11:03that jobs won't necessarily vanish, but the nature of the
- 11:06labor will drastically transform.
- 11:08How so? His specific example is a human
- 11:11chauffeur, so the car might be driving itself entirely.
- 11:15But the chauffeur doesn't just disappear.
- 11:17They evolve into what he calls a mobility assistant.
- 11:20They manage your schedule, handle complex logistical
- 11:23routing, and coordinate your day from the car while the AI
- 11:26handles the steering wheel and the pedals.
- 11:28So the human shifts from manual operation to management and
- 11:31coordination. Exactly.
- 11:32But you know, this utopian vision of robot assistance and
- 11:37seamlessly transform jobs hits a massive brick wall when you look
- 11:41at the physical and financial reality today.
- 11:43That $260 billion stock drop we talked about earlier, that
- 11:46wasn't about a lack of imagination regarding the
- 11:49future. It was about 3 very physical,
- 11:51very immediate bottlenecks. Yeah, if we connect this to the
- 11:54bigger picture, the old Stone Talks Chips analysis does a
- 11:57phenomenal job of highlighting these 3 hidden ceilings that are
- 12:00absolutely terrifying Wall Street right now.
- 12:02The first fear is customer concentration.
- 12:04This is a classic existential business vulnerability.
- 12:09Right now, roughly 90% of Nvidia's data center revenue
- 12:13comes from just five big tech giants, Meta, Microsoft, Google,
- 12:18Amazon, and maybe one or two others.
- 12:21They are buying almost everything rolling off Nvidia's
- 12:24assembly lines. And the danger goes even deeper
- 12:26than just relying on a few buyers.
- 12:28Those five giants are currently spending huge portions of their
- 12:32own revenue, potentially over $600 billion collectively by
- 12:352026, just buying Nvidia's chips.
- 12:38That is an insane amount of capital.
- 12:40It is. They are essentially working
- 12:42around the clock just to pay Nvidia's margins and
- 12:44consequently they are desperately trying to build
- 12:46their own cheaper in house alternatives.
- 12:49Google is heavily investing in their TTU architecture.
- 12:51Amazon has Tranium. It's exactly like owning the
- 12:54most profitable exclusive restaurant in the entire world,
- 12:57but your dining room only has five VIP tables.
- 12:59If even one of those VI PS decides the steak is too
- 13:02expensive and they successfully figure out how to cook at home,
- 13:05meaning they build their own internal AI ships that are just
- 13:08good enough, your restaurant is suddenly facing a massive
- 13:10revenue Cliff. Exactly, and that fragility
- 13:13leads directly into fear #2 which is the TSMC bottleneck.
- 13:17NVIDIA designs the brilliant architecture of these chips, but
- 13:20TSMC Taiwan Semiconductor Manufacturing Company actually
- 13:24physically fabricates them. And NVIDIA currently has $95
- 13:28billion in non cancellable orders tied up with TSMC.
- 13:32Wait. $95 billion in unbreakable financial vows?
- 13:36Yeah. Why is NVIDIA so utterly
- 13:38dependent on this one single manufacturer?
- 13:40It comes down to a really specific physical mechanism
- 13:43called advanced packaging, specifically a technology TSMC
- 13:47perfected called Kowas or chip on wafer on substrate, right?
- 13:51Remember earlier when we discussed how inference requires
- 13:54disaggregated chips talking to each other instantly?
- 13:56Yeah, the logic and memory chips.
- 13:58Right. You can't just put those chips
- 13:59next to each other on a standard green motherboard.
- 14:01The physical distance is too great, the signal takes too
- 14:03long. TSMC has figured out how to
- 14:05microscopically stack the logic chips and the memory chips in
- 14:08three dimensions incredibly close together so data moves at
- 14:11lightspeed. Almost nobody else on Earth can
- 14:13perform this specific manufacturing miracle at the
- 14:16scale NVIDIA needs. It's an incredible competitive
- 14:19Moat because nobody else can get the manufacturing capacity, but
- 14:23it is also a terrifying single point of failure.
- 14:26I mean, if there's a major earthquake in Taiwan, a supply
- 14:29chain shock, or any disruption at TSMC, Nvidia's entire
- 14:34pipeline freezes instantly. They cannot just take their
- 14:37blueprints to another factory down the street.
- 14:39They're completely locked in, Which brings us to fear #3 and
- 14:43from a purely infrastructural standpoint, this is the ultimate
- 14:47reality check for the AI broom the power grid.
- 14:51Yeah, according to the analysis, there are currently $80 billion
- 14:54worth of AI infrastructure orders just sitting around,
- 14:56unable to be deployed. Not because they like the
- 14:59silicon, and not because of TSMC, but because there simply
- 15:02isn't enough electricity available on the grid to plug
- 15:04them in. And to understand why this is
- 15:06happening, you really have to look at how AI factories draw
- 15:08power compared to old data centers.
- 15:10Traditional cloud servers running websites or e-mail.
- 15:13They're bursty. Yeah, they draw power when
- 15:15someone loads a page and then they idle.
- 15:18AI inference factories run at maximum utilization constantly.
- 15:22These new GPU racks are so dense and run so hot that traditional
- 15:26air conditioning fails entirely. You literally have to pump
- 15:29liquid directly to the silicon just to stop the metal from
- 15:32melting. That is wild.
- 15:34The electrical grid was built to support neighborhoods and
- 15:36traditional industry. It was not built for a single
- 15:39warehouse pulling the same continuous gigawatts as a medium
- 15:42sized city. It is crazy that the apex of
- 15:45human computing achievement is being bottlenecked by utility
- 15:48poles and copper wires that haven't fundamentally changed in
- 15:51a century. The life cycle of a new AI chip
- 15:54is measured in like 18 months. The life cycle of upgrading a
- 15:58national power grid is measured in decades.
