Taken a glance at financial markets lately? You might think Korea’s semiconductor giants have unlocked an infinite money glitch.
Any remotely tech-savvy observer has already noticed that SK Hynix (I'll just stick with Hynix since that's what's printed on the chips I knew from back in the day) and Samsung have seen their valuations soar to dizzying heights, riding a wave of unprecedented demand for High-Bandwidth Memory (HBM). Hynix recently posted operating margins hovering around an absurd 76%—margins that look less like a hardware manufacturer and more like a software monopoly. Meanwhile, hyperscalers like Microsoft, Meta, and Google are dropping hundreds of billions on AI capital expenditure, booking out every scrap of global HBM capacity through 2027. Analysts are rightly skeptical, questioning if the traditional "boom-and-bust" memory death cycle is dead forever, replaced by a seemingly endless AI supercycle.
I don't buy it. The bubble will burst, and the catalyst isn't just standard oversupply. It’s the same old game of the political pressure cooker forcing China to innovate around Western blockades at a pace the market continues to severely underestimate.
1/ The Bottleneck: Why HBM is Solid Gold Right Now
To understand why Samsung and Hynix are printing money, you have to look at how LLMs consume data.
AI models don't read words; they process tokens—sub-word chunks, syllables, and symbols translated into numbers. When an AI generates a response, it calculates output tokens one by one, continuously cycling hundreds of billions of mathematical parameters through a processor. (Someone smarter than I can opine on whether this system of tokens will remain in the future, or might change)
The primary bottleneck in modern AI isn't the raw speed of the chip—it’s data movement, a hardware bottleneck known as the Memory Wall. Apparently motherboard memory is simply "too far away" from the GPU, causing data traffic jams.
Enter High-Bandwidth Memory (HBM). By vertically stacking DRAM chips like a microscopic skyscraper and wiring them directly to the processor via Through-Silicon Vias (TSVs), HBM achieves much faster data throughput (over 1,200 GBps per stack). Because LLMs are useless without this firehose of data, tech giants are willing to pay almost any price to secure Korean HBM.
Drawn with help from AI (and HBM!):
[ Traditional PC Memory ] [ High-Bandwidth Memory (HBM) ]
┌────────┐ ┌────────┐ ┌──────┐ ┌──────┐ ┌──────┐
│ CPU │◄────►│ RAM │ │ DRAM │ │ DRAM │ │ DRAM │ (Stacked Vertically)
└────────┘ Long, Slow Bus ├──────┴─┴──────┴─┴──────┤
│ Microscopic TSVs │
├────────────────────────┤
│ GPU / AI Accelerator │
└────────────────────────┘
2/ The Illusion of the Unbreakable Western Moat
Wall Street’s current stock valuations rely on a flawed assumption: Western sanctions have permanently neutralized China's chipmaking capabilities, or at least left it helpless for the long term (5+ years).
Under export controls that have spanned across multiple presidential administrations, the U.S. has blocked China from buying ASML’s Extreme Ultraviolet (EUV) lithography machines—the $380 million, double-decker-bus-sized systems required to print silicon below the 7nm threshold. Because Korea uses EUV to manufacture cutting-edge HBM4, analysts assume China is trapped in the technological Stone Age, leaving Samsung and SK Hynix with a structural monopoly.
This model is fatally flawed because it measures technological progress through a single lens: traditional transistor miniaturization (Moore's Law). Driven by survival, China isn't trying to beat ASML at its own game—it is systematically innovating around it. Even more deeply, I think there's some kind of structural blockage in the minds of most Western policymakers that simply assumes China can only copy, not innovate. It's such a simplistic view that one has to wonder who's living in the Stone Age.
3/ How China is Bursting the Bubble: The Asymmetric Playbook
My own gut says, China's response to Western embargoes is an asymmetric strategy that will undercut existing players' margins much faster than previous industry forecasts predict. This has occurred in dozens of other industries (see the deluge of news around Chinese EVs, for example), and there's no reason why it won't happen again.
A. The Chiplet & Advanced Packaging Loophole
Instead of trying to print one giant, hyper-dense 3nm chip on a single piece of silicon, Chinese packaging giants (like JCET) are leaning into Heterogeneous Integration. They take smaller, modular pieces of silicon (chiplets)—cheaply printed using legally unbanned Deep Ultraviolet (DUV) machines—and stitch them together side-by-side or stack them on ultra-fast microscopic highways inside the casing.
