Latest / Deep Values: Value Investing Deep Dives / Elon Musk: China Could Crash AI Memory Margins
Transcript
- 0:00So what happens when the most wildly profitable AI driven tech
- 0:05sector suddenly collides head on with the world's most relentless
- 0:09manufacturing machine? I mean, historically speaking,
- 0:12it's the irresistible force meeting the immovable object.
- 0:16And the manufacturing machine usually leaves a pretty
- 0:19permanent, very disruptive mark on the landscape.
- 0:22Yeah. A highly disruptive mark and
- 0:25Elon Musk just fired this major warning shot about the memory
- 0:28chips that are basically powering our entire AI
- 0:31revolution. Right.
- 0:33And if his read on the situation is accurate, the massive profit
- 0:36margins that these tech giants have been minting over the last
- 0:39year, well, they could evaporate way faster than anyone
- 0:43anticipates. Exactly.
- 0:44So welcome to today's Deep Dive. We have a really fascinating
- 0:48stack of sources to go through today.
- 0:49We really do. We're looking at a specific set
- 0:52of posts from Elon Musk on X from October 12th, 2026 + A
- 0:57Barron's financial report from October 13th.
- 1:00Right. And some direct company
- 1:01announcements coming out of China's Changshan Memory
- 1:04Technologies or CXMT for short. So our mission for this deep
- 1:07dive is to really unpack this warning Musk issued.
- 1:11We're trying to figure out if this current AI driven memory
- 1:14chip boom is built on a solid foundation or if it's, you know,
- 1:18about to get transformed into a ruthless low margin commodity
- 1:21business overnight. It's a massive question,
- 1:23especially if you're actively tracking the markets and trying
- 1:26to separate the hype from the underlying financial reality.
- 1:29And, you know, if you're a listener looking to navigate
- 1:31these exact kinds of wild market swings, today's deep dive is
- 1:34brought to you by a suite of apps designed for the modern,
- 1:37intelligent investor. Yeah, they're phenomenal tools
- 1:40for cutting through the daily noise.
- 1:42Like if you want to calculate the Buffett style intrinsic
- 1:45value of any stock, you can just use the Deep Values app on iOS.
- 1:49Right. Or, to gauge if the whole market
- 1:51is getting dangerously overpriced, you can check out
- 1:53the Buffett Indicator app on iOS or the Bubble Watch app on
- 1:57Android. And for the full research
- 1:59platform plus access to Deep Value Intelligence, which is
- 2:03their multi agent AI analyst, just head over to Deep Values
- 2:06dot AI. You can also find longer videos
- 2:09exploring investing in life philosophy at their YouTube
- 2:11channel at The Deep Values. Definitely like and subscribe
- 2:14for one investing insight a day. Absolutely.
- 2:17And a quick reminder, as always, this deep dive is strictly for
- 2:20educational purposes, not investment advice.
- 2:23Always do your own due diligence before making investment
- 2:26decisions. Always.
- 2:27So let's get straight into this. Before we go down into the
- 2:30microscopic nanometer level engineering, we really need to
- 2:34understand the macroeconomic alarm bell that Musk is ringing
- 2:37here. Right, because the context of
- 2:39what he said is crucial. He's looking at a historical
- 2:43pattern and basically applying it to the current AI
- 2:45infrastructure build out. Yeah.
- 2:47On October 12th, he made this observation on X that instantly
- 2:51caught the attention of the financial world.
- 2:53I'm just going to quote it directly.
- 2:55Go for it, he said. When China figures out how to
- 2:57manufacture a product, their ability to scale is often
- 3:00greater than the rest of the world combined.
- 3:03And he explicitly cited solar manufacturing and lithium
- 3:06batteries as the direct precedents for what he's seeing
- 3:09right now. Which makes perfect sense.
- 3:12I mean, think about the solar industry like 15 or 20 years
- 3:15ago. Back then, manufacturing a high
- 3:18efficiency solar panel was considered specialized boutique
- 3:22technology. Oh, completely.
- 3:23The market was dominated by American, German and Japanese
- 3:26firms. The margins were really decent.
