Hook When a memory maker posts a gross margin north of 50% in Q2 2024, the market cheers. But the code doesn't. Behind the headline lies a single customer dependency ratio that would give any risk manager nightmares — and a technology alliance that both secures and cages the company.
Context SK Hynix is the world's second-largest DRAM maker, but in High Bandwidth Memory (HBM), it holds the pole position. HBM3E — the memory stack powering NVIDIA’s H100 and Blackwell GPUs — has become the company’s profit engine. The Q2 margin record is a direct result of this lock, not of broad cyclical recovery. Meanwhile, HBM4 is on the roadmap, promising custom logic and hybrid bonding, but also exposing the firm to a technology route-switch risk that few analysts quantify.
Core Let’s run the numbers. In HBM3E, SK Hynix commands roughly 50%+ market share. Its R&D efficiency — measured as HBM revenue per dollar of R&D — is 2-3x that of Samsung. The cost of this leadership? A capital expenditure of ~$5-6 billion in 2024 alone, largely directed at new fabs in Cheongju and Indiana. The depreciation drag will hit from 2026 onward, compressing margins by perhaps 10 percentage points.
The real story, however, lies in the supply chain geometry. SK Hynix’s advanced packaging (MR-MUF, and soon hybrid bonding) is its deepest moat. But that moat is tied to TSMC’s CoWoS capacity. The alliance with TSMC for HBM4 is logical — it integrates logic and memory — but it also makes SK Hynix’s HBM4 depend on an external partner’s process roadmap. If TSMC’s hybrid bonding yield lags, SK Hynix loses its first-mover advantage.
Meanwhile, the demand side is dangerously concentrated: NVIDIA accounts for over 70% of HBM purchases. The long-term agreements mentioned in the earnings release lock in volume, but not price. If Samsung’s HBM3E finally passes NVIDIA’s qualification — a real possibility by late 2024 — SK Hynix’s pricing power evaporates overnight.
Contrarian Angle The bulls will say that HBM4’s customization deepens client stickiness. They built on sand; I built on skepticism. Custom logic is a double-edged sword: it increases switching costs but also pushes SK Hynix toward a foundry-like model, reducing product fungibility. If the AI demand cycle dips, those customized stacks become inventory that no other customer can use. Also, the long-term agreements don’t guarantee margins — only volume. In a downturn, renegotiation is inevitable.
Takeaway SK Hynix is not a pure AI growth story; it is a cyclical growth story with a phenomenal tailwind. The margin record is real, but its sustainability depends on two things: keeping Samsung at bay and preventing the HBM4 technology transition from becoming a cost black hole. Cold logic cuts through the noise of FOMO. Check the oracle feeds — especially the yield numbers from Samsung and TSMC. Always.