Hook: The 50% Demand Gap
Most analysts are still modeling HBM supply-demand curves based on past cycle logic. They are wrong. SK Group Chairman Chey Tae-won just threw down a gauntlet that the semiconductor industry can't ignore: total memory demand will surge 50-60% in 2025, with AI-specific memory (HBM) exploding 60-100%. But the real signal is not the demand number itself. It’s his prediction that the supply-demand gap will widen. This is not a cyclical upswing we can model with historical data. This is structural. It means the physical world—factory concrete, EUV tool deliveries, skilled operator training—cannot keep pace with algorithmic hunger. The industry is about to hit a ceiling that no amount of capital expenditure can instantly break through. For the first time in a decade, the bottleneck is not demand. It is the physics of production.
Context: The Man and the Machine
Chey Tae-won is not a random talking head. As Chairman of SK Group, he controls SK hynix, the world’s second-largest memory chipmaker and the absolute leader in High Bandwidth Memory (HBM), the critical component for NVIDIA’s AI accelerators. His remarks, reported by Korea's Maeil Business Newspaper, carry the weight of a CEO who has already bet tens of billions on this thesis. SK hynix is currently building the M15X fab in Cheongju, spending roughly 20 trillion won, and planning a massive advanced packaging facility in the U.S. alongside NVIDIA. This is not a theoretical economist offering a forecast. This is a general preparing for a war he sees coming. His core argument is a direct attack on the traditional semiconductor playbook: stop trying to be a price-maximizing monopolist and start being a supply-maximizing partner. In his view, the era of artificially constricting output to juice margins is over. The AI ecosystem demands volume, and the first company to deliver it will own the future, even if near-term margins compress.
Core: The Seven-Dimensional Deep Dive
To understand if Chey is right, we need to strip away the rhetoric and look at the technical architecture of the problem. I’ll apply my framework across seven dimensions that matter for semiconductor strategy.
1. Technology & Process (Confidence: 5/10)
The core technical moat for SK hynix is not just the DRAM cell node (1anm or 1bnm) but the integration of that cell with advanced packaging. HBM3E’s real magic is the Through-Silicon Via (TSV) and Hybrid Bonding that stacks the dies. Chey doesn't mention these process details, but his emphasis on “facilities, manpower, and construction time” as constraints is a massive tell. It screams that the bottleneck is not in the fab’s lithography, but in the back-end assembly and test (OSAT) capacity. While Samsung and Micron can also make 1bnm DRAM, scaling the TSV and bonding capacity to millions of units per month is a different beast. The hidden insight here is that Chey rates capacity scaling above node shrinks. He’s effectively stating that for the next 2-3 years, the winner of the HBM war will be determined not by transistor density, but by how fast you can build more assembly lines for TSV and Hybrid Bonding. This is a bet against the “miniaturization-at-all-costs” dogma of the last twenty years.
2. Industrial Chain & Bargaining Power (Confidence: 7/10)
SK hynix sits in a powerful but structurally precarious position in the value chain. Its upstream dependency on ASML (EUV) and Japanese materials (TEL, Shin-Etsu) is high, but it is a “mutual hostage” relationship. The real vulnerability is downstream customer concentration. NVIDIA is the 800-pound gorilla. If NVIDIA decides to qualify Samsung’s HBM3E in volume next quarter, or worse, starts designing its own custom memory (a known risk), SK hynix’s margins evaporate overnight. Chey’s call for “chip companies to expand production” is a thinly veiled plea to his rivals (Samsung, Micron) to join him in a collective action against the hyper-concentrated buyer power of NVIDIA. He is saying: “Don’t compete for NVIDIA’s favor by keeping supply tight and prices high. Compete by offering more supply. The only way to reduce NVIDIA’s power is to make HBM so abundant that it becomes a commodity, forcing NVIDIA to design its systems around our capacity, not the other way around.” This is Machiavellian supply chain strategy disguised as an industry forecast.
3. Capacity & Capital Expenditure (Confidence: 8/10)
This is where the rubber meets the road. SK hynix’s planned CapEx for 2024 is an estimated $8-9 billion. The M15X fab is targeting initial production by 2025 H1. But the real bottleneck is equipment delivery. High-NA EUV tools from ASML have a 12-18 month lead time. TSV etchers and hybrid bonders from Tokyo Electron and Disco are similarly back-ordered. Chey’s phrase “construction time constrains capacity” is an understatement. The hidden implication is that the actual bite-size increase in HBM output over the next 18 months is already locked in, regardless of demand. You cannot accelerate a fab. This creates a fundamental supply inelasticity. Even if every HBM maker doubled their Capex tomorrow, the physical output in 2025 would only increase by a fraction of the theoretical max. The market is going to experience a real, physical shortage, not a speculative one. This supports Chey’s thesis of a widening gap.
4. Market Demand (Confidence: 9/10)
The demand side is the strongest pillar of Chey’s argument. AI model training—especially for Transformer architectures—is a massive consumer of memory bandwidth. NVIDIA’s Blackwell B200 GPU is designed to work optimally with 8 stacks of HBM3E. The demand growth from hyperscalers (AWS, Google, Meta, Microsoft) is structurally driven by their need to deploy AI, not by consumer whim. The inventory cycle is in a clear “forceful replenishment” phase. Chey’s 60-100% growth figure for HBM is in line with sell-side consensus from firms like Bernstein and Morgan Stanley. However, the risk he ignores is demand elasticity. If AI training efficiency improves faster than expected (e.g., through better algorithms or sparse computation), the required HBM per unit of compute could drop. But for now, the demand trajectory is a hockey stick. The hidden signal in his speech is that he believes the demand curve is steeper than the supply curve for at least the next 18 months. That’s why he argues for abandoning price controls—he thinks volume growth will more than compensate for any per-unit price decline.
