Upgrading to Outperform: The Pricing Engine Held, the Contracts Got Sized, and ₩100 Trillion Is Going Back to Shareholders

Results vs. Consensus

Q3 2026 Scorecard

The company does not publish quarterly EPS in its release. The figure above is net income divided by roughly 735M weighted-average shares, and should be read as approximate. The two consensus sets we track agree closely: the alternative provider's figures were ₩98.34tn for revenue and ₩76.72tn for operating profit, which put the operating beat at 8.6%.

Year-Over-Year and Quarter-Over-Quarter (K-IFRS)

Margins in this table are computed from the unrounded figures. The company's release rounds them to 80%, 76% and 47%, and describes the sequential step as three points. Through nine months, 2026 revenue has reached ₩236.5tn, 2.4 times the full-year 2025 total of ₩97.1tn, and nine-month operating profit of ₩181.5tn is 3.8 times full-year 2025's ₩47.2tn.

Quality of the Quarter

Revenue: The 6.5% beat came from price holding up better than the Street feared, not from volume surprise. DRAM bits grew about 10%, exactly as guided in July, and NAND bits declined slightly as the company steered output toward 321-layer enterprise SSDs. DRAM ASP rose by a high-20% percentage and NAND ASP by mid-20%. One widely cited pre-print estimate had DRAM ASP up about 21% and NAND about 17%, which is where the revenue gap opened. The beat was achieved against a currency headwind: management cited unfavorable exchange-rate effects from the strong won on the operating line as well as below it.

Margins: Of every additional ₩100 of revenue this quarter, about ₩90 reached operating profit. Revenue rose ₩25.3tn while the cost base, revenue less operating profit, rose only ₩2.5tn, or 13%. That is the operating leverage of a supply-constrained memory maker selling a fixed bit budget into rising prices. A 79.7% operating margin is the highest this company has ever produced, and we have now called the margin peak twice, at 72% in April and at 76% in July, and been wrong both times. We address that directly in the Key Topics below.

EPS and net income: Net non-operating expense of ₩32.6tn took pre-tax profit down to ₩50.7tn from ₩83.3tn of operating profit, and net income landed at ₩39.8tn after roughly ₩10.9tn of tax. The two named items, an ₩11.6tn FX loss and ₩22.9tn of investment-asset valuation losses, sum to more than the total, so smaller items were a modest net gain. The FX loss reflects the won's appreciation during the quarter (it moved from roughly ₩1,439/$ at the Q2 print to roughly ₩1,334/$ at this one), which reduces the won value of the company's dollar-denominated assets. Neither item says anything about memory demand or pricing.

Segment Performance

Product-Line Detail: Q3 2026 vs. Q2 2026

DRAM: Price Growth Holds in the High Twenties

The sequence of DRAM price increases across 2026 now reads mid-60% in Q1, roughly 30% in Q2 and high-20% in Q3. In July the central question was whether Q2's halving was the first step of a slide toward flat and then negative pricing, or a mix effect from high-value shipments slipping into the second half. Q3 answers it, at least for now: the rate of increase held near Q2's level while HBM4 entered the mix and server DDR5 shipments rose more than 40% sequentially.

"With continued strength in conventional DRAM prices and an expanding sales mix of HBM4, which commands higher prices than the prior generation, DRAM ASP rose by high 20%, recording a high growth rate similar to the previous quarter." — Kim Woo-hyun, CFO

Both drivers matter. Conventional DRAM pricing stayed strong on its own, which addresses the concern we raised in July that management had not claimed conventional contract prices held their trajectory. And the HBM4 mix shift that management promised for the second half showed up on schedule.

Assessment: This is the result that most changes our view. A memory price cycle turns when the rate of increase keeps falling toward zero. In DRAM it barely fell. Prices cannot rise near 30% a quarter forever. But two quarters at that rate, plus multi-year contracts with price bands, is a different cycle from the one we downgraded in April.

HBM: HBM4 Becomes the Flagship, and 2027 Is Already Priced

HBM4 moved into full-scale mass production and shipment during the quarter, and management expects it to overtake HBM3E in both shipment volume and revenue in Q4. M15X is being pushed to full capacity by year-end to supply it. On yield, HBM4 has run ahead of plan, so the internal target was raised, and HBM3E has passed its planned mature yield. The more important disclosure was commercial.

"The company has finalized negotiations with major customers for next year, including HBM volume and prices. These negotiations faithfully reflect the increased product value of HBM4. As a result, we expect the profitability gap between HBM and conventional DRAM to narrow meaningfully." — Kim Woo-hyun, CFO

In Q&A, management said HBM profitability will improve meaningfully next year. So we read the narrowing gap as HBM margins rising toward conventional DRAM margins. Our inference is that conventional prices rose faster than contracted HBM prices during 2026, which opened the gap. In July management conceded that conventional pricing could influence the 2027 HBM talks. If it did, the result was higher HBM prices, not lower.

Assessment: The 2027 HBM book is settled at higher prices before the year starts, which removes the largest single pricing uncertainty in the model. Our pricing-power pillar moves back to on track. The open question shifts from price to share, which we take up under competition below.

