SK Hynix: The $38 Billion Capex Nobody Actually Needs to Worry About

SK Hynix announced a $38 billion expansion plan on Friday. The market treated it like a warning sign. The stock dropped nearly 4% in midday trading, extending a three-week pullback that has erased roughly $230 billion from its market cap since mid-July.

The narrative forming around the move is straightforward: too much capex, too much supply coming, the AI memory boom is about to peak. It's the same script the market has run through four times before. And it's wrong this time for the same reason it's always wrong in this cycle — the market is reacting to 2029 production capacity while the earnings expansion is happening in 2026.

SK Hynix reported Q2 revenue that surged 257% year over year to 79.32 trillion won ($64.6 billion). Operating profit jumped 557% compared to the prior year. Earnings per share came in at $8.99, blowing past the $5.12 consensus estimate. That's not a company approaching a cycle top — that's a company in the explosive phase of one.

The $38 billion in capital spending the market is fixated on won't produce a single chip until the late 2020s. The Yongin DRAM plant doesn't break ground until July 2027, with the first cleanroom opening in June 2029. The Cheongju NAND plant breaks ground in February 2027, with production starting December 2028. Counterpoint Research's Neil Shah was direct about it: this spending doesn't change near-term output. Memory prices, which have surged on short supply and AI demand, aren't expected to soften before the end of 2028.

The disconnect between what the market is pricing and what the earnings trajectory shows up immediately in the valuation.

SK Hynix trades at approximately 5.5 times forward earnings. Against FY2027 consensus EPS of roughly $30, the stock sits around 4.6x. Even if you're skeptical of how high memory margins sustain, that multiple doesn't price in a stable business — it prices in a business that's about to stop growing. Yet Q2 revenue grew 51% sequentially on top of 257% annual growth.

The stock has fallen from its July 14 peak of $194.80 to around $138 today, a drop of roughly 29% in three weeks. Part of this is the Nasdaq ADR listing still establishing its trading base — SK Hynix debuted on July 10 at $149 per ADS, making it the second-largest US listing in history after raising $26.5 billion. Volatility is to be expected in the first month of US-listed trading for a company that had no direct US access until a week ago. But the scale of the pullback suggests something deeper than new-listing noise.

The bear case has two legs: historical cyclicality and capex overhang. The cyclicality argument is real — SK Hynix posted a negative 28% net margin in 2023, and the memory industry has crashed hard four times since the 1990s. The capex overhang argument is what the $38 billion announcement feeds. Both were present in April 2025 when the stock traded at 4.87x forward earnings. Both were present in September 2025 when it traded at 6x. The stock went up 9x over that period entirely through earnings growth, with almost no multiple expansion. The multiple is still at roughly the same level today.

What's changed since then is the structural nature of the demand. SK Hynix holds 58% of the high-bandwidth memory market — the memory used behind NVIDIA, AMD, Broadcom, Google, and custom silicon AI chips. HBM is shifting from a fungible commodity to a bespoke product, with HBM4 base dies designed around specific customer requirements. That increases switching costs and makes qualification a competitive moat. SK Hynix has hundreds of millions in prepayments locked in for HBM allocations through the end of the decade and exclusive supply qualification for Microsoft's Maia 200 AI chip.

The market is treating a 2029 capacity addition as a 2026 earnings problem. Those are not the same variable.

AInvest's aggregate signal labels the stock a Buy. That's not the point. The point is that 11 out of 12 covering analysts rate it Buy or Strong Buy, with an average price target of $245.50 — 78% above the current price. Rosenblatt initiated at Buy with a $320 target. Cantor Fitzgerald initiated at Overweight with $300. Even the most conservative targets sit 45% above today's level. The analyst consensus isn't the evidence, but the gap between what these estimates imply and what the market price reflects is the entire thesis.

At roughly 5.5x forward earnings, SK Hynix trades at about a third of the multiple you'd pay for AMD and less than a tenth of Broadcom. Both of those companies also sell AI infrastructure. SK Hynix doesn't design chips — it makes the memory that those chips need to function. If you believe the AI buildout is real, the memory bottleneck is the part of the supply chain where demand is tightest and pricing power is strongest.

The risks are real. The memory cycle has a habit of punishing complacency. Samsung reclaimed the number-one DRAM market share spot in Q2 2026, showing SK Hynix's lead isn't automatic. Hyperscaler capex spending is contractually committed for now, but it's not perpetual. If the AI buildout slows materially before 2028, HBM pricing power erodes and these margins compress. The company also pledged to detail shareholder return plans in the third quarter — buybacks and dividends at this scale would be a credibility signal, and the absence of one would leave the valuation unsupported by capital discipline.

The stock may need to find a bottom before an investor dives in. The pullback from $195 to $138 is still unfolding, and new-listing volatility doesn't follow a script. But the forward math is already far more attractive than the panic narrative suggests. A company growing revenue at 257% year over year, sitting on a net cash position of $45 billion, with $27.5 billion in trailing free cash flow and 58% HBM market share — that doesn't trade at 5.5x forward earnings because the business is broken. It trades at 5.5x because the market is looking at the wrong timeline.

The $38 billion capex isn't a problem for 2026 earnings. It's the company's answer to what happens after 2028. The market has confused the two.