SK Hynix (000660.KS): The Bandwidth Bottleneck of Artificial Intelligence

$000660.KS is up 65 percent in the last 30 days. (Aside: who decides the stock ticker convention in Korea? It looks like a phone number.) South Korea's SK Hynix is not really a chipmaker in the way the headlines describe it. It is the bandwidth bottleneck of artificial intelligence. Without its HBM, every NVIDIA Blackwell becomes a Ferrari with no fuel line.

That sentence is doing a lot of work, so let's unpack it. A modern AI accelerator like Blackwell or Rubin is, at the physics level, a bandwidth-bound machine. Compute scales roughly with transistor density, which keeps doubling. The rate at which memory can feed those transistors does not. That gap is exactly what High Bandwidth Memory (HBM) is designed to close. Stack the DRAM dies vertically, connect them through the silicon with thousands of through-silicon vias, sit the stack right next to the GPU on a shared interposer, and you get an order of magnitude more memory bandwidth than conventional DDR. The trick is, only a handful of fabs in the world can actually build it.

Why This Is Not Top Glove

Compare the SK Hynix rally to the Malaysian glove maker bubble of 2020. Top Glove ten-bagged inside a year, then swiftly collapsed back to roughly where it started. Why? Because a glove plant takes six months to build and anyone with a credit line can enter the market. The capacity response was almost instantaneous. Once Covid demand normalised, the price collapse came faster than the build-out ever could.

Memory fabs are not gloves. A new memory fab takes four to five years and 30 billion dollars of capital before it produces a single competitive die. HBM specifically requires process know-how that has been compounded over thirty years of running DRAM at the bleeding edge. There is no weekend-warrior version of this. The pool of companies that can credibly add HBM capacity in the next five years is not a long list.

"Capital intensity, when paired with a multi-year process learning curve, is one of the most underappreciated moats in modern markets."

The Field: One Player Holding All Three Boards

Look at the global memory field today. Samsung leads conventional DRAM but trails in HBM, with roughly 35 percent share and a TC-NCF packaging approach that has had yield issues. Micron is gaining in HBM but is essentially absent in enterprise NAND. Kioxia is NAND-only.

SK Hynix is the only one playing all three boards: HBM at 62 percent share, conventional DRAM at 33 percent, and the number two position in enterprise SSD. That tri-segment presence is what lets management speak in three-year supply horizons rather than quarter-to-quarter cycle calls.

Catalysts Stacking Through 2027

The catalyst calendar over the next twelve months is unusually dense:

• HBM4 co-developed with TSMC. For the first time, the HBM logic base die is being fabricated by TSMC rather than by the memory maker itself. This pulls SK Hynix deeper into the foundry-grade ecosystem and tightens its co-dependency with NVIDIA.
• M15X fab live four months early. Adds material capacity into a market where management says demand is already running ahead of supply.
• Yongin Phase 1 pulled forward to February 2027. The Yongin cluster is the long-dated capacity story. Pulling Phase 1 forward by months is a signal that customer commitments are binding.
• US ADR listing targeted June. Widens the pool of US capital that can hold the name without jumping through Korean settlement hoops.

Management's own line on the latest call: “Demand for the next three years already exceeds our supply capacity.” Take it with the usual grain of salt, but the capacity bring-forwards make it harder to dismiss as pure cheerleading.

The MoatMap Scorecard: Q77 V29 M66

Numbers anchor the narrative. Here is the SK Hynix MoatMap StockRank:

• Quality: 77/100. ROIC of 37 percent, ROE of 61.2 percent, operating margin of 71.5 percent. Capital efficiency at the high end of any industrial business globally.
• Value: 29/100. Trailing multiples look stretched. The score likely underweights forward earnings power if the supply-demand setup holds.
• Momentum: 66/100. RSI 91, less than 1 percent from the 52-week high. Strong, but flashing technically overbought.
• Composite StockRank: 89/100. Strong buy on the composite, dragged down only by the value factor.

A Quality-77 / Value-29 / Momentum-66 profile is the classic AI winner shape: extraordinary unit economics, elevated price, strong momentum. We covered how to think about that mix in our guide to factor investing.

The Cyclicality Question

Memory has been one of the most violently cyclical industries in technology for forty years. Boom, glut, bust, repeat. Anyone old enough to remember 1996, 2001, 2008, 2019 or 2023 knows that capacity discipline tends to break exactly when pricing peaks. Calling it different this time is the most expensive sentence in finance.

The honest question is whether AI inference is a structural floor that breaks the cycle, or just a longer rope before the same hanging. The bull case rests on three points: AI inference is becoming embedded in every consumer product, not just hyperscaler training; HBM capacity is structurally constrained for years, not quarters; and the customers building the demand (NVIDIA, the cloud providers, the sovereign-AI buyers) have shown they will pay for guaranteed supply rather than spot.

The bear case is that all three things are true and still priced in. Memory cycles peak when capacity comes online and customers blink. The 65 percent rally in 30 days is the kind of move that historically marks the late innings, not the early ones. We are not arguing the cycle is dead. We are arguing the floor is rising.

Capital Discipline Is Quietly Real

One thing the noise around the rally tends to drown out: SK Hynix held its KRW 1,200 per share dividend flat through the 2022-2023 memory losses, then raised it 25 percent in 2025. That is a meaningful signal. Most cyclicals cut at the bottom. Holding the dividend through a real downturn says something about the balance sheet and management's read of the cycle.

On top of that, the company is retiring all 50 million treasury shares, roughly 2.1 percent of float, worth about KRW 12.2 trillion. Total FY2025 returns are running above KRW 14 trillion. The headline yield is modest, but the capital allocation framework is closer to a US-style buyback-and-raise pattern than a typical Korean conglomerate.

The Bottom Line

SK Hynix is not a stock you buy because it is cheap. It is a stock you own because the global AI build-out cannot happen without it, and there is no realistic substitute on a three-to-five year horizon. The cyclical risk is real. The multi-year demand visibility is also real. The 65 percent rally compresses some of the upside but does not by itself invalidate the thesis.

For investors thinking about how a position like this fits inside a broader portfolio (concentration, factor exposure, cyclical risk budgeting), our guide to reviewing your portfolio for weak spots is the natural follow-on read. For more on how Quality, Value and Momentum interact in glamour-shaped names, see screening stocks with the QVM framework.

For the full breakdown including segment economics, customer concentration, capex trajectory, the cyclicality bear case, and the management quality assessment, the SK Hynix Deep Dive is the place to go.

Disclosure: this article is for informational purposes only and is not investment advice.