5. Geopolitics & Export Controls (Confidence: 6/10)
Geopolitics is the wildcard. SK hynix operates fabs in China (Wuxi) that are subject to US export controls on advanced equipment. Currently, it has a “validated end-user” authorization, but this is a sword of Damocles. Chey’s push for expanding capacity domestically (Korea) and in the US can be read as a geopolitical hedge. He is building capacity in America to align with NVIDIA (which is also building a US supply chain) and to avoid being caught in a China-US crossfire. His call for “not restricting supply” is also a subtle criticism of the US-led policy of fragmenting the global chip supply chain. He is implicitly saying: if you want to win the AI war, you can’t do it with localized, inefficient supply. You need global scale. The US CHIPS Act is forcing everyone to diversify, which increases costs. Chey’s vision is a global, unified supply chain that maximizes output, regardless of political borders. This is a high-risk political stance, but a logical industrial one.
6. Competitive Landscape (Confidence: 8/10)
This is a pure oligopoly (SK, Samsung, Micron). In HBM, SK holds ~45-50% market share in 2024. Samsung has ~40-45% and is behind in qualification with NVIDIA. Micron is a distant third but closing fast. The competitive game is now about customer lock-in with NVIDIA. Chey’s “don’t restrict supply” message is a strategic nudge to Samsung. He’s saying: “If you (Samsung) cut back your HBM capacity to keep prices high, you will never get NVIDIA’s trust. But if you join me in flooding the market, we can make HBM so available that our combined capacity becomes a competitive moat against anyone trying to design a custom memory solution.” It’s a call to form a tacit duopoly on supply volume, effectively freezing out any new entrant (like a potential in-house memory effort from a hyperscaler) by making the market look terrifyingly over-supplied in the long run, while simultaneously being starved in the short run. The hidden insight: Chey is not afraid of Samsung. He’s afraid of NVIDIA deciding to build its own memory. The best defense against that is to make his own capacity indispensable.
7. Financials & Valuation (Confidence: 7/10)
SK hynix is currently printing cash. Gross margins in H1 2024 were ~45%, driven by HBM which likely has a 60%+ margin. But the valuation (PE ~12-15x) is already discounted for the cyclical nature of memory. Chey’s aggressive CapEx plan puts free cash flow under pressure over the next 2 years. His logic is that the market is underestimating the duration of the AI upcycle. If HBM demand stays elevated for 3-4 years, the current valuation is cheap. If it’s a 1-year peak, the stock is a trap. The hidden layer is that his speech is designed to support the stock price. By publicly predicting massive demand and hinting at his willingness to capture it, he is giving cover for the massive CapEx which would otherwise spook investors. He is selling a narrative of “growth premium” over “cyclical risk.” This is classic CEO signaling for a capital-intensive business.
Contrarian Angle: The Latency Trap
The consensus view is that HBM demand is good, and SK hynix is a prime beneficiary. The contrarian insight is that Chey’s thesis is dangerously self-serving and could be wrong for three reasons. First, he is ignoring the commoditization risk. If he succeeds in flooding the market, HBM will become a high-margin commodity for precisely one generation before prices collapse. His strategy works only if demand grows faster than supply. If NVIDIA hits a product cycle slowdown (e.g., in 2026), the over-capacity will be brutal. Second, he underestimates Samsung’s resolve. Samsung is not just another competitor; it is a state-backed colossus with infinite patience and a willingness to lose money on memory to win a strategic war. It can easily match SK’s capacity expansion, turning a temporary advantage into a perpetual margin war. Third, the “latency trap.” Chey’s plan relies on physical factories (M15X, US packaging). But latency is the enemy of volume. By the time these fabs are fully ramped (2026-2027), the AI market may have moved to HBM4, requiring entirely different packaging techniques. He is building capacity for today’s product, which may be obsolete tomorrow. The smart money might not be on the old leader building more of the same, but on the new entrant designing a more efficient packaging solution.
Takeaway: Actionable Price Levels and Final Judgment
SK hynix is making a bold, high-conviction bet that the future of AI belongs to volume, not margin. For traders, the immediate actionable item is watching the equipment delivery cycle. If ASML reports increasing orders from SK hynix for EUV in the next quarter, the supply crunch thesis is confirmed. If they plateau, the downside risk increases. On the charts, SK hynix (000660.KS) needs to hold the ₩180,000 level to maintain its bull trend. A break above ₩220,000 would signal renewed momentum. But the real trade is not in the stock itself. It’s in the volatility of the HBM ecosystem. If Chey is right, every delay in a rival’s facility (Samsung’s HBM3E qualification) is a catalyst for SK. If he is wrong, every successful qualification by a rival is a risk.
Liquidity vanishes. Conviction remains. The market is now pricing Chey’s conviction. The question is whether the physical world will obey.
Chaos is data waiting to be quantified. And this data says one thing: build more factories.
Ego is the ultimate systemic risk. The biggest ego here is Chey’s belief that he can out-build the demand curve. He might be right. But the margin for error is zero.