NAND: Volume Traded for Mix, and the Pricing Leg Slowing Fastest

NAND bits declined slightly against a low-single-digit growth guide, a deliberate choice to prioritize 321-layer TLC enterprise SSDs over lower-value volume. ASP rose a mid-20% percentage on broad strength and a richer SSD mix, half the prior quarter's rate. The 321-layer process will pass half of domestic NAND capacity in Q4, and Q4 bits are guided up a mid-single-digit percentage.

"For NAND, although bit shipments decreased slightly compared to Q2 as we adjusted sales to focus on high value-added products such as 321-layer TLC enterprise SSDs, ASP rose by mid-20% due to overall price strength and an expanded SSD mix." — Kim Woo-hyun, CFO

The structural NAND story keeps improving. KV-cache storage for long-context inference is creating a new tier between DRAM and SSD, and the company is developing tiering software and a CXL-based hybrid memory to capture it, with high-bandwidth flash further out.

Assessment: NAND is where the deceleration we predicted is actually happening: ASP growth went from mid-70% to mid-50% to mid-20% in three quarters. It is the smaller and historically more volatile leg, and it is the line to watch for the first negative price print. It does not change the thesis while DRAM and HBM carry most of the earnings, but it is the canary.

Key Topics & Management Commentary

Overall management tone: Markedly more assertive than July, and the first call in our coverage on which the CEO spoke, opening with the Solidigm IPO decision and a shareholder-return principle before the CFO presented results. Where July's call was defensive on demand and closed with a disclosure refusal, this one led with numbers management had previously withheld: LTA coverage, a ₩100tn return figure and a settled 2027 HBM book. Management was least specific on 2027 CapEx and on how customer specification changes in HBM affect next year's bit demand.

1. A First ₩100 Trillion Quarter, and the Beat Returns

"Consequently, third quarter revenue increased by 32% quarter-on-quarter and 328% year-on-year, reaching KRW 104.6 trillion. This marks the first time quarterly revenues surpassed KRW 100 trillion, setting a new record high." — Kim Woo-hyun, CFO

The CEO's closing remark that the company has now set records for six consecutive quarters matches our own coverage history: every quarter since our Q2 2025 initiation has been a record on revenue and operating profit. The difference this time is that the record arrived above expectations rather than below them.

Assessment: In July the gap between best-ever and below-consensus was the whole story. This quarter it closed in the right direction by 6.5% on revenue and 8.3% on operating profit. A cyclical stock reprices on the trajectory of the surprise, and the surprise turned positive again.

2. The Margin Peak We Keep Calling, and Keep Missing

"Operating profit margin also improved by 3 percentage points quarter-on-quarter to reach 80%, setting new all-time quarterly records for both operating profit and operating profit margin." — Kim Woo-hyun, CFO

We downgraded in April partly on the argument that a 72% operating margin was a blow-off from which the only direction was compression. In July we dated the margin peak to Q2 at 76%. Q3 printed 79.7%.

Assessment: We have been wrong on the margin peak twice, and an upgrade is partly an admission of that. The mechanism we described was right in the narrow sense: the rate of price change peaked in Q1. We underestimated two things. When supply is short, prices can keep rising fast after the growth rate peaks. Cheaper older inventory and a richer mix also add margin on top. We are no longer willing to forecast the margin peak by quarter. We will instead watch for DRAM ASP growth turning negative.

3. Non-Operating Items Swing to a ₩32.6 Trillion Loss

"Regarding major non-operating items, foreign exchange related net loss of KRW 11.6 trillion due to exchange rates decline and valuation losses on investment assets of KRW 22.9 trillion were reflected, resulting in total net non-operating expenses of KRW 32.6 trillion." — Kim Woo-hyun, CFO

Pre-tax profit was ₩50.7tn and net income ₩39.8tn, a 38% net margin against an 80% operating margin. In Q2, net income exceeded operating profit by 55% on ₩63.3tn of investment-asset gains, chiefly from the Kioxia disposal. Neither the call nor the release named the holdings behind this quarter's ₩22.9tn valuation loss.

Assessment: One of our July signposts was that net income should fall back below operating profit once the Kioxia gain passed. It did, by more than half. We flagged the inflated prints as unrepresentative and we treat this deflated one the same way. The open item is the investment portfolio itself: a balance sheet that can swing ₩63tn up and ₩23tn down in consecutive quarters carries mark-to-market volatility that has nothing to do with memory.

4. LTAs: 14 Customers, Over Half of 2027–28 Capacity, Range Pricing

"While pricing structures vary by customer and product, they are generally designed to reflect market conditions within a price range agreed upon with the customer. Furthermore, to ensure that customers' purchase plans lead to actual orders and shipments, we have included safeguards such as deposits." — Kim Woo-hyun, CFO

The LTA customer count rose to 14 from around 10 in July, including major global cloud providers. Volume under contract now exceeds 50% of total capacity for both 2027 and 2028, management intends to keep coverage at a similar level over the medium term, and some customers are asking to extend terms beyond the typical five years.

