Hook

On October 24, 2024, SK Hynix closed its trading session down 3.4% despite reporting record quarterly revenue of 17.6 trillion won. The market’s reaction was cold, clinical, and instructive. The baseline was not missed revenue—it was missed expectations on margin quality and capital discipline.

Assumption is the adversary of verification.

Investors had priced in a perfect HBM scale-up. The actual data told a different story.

Context

SK Hynix is the world’s second-largest DRAM manufacturer and the dominant supplier of HBM (High Bandwidth Memory) for NVIDIA’s AI accelerators. Over the past 18 months, HBM delivery has been the bridge between AI chip designers and real-world inference clusters. The company’s HBM3E generation, fabricated on 1β nm DRAM dies and stacked using MR-MUF packaging, achieved first-mover status in 2023. NVIDIA’s H100, B200, and upcoming Rubin architectures all rely on SK Hynix for a significant portion of their HBM supply.

But dominance in a tight oligopoly does not guarantee fat margins. The bull case for SK Hynix was simple: HBM demand outstrips supply; therefore, any capacity added would be absorbed at premium prices. The stock doubled in 2024 on this logic. The earnings call corrected that narrative.

Core: The Forensic Takedown

Data from SK Hynix Q3 2024 earnings: - Revenue: 17.6 trillion won (beats consensus of 17.1 trillion) - Operating profit: 7.0 trillion won (in line with estimates) - Gross margin: 52% (missed sell-side estimates of 55%) - Operating cash flow: 12.2 trillion won (strong) - Capital expenditure: 14.5 trillion won (above expectations of 13.2 trillion)

The headline numbers looked healthy. Revenue grew 12% quarter-over-quarter. HBM revenue rose 30% sequentially. But the crack in the glass began to show in the details.

Gross margin compression: Despite record HBM volume, gross margin declined from 54% in Q2 to 52% in Q3. The company attributed this to rising depreciation from new cleanroom expansions and lower-than-expected HBM3E yield rates. Based on my audit experience with similar capital-intensive ramp-ups, the yield drag at 1β nm HBM3E likely sits between 15-20% compared to mature node. In plain terms: higher unit costs eroding the pricing power premium.

Capital expenditure intensity: SK Hynix disclosed CapEx of 14.5 trillion won for Q3 alone, representing 82% of revenue. This is unprecedented outside foundry businesses. The company is building M15X (Cheongju) and expanding its advanced packaging capacity in Icheon. While the long-term logic is sound—more HBM capacity = more revenue—the short-term math is punishing. Every won invested returns with a 12-18 month lag, during which the company absorbs depreciation costs on idle or semi-utilized front ends. The implied CapEx-to-revenue ratio is above 50% for the full year, versus peers like Micron at 35% and Samsung at 30%.

Gross margins on HBM are not what you think: Market consensus often assumes HBM carries a 10-15 percentage point margin premium over standard DRAM. Based on the disclosed blended margin and management commentary, I estimate that HBM’s actual gross margin lies between 45% and 50%, with mature node DRAM at 55%+. The reason is complex—packaging cost, TSV yield, test time. But the narrative that HBM is a magic margin lever is false.

Revenue concentration: SK Hynix remains heavily dependent on a single customer—NVIDIA. The top-5 customer accounts for over 75% of HBM revenue. This is the same risk profile we warned about in the 2017 ICO audits: single-point-of-failure in counterparty dependency. If NVIDIA’s GPU roadmap shifts, or if Samsung HBM3E passes qualification earlier than expected, SK Hynix faces immediate revenue downside with limited diversification.

Free cash flow turn negative: Operating cash flow was positive, but after deducting CapEx, free cash flow turned negative (-2.3 trillion won for Q3). This is the first negative quarter since Q1 2023. A negative FCF quarter during a bull market is a yellow flag, not a red one, but the pattern is important. If CapEx remains at this level for four consecutive quarters, the company will need to raise debt or equity to maintain liquidity buffer.

Hidden signal: Market disappointment was not because of a miss in revenue or profit. It was because the market realized that the scale-up is harder than expected and that the premium returns on HBM are not as high as earlier assumed. This is the same pattern I observed in the 2021 NFT minting statistics: when the early adopters (investors) extrapolate linear returns from exponential hype, the actual data reveals a step function. Here, the step function is engineering delays.

Contrarian Angle: What the Bulls Got Right

To be fair, SK Hynix is executing better than any competitor on HBM. Its MR-MUF packaging technology is more thermally efficient than Samsung’s TC-NCF. Its 1β nm yield is likely ahead of schedule. In the HBM3E generation, it has delivered over 40% of NVIDIA’s supply, with Samsung lagging by at least one quarter in final qualification. The bulls are correct that SK Hynix will retain the leading share in HBM4 (expected 2026) if it successfully transitions to hybrid bonding (Hybrid-Cu bonding).

From a forward-looking perspective, the company’s gross margin trajectory should improve as HBM3E yield stabilizes and packaging complexity reduces. The depreciation hit will also plateau once M15X is fully online in 2026. The negative FCF is a temporary symptom of an investment cycle, not a structural disease.

But the bulls miss one critical point: revenue concentration risk increases as the HBM market grows. If SK Hynix supplies 50% of 2030’s HBM market, and that market is 200 trillion won, then its single customer will be NVIDIA, which will hold 90%+ market share in AI training. A renegotiation of pricing terms by NVIDIA could eat 10% of SK Hynix’s gross margin overnight. The asymmetry of power is growing, not shrinking.

Takeaway

The SK Hynix earnings call was not a bad report. It was an honest one. It revealed that scaling HBM from niche to mainstream is an engineering challenge, not a finance one. The market had priced in frictionless growth. Reality has friction.

Investors should ask: if gross margins cannot expand further on HBM, and CapEx stays elevated, what drives the stock multiple higher? The answer may be nothing—until the cycle resets. Or until SK Hynix finds a second customer. Until then, the on-chain evidence says: the ledger remembers everything.