- 3:28But then the Chinese manufacturing apparatus decided
- 3:32to prioritize solar, and they didn't just iterate, they built
- 3:36these massive subsidized mega factories focused purely on
- 3:40scale and driving down the cost curve.
- 3:42And it completely hollowed out the western solar manufacturing
- 3:45base. Like the cost per Watt plummeted
- 3:48by something like 90% / a decade.
- 3:51Exactly, and the same exact story played out with lithium
- 3:54ion batteries for electric vehicles.
- 3:56It goes from this premium high margin technological marvel to
- 4:00just a cheap mass produced commodity that you essentially
- 4:04buy by the ton. Which brings us to the
- 4:06second-half of Musk's quote. The real kicker here?
- 4:09He added. Maybe RAM will be different, but
- 4:11RAM was a low margin commodity only roughly 3 years ago, right?
- 4:15And RAM being random access memory, the chips that act as
- 4:19the short term working memory for our phones, our laptops, and
- 4:22crucially the mass of server farms training these AI models.
- 4:26So Musk is forcing us to question the longevity of the
- 4:29current profit structures. He isn't saying that companies
- 4:32like Micron, SK Hynix or Samsung are going to file for bankruptcy
- 4:36tomorrow. No, the AI demand is incredibly
- 4:38strong right now. Right.
- 4:39But he is asking of these current highly lucrative profit
- 4:42margins are an illusion. Like, is this just a temporary
- 4:45window of tight supply before the crushing weight of global
- 4:49manufacturing capacity brings those margins violently back
- 4:52down to earth? And the timing of his comment is
- 4:54what's really interesting. He posted this in mid-october
- 4:572026. What specifically is happening
- 5:00in China right now to trigger for this alarm?
- 5:02Well, that brings us to Changshin Memory Technologies,
- 5:05CXMT. They are the major domestic
- 5:08manufacturer for DRAM in China. DRAM stands for Dynamic Random
- 5:12Access Memory, just to clarify. Yep.
- 5:14And we have to look at 2 distinct announcements they just
- 5:16made to understand the sheer speed at which this is moving.
- 5:20Back on September 20th, CXMT announced that their fifth
- 5:24generation manufacturing platform, their G5 platform, had
- 5:28officially entered mass production.
- 5:30And this includes their new 24 Gigabit LP DDR5X products.
- 5:34Exactly. Which is the low power memory
- 5:36architecture that is just absolutely crucial for mobile
- 5:39devices, laptops, anywhere that battery efficiency is paramount.
- 5:43But the staggering part of that September announcement is the
- 5:46yield improvement. They claim this G5 platform gets
- 5:5050% more gross dyes per wafer than their previous generation.
- 5:54Wait 50% more. So they are extracting vastly
- 5:58more usable chips from the exact same slice of silicon.
- 6:00Yes, that development is very real and has been corroborated
- 6:04by Reuters. But the development that really
- 6:07sparked the intense financial reporting in October was an
- 6:10announcement made by CXM TS President Kao Kanyu.
- 6:13This is the architectural right. He stated that the company has
- 6:17made significant progress on a completely new memory cell
- 6:20architecture called Four F ^2. And their stated timeline is
- 6:24super aggressive. They plan to introduce server
- 6:27DDR5 Ardim products, so the heavy duty memory modules that
- 6:31go directly into data center servers using this new design by
- 6:35the end of 2026. Right, but we need to decode 4F
- 6:38squared because to anyone outside of semiconductor
- 6:41engineering it just sounds like a random algebra equation.
- 6:44Oh absolutely. So to understand why this is
- 6:46such a threat, you have to know that conventional DRAM layouts
- 6:50take up a physical area on the silicon of about 6F squared.
- 6:54The F just represents the minimum feature size they can
- 6:56print. So the best way to frame this is
- 6:59as a real estate play. Conventional DRAM is like
- 7:03building an apartment complex on a six acre lot.
- 7:06What CXMT is claiming with four F ^2 is that they have figured
- 7:09out how to build the exact same apartment building on a lot that
- 7:12is only four acres. So it's a 33.3% reduction in the
- 7:17physical footprint of the memory cell.