Assessment: In July we called this the highest-value open question in the thesis and said the trough could not be sized without the coverage share. We now have it. More than half of two years' output is committed at prices that move within a negotiated band, with deposits enforcing the volume. The band's width is still undisclosed, so the floor is bounded rather than known. But the earnings base for 2027–28 is now substantially contracted, a structure we have not seen in prior memory cycles.

5. A ₩100 Trillion Shareholder Return, Including a Kioxia Special Dividend

"The company has decided to pursue shareholder return of approximately KRW 100 trillion through the 2026 fiscal year. We plan to execute cash dividends of KRW 30 trillion or more, and share acquisition cancellations of KRW 60 trillion or more, including the KRW 40 trillion announced in August." — Kim Woo-hyun, CFO

The August ₩40tn buyback was completed on October 21, about a month ahead of schedule, and the shares are to be cancelled shortly. For Q3 the board declared a ₩20,000 dividend per share (₩14tn), half regular and half a special dividend from the Kioxia gain, plus a new ₩20tn buyback. Half of that buyback, ₩10tn, will be cancelled. The other ₩10tn is for employee compensation and is excluded from the return total. With the ₩40tn program and the ₩24tn declared this quarter, about ₩64tn of the ₩100tn is done or declared, and the remainder falls into the fourth quarter and the year-end dividend.

Assessment: This resolves the capital-return deferral that had been our most frustrating watch item for two quarters, and at a scale beyond what we expected. A ₩100tn return on a market value near ₩1,000tn is roughly a 10% yield in one year. The policy also commits to quarterly buybacks and to more than 50% of free cash flow. The ₩10tn employee-compensation tranche is a real cost to shareholders and deserves noting, but it is disclosed and excluded from the headline.

6. Solidigm Ends Its IPO Review

"However, after carefully considering recent financial market volatility and the correction in semiconductor stocks, Solidigm has concluded that an IPO would not fully deliver the anticipated benefits at this time and has therefore decided to discontinue its review of potential pre-IPO and IPO options." — Kwak Noh-Jung, President and CEO

The CEO opened the call to explain the decision. The IPO was meant to fund Solidigm's manufacturing expansion with outside capital and so ease the parent's investment burden. Management said future financing will be pursued in a way that serves both companies' shareholders, without specifying the form.

Assessment: Selling a stake in the NAND subsidiary into a falling semiconductor tape would have crystallized value at a poor moment, so ending the review is the right call. The cost is that Solidigm's capacity expansion now draws on parent cash or structured financing that has not been described.

7. HBM Profitability Set to Rise in 2027

"Next year, share of HBM4 and its price will increase while mass production yield and cost competitiveness will further improve, making it likely that HBM revenue and profit grow significantly year-over-year, with meaningful improvement in profitability as well." — Kim Kitae, Head of HBM Sales and Marketing

Management added that customers' HBM volume requirements are more than double what the company can supply today, and that HBM profitability should stay far above past-cycle peaks.

Assessment: In July the risk was that HBM's premium would erode as conventional DRAM pricing pulled HBM negotiations around. The 2027 negotiations closed with HBM4's higher value reflected in price, and management is guiding to higher HBM profitability. HBM pricing for 2027 is no longer an open question.

8. Demand: Supply Cannot Keep Pace Through 2030

"Despite the high level of demand, considering actual supply conditions, achievable demand growth for 2027 is expected to be at a similar level to this year, with growth rates of mid-20% for DRAM and high teen percent for NAND." — Kim Woo-hyun, CFO

Management cited external estimates of more than 60% growth in global cloud-provider CapEx in 2027 and more than 100% growth in secured power capacity, and expects demand to exceed supply through 2030. PC and mobile shipments are expected to decline on memory cost and availability, more than offset by AI infrastructure.

Assessment: The framing that matters is "achievable" demand: management is guiding 2027 growth to what supply allows, not to what customers want. That is a conservative way to frame a demand outlook, and it means the 2027 growth rate is a supply forecast. The cost is the PC and mobile base, which is shrinking under price. That latent demand returns at lower prices when supply loosens, which supports volume rather than margin later in the cycle.

9. 2027 CapEx Will “Increase Significantly,” Still Unquantified

"We expect Capital Expenditure in 2027 to increase significantly compared to 2026 as we invest in equipment to expand the share of leading-edge processes and ramp up our first fab in Yongin while domestic and overseas infrastructure investments accelerate." — Kim Woo-hyun, CFO

In Q&A the CFO described the build as staged: building shells for Yongin and M17 is being accelerated, while cleanrooms will open in phases and equipment move-in and ramp will follow demand. Management also argued that more fabs no longer mean proportionally more bits, because HBM and leading-edge nodes consume more wafer area per bit.

Assessment: This remains the bear case's best mechanism, and the market is already pricing a 2029 oversupply narrative. Two things mitigate it. More than half of 2027–28 output is now contracted, and the staging leaves equipment, the expensive and reversible part, gated on demand. We would like a number. Without one, CapEx stays an emerging risk rather than a resolved one.

