SK Hynix at 2.24 Million Won: The Rally Has Already Bought the Story

The market believes SK Hynix is the cleanest way to own the AI memory cycle, and the tape agrees. Shares went from 1.06 million won to 2.24 million won in roughly ninety days, a 111% move, on a sequence of confirmations: the Yongin cluster capex, an eight-billion-dollar ASML equipment order, an early M15X ramp, a confidential US ADR filing, a record first-quarter print of 52.6 trillion won in revenue and 37.6 trillion in operating profit, and the Nvidia Vera Rubin SOCAMM2 exclusivity. Then the hyperscalers walked out and waved roughly 700 billion dollars of 2026 AI capex at the room. The stock printed a 13% single-day rally to a record. The consensus average target sits at 2.08 million won. The current price is 2.24 million. The rally has, in the most literal arithmetic sense, eaten the upside the sell side had modeled.

The story is internally consistent. It is also fully owned.

The trouble starts when you walk the analyst causal chain in the order it actually runs. Utilization is pinned. Pricing is the swing variable. Gross margin is the output, not the input. A 71.53% operating margin in a memory business is not a steady-state number you mid-cycle off of. It is the peak of the peak-to-trough EPS range, achieved when HBM is sold out, conventional DRAM is tight, and customer inventory days are thin enough that they will qualify whatever you can tape out. That is the environment now. The replacement cost of a fab does not change because Alphabet raised its capex guide by a hundred billion dollars. What changes is the multiple the market will pay on trailing earnings that, by construction, cannot be the right denominator.

The specific signal the consensus is glossing over sits inside the DRAM mix itself. HBM4 is 40% of DRAM. That is remarkable and also the source of the asymmetry. The other 60%, commodity DDR5, DDR4, mobile, legacy nodes, is still the part of the business whose pricing turns first when hyperscaler order books are revised, even modestly. HBM is binned, qualified, and contracted; commodity DRAM is spot-adjacent and behaves accordingly. A weak won, currently around 1,506 per dollar, is flattering the won-denominated revenue line on every bit shipped, including the bits that would otherwise be telling you something about volume. When the silent variable is how fast HBM versus commodity DRAM cycles through finished-goods inventory, a headline operating margin tells you little about where free cash flow inflects next.

The bull case I take seriously is structural, not cyclical. HBM is not a commodity bin. It is a co-designed, customer-qualified product with a reticle-limited photomask set and a fabless designer counterparty, Nvidia in the binding case, that cannot dual-source on a whim. If HBM4E and the SOCAMM2 module business genuinely behave like a multi-year contract pipeline rather than a memory cycle, then mid-cycle earnings power resets meaningfully higher and 21-times trailing on a peak quarter looks defensible. The market made the same assumption about memory exceptionalism in prior super-cycles, when the argument was always that this time the demand source was different: servers, then smartphones, then cloud, now accelerators. Each time the believing was sincere. Each time the binning eventually caught up with the pricing.

The trouble with that case is the arithmetic already in the price. A consensus target seven percent below spot is not a wall of worry. It is a building already leased to the top floor. Multiple target hikes in May, the trillion-dollar-club chatter, 10.4 billion dollars sitting in DRAM-themed ETFs, none of this is the texture of a misunderstood stock. Markets love stacking assumptions like that right before they rediscover gravity.

The backdrop deserves one paragraph and not more. Washington shifted from indefinite VEU waivers to annual export licenses for chipmaking equipment shipments to the China fabs for 2026, which means the China DRAM and NAND footprint now runs on a yearly regulatory renewal. Beijing’s parallel push for 50% domestic equipment sourcing inside China fabs by decade’s end adds a second compliance layer on the same assets. Neither is an acute earnings event in the next two quarters. Both quietly raise the discount rate the market should be applying to the long-duration cash flows the bulls are extrapolating. The current price is not applying that discount.

What would make me wrong is not subtle. If HBM4E samples qualify on schedule in the second half and the 2027 HBM order book lengthens to a credible two-year contracted pipeline at flat-to-up pricing, the cycle frame breaks and the stock is reasonably valued on forward earnings power rather than expensively valued on trailing. If quarterly HBM mix climbs above 50% of DRAM revenue and commodity DRAM ASPs stay flat sequentially through the year, my concern about the 60% non-HBM tail is misplaced.

If neither holds and commodity DRAM ASPs roll over by more than 10% in any quarter before HBM mix crosses 50%, the operating margin reverts mechanically and the trailing multiple becomes the wrong multiple on the wrong number.

Priced like a contracted franchise. Operates like a memory company having an exceptional quarter. These two things are not the same, and the gap between them is what the buyer at 2.24 million won is underwriting. The sell side modeled the rally and the rally arrived early. What I cannot find in the current tape is anyone being paid to ask what happens at the next inventory-days print from the hyperscalers.

© Mathstock