- 7:19Which instantly boosts your density by 50% across the entire
- 7:23silicon wafer. But wait, how do you just
- 7:26magically shrink the lot? If you aren't upgrading your
- 7:29lithography machines to print smaller lines, how do you just
- 7:32cram the same transistors and capacitors into a space that is
- 7:361/3 3rd smaller? You can't just arrange the
- 7:38furniture closer together. No, you don't arrange it closer
- 7:41together. You start building up and down.
- 7:43OK, so going vertical. Exactly, to go from six F ^2 to
- 7:46four F ^2, engineers have to move away from a traditional
- 7:49flat 2 dimensional layout. In a standard DRAM cell, the
- 7:53transistor and the capacitor are generally laid out horizontally
- 7:56on the silicon right. To achieve 4F squared, you have
- 7:59to force the architecture into the third dimension.
- 8:02You physically turn the transistor vertically, bearing
- 8:05the word lines deeper into the silicon, or stacking the
- 8:08capacitor directly on top in a much more complex 3D
- 8:12orientation. So you are fundamentally
- 8:15altering the geometry of the cell.
- 8:16It is an immense material science and engineering
- 8:19challenge. It's not just a matter of pretty
- 8:21smaller lines. It's about mastering entirely
- 8:24new ways to deposit and etch materials vertically without the
- 8:28electrical signals interfering with each other.
- 8:31And if CXMT can do this reliably, they bypass the need
- 8:34for the absolute most cutting edge export restricted
- 8:37lithography machines. They achieve higher density
- 8:40purely through architectural innovation.
- 8:42OK. I have to push back on the
- 8:43reality of this though, because an executive standing on a stage
- 8:47announcing progress on a vertical architecture is miles
- 8:51away from having a functional server DDR5 RDIMM slotted into a
- 8:54data center. Oh.
- 8:55Without a doubt. I mean the G5 mobile chips from
- 8:58September are corroborated, but 4F squared?
- 9:02What about the actual yields? It doesn't matter if you can
- 9:04build it in a lab if 80% of them fail quality control when you
- 9:07try to actually mass produce them.
- 9:09Your skepticism is entirely justified there.
- 9:12The commercial viability, the real world thermal performance
- 9:15and most importantly, the mass production yields of this 4F
- 9:18squared architecture have not been independently verified as
- 9:23of mid-october 2026. So having a plan to introduce
- 9:26these to the server market by late 2026 is just an ambition.
- 9:30Right now the proof is strictly in the production.
- 9:33Absolutely, but let's play out the worst case scenario for the
- 9:36established players anyway. Let's assume CXMT nails the
- 9:40engineering. They get the high yields, they
- 9:43achieve the 50% density boost and they start churning these
- 9:46out by the millions. OK, so if they do that, how does
- 9:49a technological win for a Chinese firm translate into a
- 9:51financial bloodbath for US and South Korean competitors?
- 9:54Well, it comes down to the brutal economic realities of a
- 9:57commodity cycle. Memory chips, standard server
- 10:00DRAM are fundamentally interchangeable.
- 10:03Right, a standard DDR5 module from Micron does the exact same
- 10:08job as one from Samsung or one from CXMT.
- 10:10Precisely. When supply is tight,
- 10:13manufacturers hold the pricing power.
- 10:14They can charge a premium, but the moment too much capacity
- 10:18comes online at the same time the market gets flooded, buyers
- 10:23know they can just go to the cheapest supplier.
- 10:25And the pricing power just vanishes instantly.
- 10:28Let's run the hypothetical math from the Barron's report on this
- 10:30because it completely reframes how I view these companies
- 10:33earnings. Yeah, walk us through the
- 10:35numbers. OK, so imagine we are in a tight
- 10:37market. A memory chip sells for $100.
- 10:41It cost the manufacturer $50.00 to make it OK.
- 10:43That leaves a gross profit of $50 per unit, a solid 50% gross
- 10:48margin. Everyone is thrilled the stock
- 10:50is flying high. Right, But then CXMT enters the
- 10:53market with massive scale oversupply hits.