10. The Strong Won: A Headwind Above and Below the Line

"Despite unfavorable exchange rate effects due to the strong Korean won, product price increases and the expanded sales of high value-added products led to an operating profit of KRW 83.3 trillion for the third quarter." — Kim Woo-hyun, CFO

The won strengthened about 7% against the dollar between the Q2 and Q3 prints. Memory is priced in dollars, so a stronger won lowers reported won revenue for a given dollar price, and it produced the ₩11.6tn FX loss below the line.

Assessment: Currency had been a tailwind through most of this cycle and is now a headwind. It does not change unit economics, but it does mean that a portion of the won-denominated growth rate the market has been extrapolating was currency. A further leg of won strength would trim reported growth without any change in the memory market.

11. Competition Arrives From Two Directions

"To apply a new cell structure into the products, we must secure not only performance, but also yield, reliability, and cost competitiveness. Not only that, the product design and customer qualification must also be completed. This is why the timing of technological development and the timing of commercialization are distinct from each other." — Park Joondeok, Head of DRAM Marketing

Two developments between the prints bear on the competitive picture. On October 13 the shares fell 10.1% after China's CXMT said it had made progress on a denser 4F² DRAM cell design. Earlier in October, Samsung's 12-high HBM4E was reported to have passed qualification at the leading accelerator customer and at major cloud providers, though Samsung did not confirm it. Management answered the DRAM-architecture question by describing a staged path that extends the current cell design before moving to 4F² or 3D DRAM, and framed its HBM position as first supplier to key customers.

"As the first supplier to our key customers, we will keep growing our HBM business, drawing on proven product competitiveness and reliable mass production capability." — Kim Kitae, Head of HBM Sales and Marketing

Assessment: The competitive threat is real but slower than the share price suggested in mid-October. CXMT's claim is unverified and aimed at server shipments by year-end, and the gap between a cell-design announcement and qualified volume at scale is measured in years. On HBM, a competitor qualifying is no longer news. The relevant test is whether SK hynix loses share or price at its lead customer, and the settled 2027 HBM book says it has not, at least for next year.

Guidance & Outlook

SK hynix guides bit shipments and frames pricing qualitatively. It gives no revenue or margin guidance.

"Consequently, fourth quarter shipment volume is expected to grow meaningfully by mid 20% sequentially." — Kim Woo-hyun, CFO

Implied ramp: A mid-20% DRAM bit guide for Q4 is more than double the roughly 10% delivered in Q3, and it comes from new capacity (M15X at full rate by year-end) rather than inventory. Second-half DRAM bit growth will far exceed the first half's, when Q1 bits were roughly flat and Q2 grew a high single digit. That was one of the July promises we said we would grade, and it is being kept.

Street at: Pre-print consensus had Q4 revenue near ₩124tn, about 19% above Q3, and Q4 operating profit near ₩87tn. On our rough split of revenue between DRAM and NAND, the bit guide alone gets close to that revenue figure at flat prices. Any positive price growth in Q4 would put the Street's numbers at risk of upward revision.

Guidance style: Conservative and reliable on DRAM, where the company has delivered its bit guide in each of the last three quarters, and more discretionary on NAND, where it under-shipped this quarter by choice. Management's habit of guiding demand to "achievable" supply rather than to customer requests means the volume guide is a capacity statement, which is the most credible kind.

Analyst Q&A Highlights

The call was interpreted consecutively from Korean. Questions and answers below are quoted from the interpreted English rendering and attributed to the analyst and to the named executive who answered.

Is AI Demand Durable If Monetization Lags?

For the second quarter running, the first question challenged the demand behind memory pricing rather than the pricing itself. Management answered with customer cloud-revenue growth and backlog data, and with a structural argument that the buyer base has broadened beyond a handful of hyperscalers.

Q: "We see that the AI investment continues to grow, but then it seems as if the monetization following the AI demand or the AI infrastructure build-out is slow to be realized. As a result, there are some concerns in the market over the sustained AI demand. Against this backdrop, my question is, what is the company's basis for believing that the AI investment momentum will continue?"
— Peter Lee, Citigroup

A: "Early AI investment was concentrated among the big tech, so a single company adjusting its investment could directly result in overall demand volatility. But today, the investor base is rapidly broadening, also including non-tech companies investing in their own AI infrastructure, Sovereign AI, or on-prem demand."
— Park Joondeok, Head of DRAM Marketing

Assessment: The diversification argument is the right one and it is better supported than in July, because more than half of 2027–28 capacity is now under contract with 14 customers. It does not address the market's actual fear, which is valuation multiples on AI infrastructure rather than memory volumes. Two days after this call a major cloud provider's results sent the stock down 7.1% without any change in memory fundamentals.

How the Return Splits Between Dividends and Buybacks

An analyst asked how the board set the proportion between cash dividends and share repurchases. Management described a sustainable dividend anchored to earnings and peer yields, special dividends for investment gains, and buybacks for the remainder because it considers the stock undervalued.

