SK hynix Posts Record 76% Operating Margin as AI Memory Shortage Deepens, Discloses Multiyear Supply Deals With Roughly 10 Customers
Q2 2026 earnings call, July 28, 2026
SK hynix delivered the strongest quarter in its history and, in the process, offered investors a rare level of detail on how it is locking in years of AI-driven memory demand through long-term agreements, an accelerated capacity build-out, and a landmark Nasdaq ADR listing. Revenue hit KRW 79.3 trillion, up 51% quarter-on-quarter and 257% year-on-year, while operating income reached KRW 60.5 trillion for a 76% operating margin, a level President Song Hyunjong called an all-time high for both metrics. EBITDA margin came in at 81%. Net profit was KRW 93.9 trillion, inflated by KRW 63.3 trillion in gains from asset sales and revaluations, including the Kioxia stake disposal, pushing the net margin to an eye-popping 118%.
Cash and equivalents jumped KRW 33.6 trillion sequentially to KRW 88 trillion, and net cash swelled to KRW 69.4 trillion with a debt-to-equity ratio of just 7%. That balance sheet strength is now central to the equity story: management said it is "currently reviewing various additional execution measures for shareholder returns" and plans to communicate specifics "within the year," though it declined to give size, format or timing given ongoing ADR-related disclosure restrictions.
Long-Term Agreements Emerge as the New Industry Architecture
The most consequential disclosure was the scale and structure of SK hynix's long-term agreements, or LTAs, with customers. Management confirmed it has "concluded LTA negotiations with around 10 customers, including our key customers," with more discussions ongoing across the industry's major players. Contract terms run roughly five years, with pricing mechanisms tailored by customer and product rather than a uniform formula, and critically, the deals include deposit mechanisms designed to "support contract fulfillment and enhance the visibility and reliability of customers' mid- to long-term demand plans." Management would not disclose what share of total sales these contracts will eventually cover, but said the goal is to use LTAs to build downside resilience into earnings while preserving flexibility to capture spot-market upside when conditions are favorable. This represents a structural shift for a historically cyclical, spot-price-driven industry, and suggests the memory majors are increasingly behaving like foundries with take-or-pay commitments rather than commodity producers.
Capex Raised as Capacity Pulled Forward
SK hynix now expects 2026 capital expenditure to land in the "high KRW 40 trillion range," an upward revision driven by a decision to pull forward the mass production schedule for its M15X fab and accelerate capacity expansion following the Yongin Fab 1 clean room opening in early 2027. The company has also announced new investments in P&T7, an advanced packaging facility, M17, a new NAND fab, and a longer-horizon domestic semiconductor cluster intended to serve demand beyond Yongin. Management was careful to frame this as demand-led rather than speculative, noting that capacity build-out "will be pursued in stages comprehensively considering customer demand visibility, investment efficiency." Asked directly about oversupply risk given the scale of announced investment, one executive pushed back firmly, arguing that "given that our capacity expansion will be executed flexibly in alignment with confirmed customer demand, we do not believe our medium- to long-term investment plans will lead to oversupply right away."
HBM4 Ramping, HBM4E Samples Out, and a New Cooling Technology in the Wings
HBM4 mass production shipments began in the second quarter, with a "full-fledged ramp-up" planned for the second half. Management said yield and quality for HBM4 are "nearing the levels of HBM3E," which is already in its maturity stage, an important signal given persistent market chatter about competitors closing the technology gap. HBM4E samples have already been delivered to a major customer, with volume production targeted for 2027. Looking further out, SK hynix disclosed it is developing "iHBM" technology for HBM5 and beyond, which integrates cooling elements directly into the package and is expected to reduce thermal resistance by more than 30%, a meaningful engineering claim for an industry increasingly bottlenecked by power and thermal constraints in AI accelerators. On pricing, when pressed by UBS's Nicolas Gaudois on 2027 HBM contract negotiations, management acknowledged that surging conventional DRAM prices "may also have some influence" on HBM pricing discussions, but insisted HBM pricing reflects the "resources, opportunity costs, technical complexity and value" the product delivers rather than tracking commodity DRAM mechanically.
Demand Narrative Shifts From Training to Inference and Agentic Workloads
Management pushed back explicitly on recent investor concerns that data center leasing arrangements and more efficient AI models could signal an infrastructure slowdown, characterizing these developments instead as "a transition towards higher utilization of the AI infrastructure that has already been built at scale." The company argued that model efficiency gains historically broaden rather than shrink the addressable market, pointing to "explosive demand for the recent high-efficiency AI models" as evidence that lower costs expand adoption rather than reduce infrastructure needs. On the product side, this is showing up as demand for server DRAM to support "agentic" AI workloads and a rapidly growing role for enterprise SSDs, which management said are becoming "a core component of the AI memory hierarchy" as the industry shifts from training-centric to inference-centric compute. Enterprise SSD revenue doubled quarter-on-quarter, and Solidigm's 30-terabyte-plus high-capacity SSD revenue more than tripled. The company said it is deliberately avoiding a single-technology NAND strategy, instead building out QLC, TLC and SLC-mode SSD product lines to match workloads ranging from HDD replacement to KV cache offloading near GPUs.
DRAM ASP Miss Explained as Mix Shift, Not Demand Weakness
Second-quarter DRAM ASP growth of roughly 30% came in below some market expectations, which management attributed to a deliberate push of high-value product shipments into the second half rather than any softening in underlying demand. Bit shipment growth guidance for the third quarter was set at approximately 10% quarter-on-quarter for DRAM and low single digits for NAND, with management expecting second-half bit growth to exceed the first half as HBM4 ramps and 1c-nanometer conventional DRAM shipments increase, both of which should also lift blended ASP.
Nasdaq ADR Listing and the Mechanics Investors Should Watch
SK hynix completed its ADR listing on Nasdaq on July 10, which the company said was "the largest ever for a foreign company IPO in the U.S." Management provided specific mechanical detail relevant to arbitrage and liquidity dynamics: ADRs can be freely converted into Korean shares starting July 30, but converting shares into ADRs may face delays of several weeks due to regulatory filing requirements, and the total ADR pool is capped at 17,790,000 shares, equivalent to the original offering size. Any increase in that conversion limit remains undecided and subject to regulatory review, meaning the two-way fungibility of the ADR is currently asymmetric, a technical detail that could matter for cross-listing valuation gaps in the near term.