Hook
SK Hynix reported its most profitable quarter in history. Revenue hit 20 trillion won. Operating profit surged to 8 trillion won. The market responded by dumping the stock. Down 4% in a single session. The narrative: “missed expectations.”
This is the paradox of the AI boom. A company that owns 50% of the HBM market, the critical memory for every Nvidia GPU, posts an all-time high. Yet the algos see a flaw. A structural discount baked into the price.
I do not trust the pitch. I audit the structure.
Context
SK Hynix is a memory IDM — integrated device manufacturer. It designs, fabs, and packages DRAM and NAND. Its crown jewel is High Bandwidth Memory (HBM), a 3D-stacked DRAM solution that sits next to AI accelerators. HBM3E, the current generation, is the bottleneck in every H100 and B100 GPU. Supply is tight. Pricing is high.
For the past two years, Hynix has ridden the AI wave. Revenue doubled. Margins swung from negative to 40%. Capital expenditures soared to 12 trillion won per year. The company is building new fabs — M15X in Cheongju, converting lines in Icheon, planning a mega-cluster in Yongin.
The bull case is simple: AI training demand is exponential. Hynix is the leader. Earnings will compound.
The market is now saying: not enough.
Core
Let me decompose the “miss” into three structural layers. Each reveals why the market is pricing in a discount, not a premium.
Layer 1: Capital Intensity
Hynix’s operating cash flow in Q2 2024 is approximately 9 trillion won. Capital expenditure for the year is 12 trillion won. That means free cash flow is negative — roughly -3 trillion won.
The company is borrowing to grow. It is issuing bonds, raising equity, and diluting future cash flows to fund its expansion. The market sees this. A company that earns 8 trillion in profit but needs to invest 12 trillion to sustain that growth is not a cash machine. It is a capital furnace.
Record profits do not equal value creation if the incremental capital employed generates returns below the cost of capital. Hynix’s ROIC is currently around 15%. That exceeds its WACC of 10%. But the differential is thin. Any slowdown in HBM demand or rise in competition will compress this spread. The market is pricing that risk.
Layer 2: Customer Concentration
Nvidia accounts for an estimated 60-70% of Hynix’s HBM revenue. That is a single point of failure. Hynix’s “record” is a reflection of Nvidia’s success, not its own diversified strength.
In crypto, we call this a single-collateral system. One oracle failure and the whole DeFi protocol liquidates. Here, one shift in Nvidia’s procurement strategy — a pivot to Samsung for HBM4, or a move to vertical integration — and Hynix’s top line collapses.
Market makers understand this. They apply a conglomerate discount to single-client exposures. Hynix’s P/E of 12x reflects that discount. A diversified company with the same cash flows would trade at 15-18x.
Layer 3: The Capex Damocles
Hynix’s current generation of HBM3E relies on MR-MUF (Mass Reflow Molded Underfill) packaging. It gives them a 6-12 month lead in yield and thermal performance. But Samsung and Micron are closing. Samsung’s TC-NCF process is catching up. Micron is ramping HBM3E mass production in 2024.
The next generation, HBM4, will require hybrid bonding and a custom logic die. Hynix is partnering with TSMC for that logic. Samsung is building its own. The competition intensifies.
Meanwhile, the capital already sunk into M15X and Icheon conversions — over 20 trillion won — will begin to depreciate in 2025. Depreciation will eat into gross margins by 5-8 percentage points. To maintain a 40% gross margin, Hynix must either raise prices (harder with multiple suppliers) or achieve scale faster than depreciation ramps.
The market is betting that scale will be insufficient. Hence the price drop on record earnings.
Contrarian
Let me play the bull’s advocate. The market may be overcorrecting.
First, the demand for HBM is not cyclical — it is structural. Every major CSP (Amazon, Google, Microsoft, Meta) is doubling down on AI infrastructure. Nvidia’s B100 and B200 GPUs will ship in volume in 2025. Each GPU requires eight HBM3E stacks. Supply remains constrained through 2025. Hynix’s lead in yield gives it a pricing advantage that will persist into early 2026.
Second, the partnership with TSMC for HBM4 is a strategic moat. By co-designing the logic die with TSMC, Hynix locks Nvidia into a customized solution. Switching costs become high. Nvidia cannot easily swap to Samsung if the interface and physical design are deeply integrated with TSMC’s process. This is analogous to how Apple locks in suppliers: custom silicon, shared IP, tight co-engineering.
Third, the “negative free cash flow” narrative is temporary. Once the M15X fab reaches high-volume manufacturing in late 2025, depreciation peaks and then stabilizes. Operating cash flow will grow faster than capex. Free cash flow turns positive by 2026. The market may be ignoring the convexity of the capex profile.
Emotion is a variable I exclude from the equation. But the equation itself may be missing a term: the option value of AI dominance. If AI follows a Moore’s Law trajectory, Hynix’s cumulative cash flows over the next decade dwarf the current capex. The negative FCF is a front-loaded investment, not a structural flaw.
Takeaway
Hynix is not a growth stock. It is a capital-intensive commodity producer with a temporary monopoly on a critical input. The market is correct to discount the headline profit. But it may be over-discounting the duration of the monopoly.
The question is not whether Hynix can earn more next quarter. It can. The question is whether the incremental won invested today will generate a return that exceeds the cost of capital over a full cycle.
Based on my audit experience — having watched ICOs raise $50 million on code that could not pass a basic reentrancy test — I know that the most dangerous narratives are the ones that sound rational. “AI demand is infinite” sounds rational. It is also a trap. The balance sheet reveals what the income statement conceals.