Q: "What is the basis for the company to determine the proportion between the dividend payout and share acquisition out of the total shareholder return?"
— Sun-woo Kim, Meritz Securities

A: "After the cash dividend, the rest of the total shareholder return is allocated to share acquisition. We believe our share price today is undervalued, not adequately reflecting our earnings and cash flow."
— Kim Woo-hyun, CFO

Assessment: An explicit statement from the CFO that the shares are undervalued, backed by ₩60tn or more of cancellations this year, is the strongest signal management can send. The company completed the August program at an average near ₩1.69M per share, about 19% above the October 30 close, which suggests it will keep buying at these levels.

What 50% LTA Coverage Means for Growth

An analyst asked what the contracted share of 2027 and 2028 capacity implies for performance and for incremental growth. The answer went further than the prepared remarks.

Q: "It was explained earlier that currently the volume secured through the LTAs is about 50% of the volume of the total capacity for 2027 and 2028. What would it mean for the company's performance and potential for increment growth in the longer term?"
— Giuni Lee, Goldman Sachs

A: "The volume already secured under LTAs and the demand currently being discussed with customers taken together means that our DRAM and HBM volume for the next three years through 2029 are effectively sold out."
— Park Joondeok, Head of DRAM Marketing

Assessment: "Effectively sold out" through 2029 combines signed contracts with negotiations in progress, so it is a stronger claim than the LTA figure alone and should be discounted accordingly. The signed portion, more than half of capacity, is the part we underwrite. Management's plan to keep the remaining capacity flexible for spot pricing and next-generation products is sensible: it keeps upside exposure while the contracts protect the downside.

What Happens If Demand Falls Short of the Build

An analyst raised the overproduction risk directly, given the Yongin and other fab investments, and asked for the contingency plan. Management described a staged investment model in which shells are built ahead of demand and equipment follows it.

Q: "What if demand were to fall short of expectations, then what would be the company's strategy or the management's plan against it?"
— Simon Woo, Bank of America

A: "Our investment will be executed in stages. The foundational infrastructure will be readied in advance for medium to long term demand, but the key resources needed to increase supply, such as equipment moving and mass production ramp-up, will be adjusted in line with demand. For example, for new fabs, including the Yongin cluster and M17, we plan to accelerate construction of the shell structure, but clean rooms will be opened in phases."
— Kim Woo-hyun, CFO

Assessment: This is the most credible version of the capacity answer we have heard from this company, because it separates the irreversible spend (shells) from the bit-adding spend (tools). It remains a stated intention. The cycle test is whether the company actually slows tool move-in when spot pricing softens, and that test has not arrived.

HBM4 Supply, HBM4E and Custom HBM

An analyst asked for HBM4 supply specifics and the roadmap beyond it. Management said HBM4 production is running without disruption, that yields have exceeded plan at both the front end and the full back-end process, and that HBM4E qualification is tracking customers' timelines, with custom HBM projects under way.

Q: "The HBM4 supply was mentioned earlier, but I would like to get an update from the company on the specifics of the HBM4 supply. Can you also share your plan for the follow-up, the HBM4E and custom HBM?"
— Jay Kwon, JPMorgan

A: "On this basis, HBM4 already exceeded its planned yield, so the target was revised up. HBM3E has also surpassed the planned mature yield."
— Kim Kitae, Head of HBM Sales and Marketing

Assessment: Yield above plan on a first-generation HBM4 ramp is the execution evidence behind the HBM leadership pillar. Specifying that the yield refers to the entire back-end process, not individual steps, reads as a deliberate contrast with competitors' yield claims.

Shorter HBM Stacks

An analyst asked about reports that customers may move to shorter HBM stacks and what that would mean for shipments and profitability. Management framed the change as evidence of a supply shortage and argued that higher capacity per die offsets much of the impact.

Q: "Thank you for taking my question, which is also on the HBM, particularly on the possibility of lowering the HBM stacks. What is the company's view on this possibility of having shorter HBM stack? What would be the potential impact on the HBM business, including HBM shipment and profitability?"
— Dong-hee Han, SK Securities

A: "Having said that, lower stack does not necessarily mean a proportionate decrease in HBM capacity. Even if it goes from 12-stack HBM4 to eight-stack HBM4E, the increase in capacity per die can offset a substantial part of the impact."
— Kim Kitae, Head of HBM Sales and Marketing

Assessment: "A substantial part" is not all. A move from 12-high to 8-high stacks means fewer DRAM dies per accelerator, partly offset by denser dies, so bit demand per GPU likely falls somewhat even as cube count rises. Management says it has already built this into its supply plan. We treat it as a modest headwind to 2027–28 HBM bit demand that the over-2x demand-to-supply gap can absorb.

The 4F² Question

An analyst noted that competitors are developing new DRAM cell structures and asked about the company's 4F² preparation. Management said it will extend its current cell design through EUV-based scaling before moving to 4F² or 3D DRAM, and stressed that development timing and commercialization timing differ.