- 16:01Physics and infrastructure are incredibly stubborn limits, but
- 16:04compounding all of these physical bottlenecks is the
- 16:06fracturing of the global landscape.
- 16:08The supply chain isn't just limited by the laws of
- 16:10thermodynamics anymore, it is actively being limited by
- 16:13politics. Right now, the Old Stone
- 16:16analysis takes a hard look at the geopolitical side of this,
- 16:20specifically the contrasting tech policies of the Biden
- 16:23administration and the Trump administration.
- 16:26And just to be super clear with you listening, we are simply
- 16:29unpacking their strategic analysis here from the source
- 16:32material. We are absolutely not taking a
- 16:34political stance or endorsing any viewpoints.
- 16:37We're remaining completely impartial.
- 16:39But the core argument the source makes is about the tension
- 16:42between global dominance and isolation.
- 16:45The All In interview highlights a desire from Trump for the US
- 16:48tech industry to entirely dominate the global landscape.
- 16:52But the immediate reality on the ground is that due to current
- 16:54export restrictions from the Biden administration, NVIDIA
- 16:57basically has a 0% market share in China, which is the second
- 17:00largest market in the world. And when you artificially create
- 17:03a vacuum in a massive market through blockades, someone is
- 17:06going to fill it. The tech restrictions are
- 17:09effectively forcing the creation of a parallel, completely
- 17:12independent tech ecosystem. China isn't just sitting still
- 17:15without NVIDIA chips. The sources point out that
- 17:18Chinese robotics and AI hardware are advancing rapidly on their
- 17:22own. And the hidden leverage there is
- 17:24that they control the underlying physical supply chains that the
- 17:27rest of the world relies on. You can't build advanced robots
- 17:30without rare earth minerals, high efficiency motors, and
- 17:33specialized magnets. These are all areas where China
- 17:36holds massive dominance. Yeah, and this geopolitical
- 17:39anxiety isn't limited to just the US and China, either.
- 17:42The fear of being cut off from compute power has sparked a new
- 17:45global trend that Huang calls sovereign AI.
- 17:48Countries are looking at these blockades and realizing that
- 17:50relying on a handful of foreign tech giants is a massive
- 17:53national security risk. So what does this all mean?
- 17:56Do blockades actually work, or do they just accelerate the
- 17:59creation of formidable independent rivals?
- 18:02Because India is spending billions of dollars right now to
- 18:05build their own national AI infrastructure.
- 18:08Japan is doing the exact same thing.
- 18:10Right, because they realize that AI isn't just a fun app on your
- 18:13phone, it is the fundamental future infrastructure of their
- 18:16healthcare systems, their defense grids, and their
- 18:18education. They just refuse to let their
- 18:21citizens data and their economic engine be held hostage on
- 18:25foreign servers. The ultimate ceiling for the
- 18:27entire AI era might not be the limits of silicon architecture
- 18:31or even the power grid. It might be the fragmented,
- 18:33siloed world where global cooperation breaks down into
- 18:36isolated islands of technology fighting over raw resources.
- 18:40We are trying to build this unified, hyper intelligent
- 18:43future on top of a highly fractured analog foundation.
- 18:47It is a phenomenal amount to process.
- 18:49It really is. So let's do a quick recap of the
- 18:52journey we just took. We started with a paradox of
- 18:55Nvidia's flawless earnings report resulting in a massive
- 18:59$260 billion stock drop. We explored the mechanics of
- 19:03their radical transformation from simply making GPU's to
- 19:07building fully integrated AI factories.
- 19:09We looked at the shift from training to inference and that
- 19:12explosion of AI agents, your new digital Co workers doing real
- 19:17measurable work requiring disaggregated computing.
- 19:20And we examine the massive hurdles threatening all this
- 19:23growth, the severe lack of electricity, the terrifying
- 19:27reliance on advanced packaging at one manufacturing plant, and
- 19:30the deeply fragmented geopolitical landscape driving
- 19:33sovereign AI. Navigating those physical and
- 19:36political bottlenecks is undoubtedly the defining
- 19:38challenge for the tech industry over the next decade.
- 19:40But bringing it back to you, the listener, and your daily life,
- 19:43Jensen Wong gave a piece of advice that really cuts through
- 19:46the noise. He said you shouldn't worry
- 19:48about losing your job to an AI. You should worry about losing
- 19:50your job to someone who is using an AI.
- 19:52The mandate for all of us isn't to figure out how to compete
- 19:55with the machines token output, it's to become an expert at
- 19:59managing these AI agents. To be the creative director
- 20:03coordinating the Symphony of tokens.
- 20:05The human element shifts entirely from manual execution
- 20:08and data processing to high level strategy and direction.
- 20:12Well said. I want to leave you with one
- 20:14final, slightly provocative thought to Mull over.
- 20:17Building on everything we just discussed, we spend so much time
- 20:21obsessing over the intelligence of the software models, the
- 20:23microscopic stacking of the silicon, and the agentic
- 20:26workflows. But think back to that brilliant
- 20:29student who failed because of the system around them.
- 20:32If the ultimate unbreakable bottleneck for this entire
- 20:34trillion dollar technological revolution is actually the 100
- 20:38year old electrical grid, could the next great global tech
- 20:41monopoly actually end up being a traditional, seemingly boring
- 20:44utility company that finally figures out how to power the AI
- 20:48era? It really forces you to
- 20:49reconsider where the true power lies in the modern world.
- 20:52Literally and figuratively. That is our deep dive for today.
- 20:56Keep questioning the bottlenecks, yeah.