- 10:57To stay competitive and actually move their inventory, are
- 11:00established manufacturers forced to drop their selling price by
- 11:0335%? So that chip now sells for $65
- 11:07on the open market. Exactly.
- 11:09But the fixed cost, the massive fabrication plants, the
- 11:11materials, the labor, none of that has changed.
- 11:14It still costs $50.00 to manufacture that chip.
- 11:17Wait, so if the price drops to $65 and it costs fifty to make,
- 11:21the profit is only $15? Yep, the price of the chip only
- 11:25dropped 35%, but that completely wiped out 70% of their actual
- 11:30profit. That is terrifying math.
- 11:33A relatively standard cyclical price adjustment basically
- 11:36destroys the bottom line. And this is exactly what Elon
- 11:39Musk is warning about. You can be shipping record
- 11:42volumes of memory chips, running your factories at maximum
- 11:45capacity, and still watch your earnings get absolutely crushed.
- 11:49And if you look at a giant like Micron, the Barron's report
- 11:51highlights that standard DRAM represents approximately 3/4 of
- 11:55their total revenue. If that standard memory market
- 11:58floods, they are so heavily exposed.
- 12:01They are. But is it really just impending
- 12:04doom? Because if I look at analysts
- 12:06from Morgan Stanley, they're looking at this exact same
- 12:08landscape and they are panicking.
- 12:10Right. They still see substantial
- 12:12memory supply constraints ahead. I mean if AI data centers just
- 12:15keep getting built at this brickneck speed, maybe CXMTS new
- 12:18chips just get absorbed by the market without crashing the
- 12:21price. That's a huge factor.
- 12:23Plus we are ignoring the absolute elephant in the room
- 12:26here, high bandwidth memory. Right HBM.
- 12:30HBM is the massive defensive wall protecting the established
- 12:33giants. Morgan Stanley is right to point
- 12:36out that advanced AI memory seems to be playing by
- 12:38completely different economic rules than standard DRAM.
- 12:42So I wanted to ask you about this.
- 12:44If standard DDR5D RAM is like building identical single story
- 12:49suburban houses which China is clearly gearing up to mass
- 12:52produce, is HBM like trying to build a highly engineered
- 12:56bespoke skyscraper? I'd love that analogy.
- 13:00It's spot on because to build that skyscraper you aren't just
- 13:04pouring a concrete foundation and repeating a floor plan.
- 13:07HBM is used alongside the most advanced AI accelerators.
- 13:11The chips from NVIDIA and AMD doing the heaviest lifting for
- 13:14artificial intelligence. So it involves taking multiple
- 13:18highly advanced memory dyes and physically stacking them on top
- 13:22of each other. Yes, and that means they have to
- 13:24communicate with each other vertically.
- 13:25It's not just about making the cells, it requires something
- 13:28called through silicon vias or TSVS.
- 13:32Through silicon vias. Right.
- 13:34You are essentially drilling microscopic elevator shafts
- 13:37entirely through the silicon wafers and filling them with
- 13:39copper so electrical signals can travel straight up and down the
- 13:42stack. Wow.
- 13:44You have to align and bond these layers with atomic precision.
- 13:48And the hardest part? Thermal management.
- 13:50Because they get so hot. Extremely hot.
- 13:52You have a stack of highly active chips generating a
- 13:55massive amount of heat, and that heat is just trapped in the
- 13:58middle of the skyscraper. Dissipating that heat requires
- 14:01incredible engineering for the more the testing is ruthless.
- 14:05If one single layer in an 8 high or 12 high stack fails, you have
- 14:10to throw the entire skyscraper away.
- 14:11And Musk actually specifically called this out in his post,
- 14:14right? He noted that memory stacked
- 14:16above logic with adequate cooling is considerably harder
- 14:19to manufacture. It is not something you can just
- 14:21reverse engineer and scale up in a few months.
- 14:23Not at all. The barriers to entry are
- 14:26astronomical, which gives Micron, Samsung and SK Hynix a
- 14:30fortress. This specific high margin
- 14:33segment of the market is severely constrained and highly
- 14:36defensible for the near future. It essentially acts as a
- 14:39counterweight to the threat in the standard DRAM market.