Q: "We see that other companies are also working on new cell structures and next-generation DRAM technology, which is heightening interest in long-term technology competition."
— Young-gun Kim, Mirae Asset Securities

A: "Each technology will be applied to the right product at the right time. The right time, meaning when it is ready technologically and economically."
— Park Joondeok, Head of DRAM Marketing

Assessment: The question was prompted by the CXMT announcement that knocked 10% off the shares in one session, and management chose not to engage with it by name. The answer is reasonable engineering logic and gives no date, which is what an incumbent says when it does not want to be seen racing. If a competitor ships 4F² server DRAM in volume during 2027, this answer will look slow.

KV Cache and the NAND Opportunity

An analyst asked how KV-cache growth for AI inference will affect NAND demand. Management described a tiered memory architecture in which SSDs hold and share inference context, and outlined software and CXL products to capture it.

Q: "Now, we see these days that the size of KV cache is growing rapidly, and now there is also work underway to store and reuse it across multiple memory tiers. How does the company expect this to affect future NAND demand, and what type of preparation is being made?"
— Young-ho Ryu, NH Investment & Securities

A: "As agentic AI and long context use cases spread, the size of KV cache is growing rapidly, making it difficult to store it all in HBM or DRAM due to capacity and cost constraints. As a result, a tiered memory architecture is becoming increasingly important, which distributes data across HBM, DRAM and SSD according to access frequency."
— Song Chang-seok, Head of NAND Marketing

Assessment: This is the clearest articulation yet of why NAND has its own AI driver rather than an AI-adjacent pricing tailwind. It is a 2027–28 story, though. Near term, NAND ASP growth is slowing fastest of any product line, and the architecture shift does not change that.

What They're NOT Saying

1. The width of the LTA price range. Contracts price "within a price range agreed upon with the customer," but neither the floor nor the ceiling was described. The floor is the single number that determines how much downside protection the contracted half of capacity actually provides.
2. A 2027 CapEx figure, or a refreshed 2026 figure. 2027 will "increase significantly." The high-₩40tn 2026 figure from July was not repeated or updated on the call or in the release.
3. Which holdings produced the ₩22.9tn valuation loss. A loss of this size on investment assets deserved identification. Its absence leaves the volatility of the portfolio unexplained.
4. HBM share at the lead customer. Management described itself as "first supplier to our key customers" but gave no share figure, and did not address the reported qualification of a competitor's 12-high HBM4E.
5. The bit-demand impact of customer specification changes. Management acknowledged customer moves to adjust HBM specifications but quantified neither the stack-height shift nor any bandwidth changes in terms of 2027 HBM bit demand.
6. How Solidigm's expansion will now be funded. The IPO review ended with no replacement plan beyond a commitment to serve both companies' shareholders.
7. HBM's share of revenue or margin. Still undisclosed. Management's guidance that the HBM-versus-conventional profitability gap narrows cannot be checked from outside without it.

Market Reaction

All price levels are KRX closing prices verified against end-of-day data; the benchmark is the KOSPI. SK hynix discloses before the Korean market open, so the reaction is the print-day session. Two further sessions have closed since the print and are included.

• Between the prints: From ₩1,322,000 the day after the Q2 print, the shares rallied 41.3% to a ₩1,868,000 close on September 21, helped by the ₩40tn buyback announcement in August. They then fell 34.5% to ₩1,223,000 on October 20 and again on October 22, which is exactly our April downgrade level. The low came after a 10.1% drop on October 13 when China's CXMT announced its 4F² progress, and an 11.1% drop on October 15.
• Pre-print setup: ₩1,341,000 close on October 27 (~$1,006), −24.2% over the trailing 30 days, +106.0% year-to-date and +157.4% over twelve months. The KOSPI was +33.8% year-to-date.
• Reaction (October 28): Opened +3.7% at ₩1,390,000 and closed at ₩1,487,000, +10.9% (~$1,115), near the session high of ₩1,495,000. The KOSPI rose 6.6% and recovered 6,000 for the first time in ten sessions, with SK hynix a large part of that move. Samsung Electronics rose 9.0%.
• Volume: 4.2M shares against a 3.6M 30-day average, 1.2x, a far lower-conviction session than the 2.0x selloff at the Q2 print.
• Next two sessions: ₩1,526,000 on October 29 (+2.6%, KOSPI −1.2%), then ₩1,418,000 on October 30 (−7.1%) as Alphabet's results revived AI-spending fears and the KOSPI fell 3.5%, triggering a sell-side sidecar.
• Net: +5.7% from the pre-print close against +1.6% for the KOSPI. Since our April downgrade at ₩1,223,000 the stock is +15.9% while the KOSPI is −10.7%. Since our Q2 report it is +7.3% against +2.5%.

The print-day move was a straightforward repricing of a clean beat plus a capital return that equals about a tenth of the company's value. What matters more is the context. The stock entered the print down 24% in a month on two narratives: a Chinese competitor's DRAM architecture claim and a 2029 oversupply thesis. It rose 10.9% when the company answered both with contracts and cash. Two sessions later it fell 7.1% on a US cloud company's results, which said nothing new about memory, leaving it 5.7% above the pre-print close.

