The market is punishing Samsung Electronics and SK Hynix. The narrative is familiar: peak memory cycle, slowing demand, geopolitical overhang. But on-chain data—or in this case, on-the-ground semiconductor fundamentals—tells a different story. Meritz Securities analyst Kim Sunwoo recently published a deep-dive report arguing that the current sell-off is a "misunderstanding" and "over-correction," driven by a failure to grasp the structural shift in DRAM demand powered by AI.
I’ve spent 20 years in semiconductor research. I’ve tracked every cycle from 256Mb DRAM to HBM3E. This report is not just another buy thesis. It’s a roadmap of what happens when the market misprices the most critical component of the AI stack: memory bandwidth. Let me break down why Kim’s core argument holds water, where it’s dangerously silent, and what data points you need to watch over the next 12 months.
The Hook: A 60–75% Supply Fulfillment Rate
Here’s the number that should jolt you: Kim projects that DRAM supply will only meet 60–75% of demand in 2024–2025. In any commodity market, a 25–40% supply deficit triggers a pricing supercycle. Memory is no exception. If this holds, we’re looking at a sustained upward price trajectory that current analyst consensus has completely failed to price in.
Why such a gap? It’s not just about HBM—though that’s the flashiest driver. It’s about the reallocation of fab capacity. Samsung and SK Hynix are converting legacy DRAM lines to HBM and DDR5, which consumes more wafer area per bit. At the same time, non-AI demand (PCs, smartphones, servers) is recovering mildly. The result: the industry is running at full utilization but still can’t keep up.
Context: The Data Methodology Behind the Call
Kim’s analysis is built on three pillars. First, AI capital expenditure trajectory. He tracks CapEx guidance from Microsoft, Google, Amazon, Meta, and Nvidia. These five firms alone are expected to spend over $200 billion on AI infrastructure in 2025, up from ~$150 billion in 2024. Each dollar spent on GPUs requires a proportional spend on HBM and high-speed DRAM.
Second, long-term supply agreements (LTA) . SK Hynix and Samsung have signed multi-year contracts with hyperscalers that lock in volume but also guarantee pricing floors. Kim argues these LTAs are a strategic moat—they prevent spot price crashes and enable capital-intensive capacity expansion.
Third, shareholder return catalysts. Samsung has announced a stock buyback and cancellation program, while SK Hynix is expected to boost dividends. In a bearish tape, these actions provide a floor for equity valuations.
Core: The On-Chain Evidence—Wait, It’s Not On-Chain, But the Analogy Works
In crypto, we say "follow the gas, not the hype." In semiconductors, follow the bit supply growth vs. bit demand growth. Kim’s report is essentially a liquidity analysis of the memory market. Let me translate it into a framework I use for protocol valuations:
- Supply side: DRAM bit supply growth is slowing to 10–15% in 2025, down from historical 20%+ due to node migration difficulties and limited EUV tool availability. This is like an L1 blockchain hitting a block size limit without a scalability upgrade.
- Demand side: AI-driven bit demand is growing at 30–50% CAGR. HBM consumption alone will account for 20–25% of total DRAM bits by 2025, up from 5% in 2023. That’s a demand explosion that cannot be met by existing fabs.
- Price elasticity: Unlike NAND, DRAM pricing is less elastic because substitution is difficult. You can’t easily replace HBM with a slower memory type in an AI accelerator. This creates a pricing power dynamic closer to a monopolistic bottleneck than a commodity.
Kim projects that even with full utilization and new fab construction (Pyeongtaek, Taylor), the supply gap persists until late 2026. This is the structural thesis.
Contrarian: Correlation ≠ Causation—And What the Report Ignores
Here’s where I push back. Kim builds a compelling chain of logic, but he treats the causal link between AI CapEx and DRAM pricing as nearly deterministic. It’s not. There are three blind spots:
- AI CapEx is not irreversible. If the US economy enters a recession in H2 2025—something the bond market is already pricing—hyperscalers could pause or slow their AI buildout. Kim does not attach a probability to this. He treats it as a tail risk, but I’d argue it’s a 30–40% probability. If that happens, the 60–75% fulfillment rate flips to oversupply.
- China’s memory capacity expansion is real. YMTC and CXMT (ChangXin Memory Technologies) are scaling up 1x nm DRAM and 200+ layer NAND. They face US equipment restrictions, but they are innovating with self-aligned patterning and hybrid bonding. A successful breakthrough could flood the mid-range market, compressing margins for Samsung and SK Hynix. Kim completely ignores this.
- HBM technology gaps are narrowing. SK Hynix leads in HBM3E with 16-layer stack technology. Samsung is behind but catching up. However, if Samsung fails to secure Nvidia qualification for its HBM3E in Q2 2025, its earnings trajectory will diverge sharply. The report does not address this binary outcome.
The Real Alpha: Tracking the Next Week’s Signal
For the next 30 days, ignore the noise about "memory peak" vs. "supercycle." Watch these three on-the-ground data points:
- DRAMeXchange spot prices for DDR5 and HBM contract renewals. If spot prices hold or dip less than 3% in April, the supply tightness is real.
- SK Hynix’s Q2 2025 earnings pre-announcement (expected late May). If they guide HBM revenue above 40% of total DRAM, the thesis is intact.
- Samsung’s HBM qualification news flow. Any official press release about Nvidia certification is a strong catalyst.
Takeaway: The Market Is Short Volatility, Not Memory
The current valuation of Samsung and SK Hynix implies that DRAM pricing will revert to historical mean (60–70% gross margins) within 12 months. Kim’s data suggests otherwise. The next 6 months will determine whether 2024–2025 becomes a memory supercycle or a false dawn. My base case: the bull thesis is correct for SK Hynix but overextended for Samsung due to HBM execution risk. Hedge by pairing a long SK Hynix position with a Samsung put spread.
"Alpha hides in the margins." The margin here is the gap between AI CapEx guidance and DRAM supply growth. Follow that gap, not the hype.