To the software, it performs like a single, cutting-edge sub-3nm chip. By bypassing the physical transistor limit, China can build viable AI hardware without ever touching an EUV machine.
B. The Volumetric Memory Attack
While SK Hynix and Samsung prioritize high-margin HBM for Western AI data centers, China's memory champion, CXMT, is executing a massive market-share offensive. CXMT is flooding the global consumer market with commodity DDR5 and LPDDR5X memory built on DUV nodes.
By drastically undercutting global prices, CXMT has been squeezing Korean margins on standard memory. But flooding the market step one — old news. More importantly, CXMT is rapidly moving from HBM2 to sampling 8-layer HBM3 to power domestic AI processors like Huawei’s Ascend, insulating China’s internal tech ecosystem from Western pricing power. That's step two. Step three is beating the oligopoly at its own game, and I think that's going to start playing out very soon.
C. System-Level Software Efficiency (MoE)
Because Chinese engineers know their hardware is less efficient — it runs slightly hotter and slower, they have to be leaders in software optimization. Chinese AI architectures rely heavily on what are called Mixture-of-Experts (MoE) models, which only activate specific sub-networks of a neural model per query rather than firing the entire system. This drastically cuts the volume of data that needs to pass through the memory chips, allowing lower-bandwidth hardware to run frontier-class AI smoothly. So it's banking on at least slightly better software to outweigh slightly worse tools.
D. Paradigm-Shifting Lithography Alternatives
Looking further down the horizon, China isn't even waiting to master 13.5nm EUV light.
Nanoimprint Lithography (NIL): Startups like Prinano are deploying commercial systems that physically stamp circuit patterns into silicon, eliminating the need for optical lasers entirely for high-density 3D NAND storage. Now, if this were a Silicon Valley startup, I would take any claim of innovation with a huge grain of salt. But it's a Chinese startup and has presumably has the backing of the central government on an ultra-strategic technology. I don't think failure is an option here.
Soft X-Rays via Particle Accelerators: State research institutes in Beijing and Shanghai are building Steady-State Micro-Bunching (SSMB) particle accelerators designed to power entire industrial fabs with 6.7nm Soft X-Rays, aiming to bypass ASML's localized laser monopoly by the early 2030s.
4/ The Rebalancing Horizon
The current financial multiples for Hynix and Samsung are pricing in a long-term monopoly. But history—and physics—tell us otherwise.
Capacity Overhang: Driven by record cash flows, global foundries are executing massive capital expenditure projects. When those new cleanrooms reach full mass production around late 2027 or 2028, high-margin shortages will inevitably turn into market oversupply.
Loss of the Chinese Market: Nvidia’s market share for data center silicon in China has essentially evaporated due to export restrictions, forcing Chinese tech giants to spend billions inside their own borders. Huawei's domestic AI ecosystem is projected to absorb a massive chunk of regional demand, permanently cutting Western and Korean suppliers out of a market that once represented a third of global semiconductor consumption.
Pushback at Home: I don't want to speculate how the AI market will evolve far out into the future, let alone the pace of progress. But I do think that it's going to be a bumpy road and pushback will be plentiful. Issues in the West include environmental concerns, labor market disruption, infrastructure disruption, etc. Not to mention, the last round of earnings have exposed deep investor concern with granting tech giants a blank check to spend on these projects.
The Bottom Line
I'll write a sister article on a more human/psychological side to this bubble soon. But in general, I'm cautious. Stakeholders of the chip oligopolies should enjoy the record earnings while they last. The current stock surge for Korea’s memory giants is real, but it is built on an artificially constrained supply chain and a temporary technological monopoly.
As China’s chiplet architectures mature, its domestic HBM production ramps up, and alternative lithography paradigms come online, the global semiconductor market will rebalance. The cyclicality of the semiconductor industry isn't dead—it's just building up momentum for its next turn.
This is meant to be a criticism, but not a permanent indictment, of what's been happening here in the West. We're playing a tortoise-and-hare game right now, and (newsflash) the West is not the tortoise. There's pressure to innovate all around, but from this angle, that same pressure is much greater on China. The state can also act much more efficiently as a coordinated apparatus, and that's not something to underestimate. From the lessons of the past, one of the paradoxes is that restrictions on technology have, on every single occasion I can recall, resulted in a higher development pressure on the homefront. It's not exclusive to China, but in the Chinese example, key players in aircraft, high speed rail, automotive, solar, and a slew of other industries have had bitter lessons to learn and might be able to shed light on how to compete more effectively.
No comments:
Post a Comment