That last move is the clearest statement of where the stock now trades: on the market's appetite for AI infrastructure risk, not on SK hynix's own earnings. That cuts both ways. It means the shares will stay volatile and correlated with AI-capex sentiment regardless of execution. It also means the stock trades on other companies' AI spending rather than on its own results, even though those results now include contracts for over half of 2027–28 output and a roughly 5% share retirement this year.

Street Perspective

Debate: Is 80% the Peak, or Do the Contracts Make It a Plateau?

Bull view: Four consecutive margin records, a DRAM pricing engine that stabilized at a high-20% rate, and more than half of 2027–28 capacity contracted within price bands mean this cycle will not mean-revert like 2018 or 2022. Earnings plateau at a high level rather than collapsing, and the stock deserves a multiple on durable earnings rather than a peak-cycle discount.

Bear view: Every memory cycle has had a reason it was different, and an 80% operating margin is the highest this company has ever printed. Contracts signed at the top get renegotiated at the bottom, the price bands are undisclosed, and NAND pricing is already decelerating sharply. Industry capacity additions in 2027–28 will meet AI demand that is growing more slowly than today.

Our take: The bear case is right that 80% is not a durable margin, and we do not underwrite it. The bull case is right that the contract structure changes the shape of the downturn. We do not need a plateau at 80% for the stock to work. We need normalized operating profit well above the ₩80–100tn we assumed in July, and contracted volume at banded prices on more than half of capacity makes that likely.

Debate: Do Samsung's HBM4E and Shorter Stacks Erode the HBM Lead?

Bull view: SK hynix's 2027 HBM volume and pricing are already settled with major customers, HBM4 yields are running above plan, and customers want more than twice what the company can supply. A competitor qualifying is necessary for customers to meet demand, not a threat to the leader's allocation. Shorter stacks are a symptom of shortage.

Bear view: HBM's premium rested on SK hynix being the only reliable supplier. With Samsung reportedly qualified on 12-high HBM4E and the lead customer reportedly redesigning toward 8-high stacks and lower bandwidth, the technology bar is falling to where more suppliers can clear it. Share and premium compress from 2028.

Our take: The bear case is a 2028 argument, and the 2027 book is closed. We moved HBM competition to an active risk in July and it stays there. A competitor qualifying is not a downgrade trigger, and one had already qualified when we wrote our July trigger. We would downgrade on lost share or reopened pricing at the lead customer.

Debate: Is a ₩100 Trillion Return Enough to Re-Rate a Cyclical?

Bull view: Returning roughly 10% of market value in a single year, with a stated policy of quarterly buybacks and special dividends from investment gains, transforms the equity story from a cyclical price bet into a cash-yield compounder. The company is telling the market its shares are undervalued and spending ₩60tn or more to prove it.

Bear view: Peak-cycle buybacks are a classic capital-allocation mistake, and the company bought ₩40tn of stock at an average well above today's price. A 10% yield on peak earnings becomes a much lower yield on trough earnings, and the employee-compensation tranche dilutes the cancellation benefit.

Our take: The bear framing of peak-cycle buybacks would be decisive if earnings were about to fall sharply. With more than half of 2027–28 output contracted and net cash of ₩98.7tn after the August program began, the return is funded from contracted cash flow rather than borrowed against the cycle. We treat it as a meaningful floor on the share price, though not a reason on its own to own the stock.

Model Update Needed

Valuation framework (KRW). At the October 30 close of ₩1,418,000 on roughly 705M shares outstanding, market value is about ₩1,000tn (~$750B) and enterprise value about ₩901tn net of September 30 net cash. Trailing four-quarter operating profit is ₩200.7tn, so the stock trades at about 4.5x trailing operating profit, against 7.0x at the equivalent point after our July report. On nine-month actuals plus pre-print Q4 consensus, FY26 operating profit is about ₩268.8tn and the multiple is about 3.4x. The ₩14tn dividend and the October share purchases will lower net cash before fourth-quarter cash flow rebuilds it. That raises these multiples by about 0.1x to 0.2x.

Why the peak-earnings trap no longer binds. In July we argued that a single-digit multiple of peak earnings was a trap, and we anchored instead to normalized operating profit of ₩80–100tn. We now raise that to ₩130–160tn, roughly half of FY26's run-rate and under half of Q3 annualized. The reasons are the LTA coverage disclosed this quarter, the settled 2027 HBM book, a 2027 DRAM demand outlook of mid-20% growth, and a price level that would still sit well above late-2025 levels after a substantial correction. On that range the current enterprise value is 5.6x to 6.9x normalized operating profit. In July we said we would buy at ₩850,000–950,000. That price valued the business at 5.5x to 7.8x our old normalized profit. Today's price sits inside that range on our new estimate. The earnings base moved up; the price did not follow.

