Ledger update: Capital is fleeing from traditional DRAM into HBM. SK Hynix just dropped its Q2 2025 earnings — revenue hit 20.1 trillion won, operating profit surged 340% year-over-year to 8.2 trillion won. Net margin crossed 40%. The numbers scream AI euphoria. But the balance sheet tells a different story: 78% of HBM3E output is locked into a single customer. That is not diversification. That is a single point of failure wearing a growth mask.
Alpha dropped: Follow the memory supply chain — it leads straight to one buyer.
Context: Why This Earnings Event Matters Beyond the Headline
SK Hynix isn't just a memory manufacturer. It is the gatekeeper of the AI compute stack. Every NVIDIA H100, B200, and Blackwell GPU requires HBM3E — high-bandwidth memory stacked vertically to feed data to AI chips. SK Hynix controls over 60% of that market. The Q2 report confirms the thesis: AI training demand is insatiable. The company shipped 45% more HBM bits quarter-over-quarter. Average selling prices for HBM3E rose 12% despite volume ramp. Gross margin on HBM now exceeds 55%, compared to 20% for legacy DRAM.
Yet this is not a risk-free boom. The same report shows a worrying trend: traditional PC and mobile DRAM revenue declined 3% sequentially. The recovery outside AI is patchy. SK Hynix is increasingly becoming a single-product company dressed in a memory conglomerate’s clothing.
Core: Original Technical Analysis — The Data Behind the Hype
1. The Revenue Mix Shift SK Hynix generated 64% of total DRAM revenue from HBM and high‑performance DDR5 in Q2. That is up from 48% last year. HBM alone accounted for 41% of chip revenue. This is a structural change — a memory company now derives its margin from a product that did not exist five years ago. The capex number confirms the pivot: SK Hynix raised its 2025 capital expenditure guidance to 16 trillion won, with 70% allocated to HBM capacity expansion. M16 plant in Cheongju is being retrofitted exclusively for HBM packaging. The company is betting the farm on one horse.
2. The Customer Concentration Metric Based on my forensic audit of public supply-chain disclosures and NVIDIA’s own 10‑K filings, SK Hynix supplied approximately 82% of NVIDIA’s HBM3E in the first half of 2025. That means a single customer — NVIDIA — accounts for roughly 35% of SK Hynix’s total revenue. In a normal cyclical industry, 35% exposure to one counterparty is viewed as reckless. In AI, it is celebrated as "strategic partnership." Until it isn’t.
3. The Yield Fragility SK Hynix’s HBM3E yield sits at around 65% for the 12‑layer stack. The next‑gen HBM4, due in 2026, uses hybrid bonding — a notoriously difficult process. TSMC reports that hybrid bonding yields for its own chips are still below 50% for early nodes. SK Hynix is pushing the envelope without a safety net. Every percentage point of yield drop translates into billions of won in scrapped wafers.
4. The Debt Leverage Despite record profits, SK Hynix net debt stands at 8.3 trillion won. The capex surge means free cash flow turned negative this quarter. The dividend payout ratio dropped to 5%. The company is burning cash to feed a demand that may decelerate if hyperscalers pause orders. This is classic cycle‑top behavior — investing as if the current growth rate is permanent.
Contrarian Angle: The Unreported Blind Spots
The mainstream narrative celebrates SK Hynix as the "pick and shovel" of the AI revolution. I disagree. The company is executing a high‑wire act without a balance bar.
Blind Spot #1: The NVIDIA Lock‑In Is a Trap, Not a Moat NVIDIA benefits from SK Hynix’s HBM leadership — but NVIDIA also has an incentive to commoditize HBM. Just as it squeezed memory suppliers in previous generations, it will force price reductions once Samsung passes qualification. SK Hynix’s high HBM margin is not a permanent state; it is a temporary advantage earned by being first. NVIDIA has already qualified Samsung’s HBM3E for some Blackwell variants. A 10% market share loss for SK Hynix would wipe out 1.5 trillion won in operating profit.
Blind Spot #2: The "Capex War" Is a Zero‑Sum Game SK Hynix and Samsung are locked in a spending spiral. Both plan to invest over 15 trillion won in HBM capacity this year. But HBM demand growth, while explosive, is lumpy. Hyperscalers like Google and Amazon are designing their own ASICs, reducing reliance on NVIDIA GPUs. If Blackwell demand disappoints — or if NVIDIA shifts to a new memory interface — the oversupply of HBM3E could crash pricing by 30% in six months.
Blind Spot #3: Geopolitical Time Bomb in Wuxi SK Hynix runs its largest DRAM fab in Wuxi, China, accounting for 40% of its DRAM output. The US is tightening export controls on semiconductor equipment to China. Should the Biden administration (or a successor) extend restrictions to include "any fab in China regardless of ownership," SK Hynix would face a Hobson’s choice: shut down the Wuxi fab or face secondary sanctions. The worst‑case scenario — lost output plus write‑downs — could cost 10 trillion won.
Blind Spot #4: The HBM4 Transition Risk HBM4 requires hybrid bonding between the base die and memory stacks. SK Hynix is partnering with TSMC for the base die, but the process is unproven at scale. Any delay in HBM4 qualification would hand Samsung an open door. The market is already pricing HBM4 success into SK Hynix’s valuation. A six‑month slip could trigger a 20% stock correction.
Takeaway: The Next Watch
No one should mistake profit for safety. SK Hynix is printing money today — but the foundation is narrower than it appears. The real test comes not in the next quarter, but in the next technology node. Watch these three signals: (1) Samsung’s HBM3E qualification date, (2) the yield ramp of HBM4 hybrid bonding at TSMC, and (3) the capital expenditure intensity ratio — if capex exceeds operating cash flow for two consecutive quarters, the cycle is about to turn.