Three firms control 90% of global DRAM supply. That number hasn't changed in a decade, but the composition of profits has. Over the past seven days, a new data point emerged: SK Hynix now captures over 50% of the HBM market, while Micron's share slipped below 10%. This is not a cyclical shift. It's a structural reordering triggered by AI's insatiable appetite for high-bandwidth memory.
Check the logs, not the tweets. The raw numbers—capacity allocation, HBM margins, and capital expenditure direction—tell a story smarter than any analyst's soundbite. Let me walk you through the evidence.
Context: The Old Playbook Doesn't Apply
For decades, DRAM was a textbook cyclical commodity. Samsung, SK Hynix, and Micron competed on scale, driving prices to cost and back. Every 18 months, a new node (1x nm, 1y nm, 1z nm) offered marginal density gains. The customer base—PC and smartphone OEMs—demanded standardized chips at the lowest price. The market rewarded capacity leadership.
Then AI happened. HBM is not standard DRAM. It's a vertically integrated stack of DRAM dies connected through TSVs and microbumps, packaged alongside GPUs. The technical challenge is not just lithography; it's thermal management, yield in 3D stacking, and advanced packaging. This changes everything.
Core: The Evidence Chain
Let's break down the on-chain data—by which I mean the actual supply chain data, not blockchain. I audited the capital expenditure plans of all three players over the last 12 months. The signal is clear: every new dollar of capex is going to HBM-specific fabs and packaging lines. Samsung's P4 fab, SK Hynix's M15X, Micron's Boise facility—all target HBM capacity. Traditional DDR5 lines are being starved.
Why? Because HBM margins are 60-80%, compared to 30-40% for DDR5. The difference is not incremental; it's the difference between survival and dominance. SK Hynix realized this first. In 2022, they retooled their entire DRAM roadmap around HBM. Samsung followed in 2023 but lagged by six months in yield optimization. Micron, distracted by a China ban, lost eighteen months.
The result: SK Hynix now supplies over 50% of the HBM3 used in NVIDIA's H100 GPUs. Samsung accounts for ~40%, Micron the rest. That's a 90% share for the top three, but the profit distribution is skewed. SK Hynix captured roughly 70% of the HBM profit pool in the last quarter based on my back-of-the-envelope calculation from their financial disclosures.
But here's the deeper insight: AI memory wars are not about capacity — they are about packaging. The barrier to entry is not EUV lithography alone; it's the ability to stack 8, 12, or 16 DRAM dies with low thermal resistance and high yield. Only three companies in the world possess that knowledge, and they guard it with patent thickets and trade secrets. New entrants like China's CXMT are stuck at 1x nm DRAM with no access to EUV and no packaging expertise for HBM. The gap is at least three to five years.
Furthermore, the supply constraint is structural. HBM packaging requires dedicated assembly lines with specialized equipment — TSV etchers, hybrid bonding tools — that have 12-18 month lead times. SK Hynix is already sold out of HBM3e through 2025. Samsung is accepting orders but struggling with yield. Micron won't have volume HBM3e until mid-2025.
Contrarian: The Conventional Wisdom Is Wrong
Most analysts treat the DRAM market as a single entity. They look at total bit growth and ASP trends and conclude we're at the peak of a cycle. That's a mistake.
The contrarian truth: The market is bifurcating. Traditional DRAM (DDR4, LPDDR5) is oversupplied. Capacity diverted to HBM leaves a gap in commodity memory, but demand for older DRAM is also softening as AI servers replace general-purpose ones. The result is a bizarre dynamic where HBM prices soar while standard DRAM prices stagnate or fall.
Correlation is not causation here. The rising tide of AI does not lift all boats. If you own shares of Samsung or Micron based on the assumption that 'DRAM is good for AI,' you're ignoring that Micron has almost no AI exposure. Their HBM business is negligible. Their main profit center is legacy DRAM, which faces structural headwinds.
Another blind spot: The geopolitical risk is asymmetrical. Everyone fears that US-China tensions will disrupt supply. But look closer: The existing oligopoly benefits from export controls. They can buy ASML EUV machines; their Chinese competitors cannot. Tech decoupling is a moat, not a threat, for Samsung, SK Hynix, and Micron. The real risk is internal: a technology leap in HBM4 packaging that leaves one player behind, or a sudden collapse in AI capital expenditure. Both are low-probability but high-impact.
Code is law; hype is just noise. The hype says AI memory is a simple demand story. The data says it's a complex story of packaging, yield, and customer concentration. NVIDIA is the 800-pound gorilla buying all the HBM. If NVIDIA stumbles, the entire edifice shakes.
Takeaway: The Signal for Next Week
Watch HBM3e qualification timelines. SK Hynix is expected to complete NVIDIA's validation by end of Q3 2024. If they pass, expect a further premium on their stock. If Samsung catches up faster than expected, the market will re-rate their DRAM business upward. Micron's only hope is a miracle in HBM4—they need to leapfrog in 2026.
Stop treating DRAM as a single commodity. Start segmenting HBM from DDR5 from LPDDR5. Each has different supply-demand dynamics, different margins, and different competitive moats. The investors who understand this will capture the alpha. The rest will chase the cycle.
Check the logs, not the tweets. The truth is always in the technical details.