Fair-value range: base ₩1.6M to ₩1.9M (normalized operating profit of ₩130–160tn at 8x, plus about ₩85tn of net cash: the September 30 balance less the declared dividend, before the October share purchases, which would trim the range by roughly 1%), or 13% to 34% above the October 30 close. Bull case ~₩2.5M (+76%) if the LTA floors hold through a correction and normalized profit sits nearer ₩200tn at a modestly higher multiple. Bear case ~₩900,000 (roughly −37%) if pricing returns toward late-2025 levels as 2027–28 capacity arrives and normalized profit falls to ₩70–80tn. Weighting the scenarios 25/50/25 gives about ₩1.73M, roughly 22% upside, with the buyback adding support beneath the base case.

Thesis Scorecard: 2026_Q2 Signposts Revisited

Our July report set eight signposts for this quarter. Five graded bullish, two mixed, and one turned out to have been mis-specified when we wrote it.

Scorecard summary: The quarter graded almost exactly opposite to the second quarter. The variable we called decisive, blended pricing, held its pace in DRAM, where the mix-and-timing explanation applied, though NAND kept slowing. The two structural items we said we needed, LTA coverage and a concrete capital return, both arrived, and at larger scale than we expected. One signpost needs correcting. Our July Underperform trigger named a competitor HBM4 qualification at a major accelerator customer. That condition had already been met in June, when the leading accelerator customer publicly certified all three suppliers for its next platform. A trigger that had already fired when we wrote it was not a useful signal. We replace it with a test of share and pricing at the lead customer, and that test passed for 2027.

Pillar status changes: pricing power moves from at risk back to on track. The LTA architecture moves from emerging and unproven to on track: sized but untested through a downturn. The capital-return deferral risk is retired. The price-normalization-as-mix risk moves from active to contained. HBM competition stays active, AI-capex multiple compression stays materializing, and we add Chinese DRAM competition as a new emerging risk.

Overall: Thesis strengthened materially. The standing view since April has been a flawless franchise at a cyclical peak. The franchise half has never been in doubt. The peak half has now been wrong on margin for two quarters, and the evidence this quarter, contracted volume at banded prices and stabilized DRAM pricing, says the cycle is longer and its downturn shallower than we assumed.

Action: Buy. Upgrade to Outperform, and add on further AI-sentiment drawdowns while the operating evidence holds.

Bottom Line: Upgrading to Outperform

Rating decision: We upgrade SK hynix to Outperform from Hold. We initiated at Outperform in July 2025 at ₩268,500, maintained it through the Q3 and Q4 2025 records, downgraded to Hold in April 2026 at ₩1,223,000 on cycle position and price, and maintained Hold in July at ₩1,322,000 after the first double miss.

An honest accounting of the Hold. The April downgrade was early: the stock rose 139% to its June peak before round-tripping. Since then it has done what a Hold should do, rising 15.9% while the KOSPI fell 10.7%, and touching our downgrade price again on October 20 and 22. The reasoning behind it has aged less well than the price call. We said 72% was a blow-off and then dated the peak to 76%. Q3 printed 79.7%. We underestimated how long a supply-constrained market can sustain high price growth once contracts begin to replace spot exposure. The lesson we carry forward is to watch for negative price growth rather than predict a margin peak by quarter.

Why upgrade now. Our July note listed three routes back to Outperform: a pullback to ₩850,000–950,000, Q3 pricing showing the Q2 undershoot was not the turn, or hard evidence that the deposit-bearing LTAs hold pricing through a real decline. The second is met for DRAM, where the July explanation applied: DRAM pricing held at a high-20% rate, and the deferred high-value shipments arrived. NAND pricing kept slowing, and we treat it as the smaller leg. The third is half met: the contracts are now sized at more than half of 2027–28 capacity with range pricing, though no decline has yet tested them. On top of that the company announced roughly ₩100tn of returns and settled its 2027 HBM book at higher prices. Meanwhile the stock is only 16% above where we downgraded it, while trailing operating profit has grown 2.6x.

Why not wait for the price. The ₩850,000–950,000 zone was a multiple expressed as a price, calibrated to normalized earnings of ₩80–100tn. The disclosures this quarter justify a higher normalized base, and at ₩1,418,000 the stock already trades within the multiple range that zone implied. The price zone stood in for a multiple. Earnings have risen, so waiting for the old price would follow the rule but miss its purpose. We said in July that we would rather be early to upgrade than clever about the bottom, and we hold to that.

What would move us back to Hold: DRAM blended ASP growth turning negative with operating margin falling below 70% before the LTA floors are visible; evidence that LTA customers are renegotiating; a reported HBM share loss or reopened 2027 pricing at the lead accelerator customer; a 2027 CapEx figure implying bit capacity growth above the mid-20% demand outlook; or the price reaching the top of our base range without new evidence.

What would move us to Underperform: LTA price floors proving too low to protect margins in a genuine downturn, operating margin falling below 50%, or a combination of competitor HBM share gains and Chinese DRAM capacity reaching volume while AI infrastructure spending is visibly cut.

Signposts for Q4 2026 earnings (late January 2027):