- 14:41Exactly. So if I'm a listener using the
- 14:44Deep Values app trying to cut through the noise of tech
- 14:47headlines, I need a framework to track this.
- 14:50I'm not just looking at the hype.
- 14:52It sounds like there are two distinct scenarios playing out.
- 14:55There are. You have the bull case and the
- 14:56bear case. The bull case is that AI demand
- 14:59is so immense that it simply absorbs whatever new supply CXMT
- 15:03brings online. The pricing for standard DRAM
- 15:06stays healthy and Micron retains its pricing power and margins
- 15:09because of the HBM demand. OK.
- 15:11And the bear case? The bear case is that CXMT
- 15:14achieves competitive yields on 4F squared, floods the market,
- 15:18and weakens conventional DRAM pricing that drags down the
- 15:21baseline across the industry. So to monitor which reality is
- 15:24taking shape, I need to look at 4 specific leading indicators.
- 15:28First. I need to see independent
- 15:30verification of CXM TS actual shipping yields for the four F
- 15:34^2 architecture, not just press releases from executives.
- 15:37Absolutely crucial. Second, I have to keep a hawk
- 15:40eye on global DRAM spot pricing. If the baseline price per bit
- 15:44starts to slide, that brutal margin crush we calculated is
- 15:48beginning. 3rd, you listen very closely to the gross margin
- 15:51guidance from companies like Micron on their quarterly
- 15:54earnings calls. Right, because if they start
- 15:56projecting lower future margins despite talking about record AI
- 15:59demand, the oversupply is already hurting them.
- 16:01Exactly. And 4th, I'm watching the
- 16:03revenue mix. I want to see how fast Micron
- 16:06can shift their business away from standard DRAM and into that
- 16:09defensible HBM skyscraper. The faster HBM grows as a
- 16:12percentage of their total revenue, the more insulated they
- 16:15are from the commodity cycle. And that is the ultimate race.
- 16:19Can the established players run fast enough of the extreme high
- 16:23end of the technology curve to escape the gravitational pull of
- 16:26Chinese manufacturing scale? It's intense.
- 16:30It is because you have to separate technological progress
- 16:33from economically competitive mass production, and you have to
- 16:37separate mass production from global oversupply.
- 16:40They are three distinct hurdles. It fundamentally changes how you
- 16:44look at the semiconductor industry.
- 16:46To wrap up our journey today, we started with Elon Musk's
- 16:49historical warning about the sheer overwhelming scale of
- 16:52China's manufacturing capabilities.
- 16:54Looking back at solar and lithium batteries, we explored
- 16:57CXMTS leap into vertical engineering with the four F ^2
- 17:00architecture to boost density by 50%.
- 17:03We worked through the terrifying math of margin compression,
- 17:06where a 35% price drop wipes out 70% of profits.
- 17:10And we examine the incredible engineering mode of high
- 17:12bandwidth memory with its microscopic elevator shafts and
- 17:15immense cooling challenges. But there is a lingering
- 17:18vulnerability in that Moat, isn't there?
- 17:20Oh, definitely. Because, as you noted earlier,
- 17:23HBM, as advanced as it is, still fundamentally relies on
- 17:27underlying standard DRAM technology.
- 17:29Exactly. The skyscraper is ultimately
- 17:31built out of standard memory bricks.
- 17:34Which leaves us with a final provocative thought to Mull over
- 17:37today. If Chinese manufacturers
- 17:39successfully crash the cost and massively increase the density
- 17:42of those foundational baseline DRAM cells, how long until those
- 17:47baseline improvements empower emerging competitors to breach
- 17:50the highly defensible HBM mode as well?
- 17:53If the foundation gets cheap enough, building the skyscraper
- 17:56starts to look a lot less intimidating.
- 17:58It really does. Something to keep at the
- 18:00forefront of your mind the next time you see headlines declaring
- 18:03the AI hardware boom to be unstoppable.
- 18:06Thank you so much for joining us for this deep dive today.
- 18:08And be sure to check out the sponsor apps and the YouTube
- 18:11channel to help track these structural market shifts.
- 18:14And above all, keep questioning the consensus.
- 18:16We'll see you next time.