Record Quarter, Clean Print, Borrowed Demand: An All-Time-High ₩22.2tn at a 41% Operating Margin with HBM on Track to Double, and the Only Asterisk Is Tariff Pull-Forward — Initiating at Outperform

Results vs. Consensus

Q2 2025 Scorecard

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

Quality of Beat

Revenue: The +6.9% beat versus consensus is high quality and broad. Both DRAM and NAND shipments came in above the company's own guidance, and the YoY growth rate of +35% sits on a 2Q24 base that was already in recovery. The composition has one caveat addressed below (some of the volume was pulled forward), but the headline magnitude, a first-ever quarter at this revenue level, is not in doubt.

Margins: The 41% operating margin slipped one point QoQ, and the reason is instructive rather than alarming: the revenue mix tilted toward lower-priced conventional DRAM (server and PC strength) and NAND solution ASP fell high-single-digit, so the blended margin gave back a point even as absolute operating profit set a record. This is the normal arithmetic of a quarter where volume and conventional product carried the upside while HBM held the high-margin floor.

EPS / net income: The EPS beat (+5.5%) is clean and operating-driven. The net-income decline is purely a below-the-line swing: net non-operating moved from a gain in 1Q25 to a ₩0.49tn net loss in 2Q25, dominated by a ₩0.61tn FX loss as the won appreciated. Pre-tax income was ₩8.7tn. We would not let the net-income optics distract from a record operating quarter.

Segment Performance

Product-Line Detail — Q2 2025

DRAM — Volume-Led, with HBM Holding the Margin Floor

DRAM shipments grew near 20% QoQ, well above guidance, on full-scale 12-high HBM3E and strong server and PC demand. ASP rose only low-single-digit because the upside skewed toward relatively lower-priced conventional DRAM. The strategic picture is two-track: HBM and high-density server modules (128GB and above) hold the high-margin core, while conventional DDR5 (and even legacy DDR4, where prices spiked on supplier end-of-life decisions) provided the volume. Management is also building out the conventional portfolio: high-speed DDR5 above 8,000 Mbps, LPDDR5X for flagships, LPDDR4X for China, LPDDR-based server modules within the year, and 16Gb plus new 24Gb GDDR7 for GPUs.

Assessment: A healthy, volume-led DRAM quarter with the high-margin HBM core intact. The legacy DDR4 price spike is a near-term tailwind but management correctly characterized it as a temporary transition-phase squeeze (suppliers exiting DDR4), not a structural shift. DDR4 has fallen from a double-digit to a single-digit revenue share and is being phased out for the mass market.

NAND — A 70% Shipment Surge Off a Low Base, Pricing Still Soft

NAND shipments jumped over 70% QoQ off a depressed 1Q base, with demand across all applications: enterprise SSD as hyperscalers extended AI investment, pull-forward in components on tariff uncertainty, and a China mobile set-build bump on promotions. ASP still fell high-single-digit as solution pricing stayed weak. Management was clear-eyed that the NAND market's structural AI-demand inflection has not yet arrived, but pointed to enterprise SSD (offloading AI inference cache, 120TB-plus QLC, 321-layer products) as the coming driver over the next two to three years.

Assessment: The shipment number flatters NAND; the pricing tells the real story. This was a volume-and-pull-forward quarter, not a NAND pricing recovery. We treat NAND as the lower-quality, higher-beta leg and weight it below DRAM/HBM in the thesis. Management's "profitability-first, prudent investment" posture for NAND is the right one.

HBM — Doubling, and Defended by More Than Speed

HBM is the franchise. Management reaffirmed the plan to double HBM revenue YoY in 2025, said 12-high HBM3E is receiving high marks from all customers on performance and supply reliability, and confirmed the industry's first HBM4 samples shipped to customers in March, with system-level optimization underway and delivery aligned to customer mass-production schedules. On the recurring profitability worry, management was direct: HBM4 carries a real cost increase (higher IO counts, new power-efficiency designs, a logic-process base die), and the plan is to reflect that in HBM4 pricing to hold current profitability.

Assessment: This is the strongest part of the story and the core of the Outperform. Management's framing, that HBM demand is "very sticky" and that leadership rests on a hard-to-copy bundle of product, yield, supply reliability, and customer engagement rather than any single spec, is consistent with a durable competitive position. The HBM4 cost/pricing question is the right thing to monitor, but a leader negotiating price increases into a supply-short market is a position of strength, not weakness.

Key Topics & Management Commentary

Overall management tone: Confident on AI memory, disciplined on the cycle, and unusually candid about the quality of the quarter. Management volunteered that part of the upside was preemptive, tariff-driven purchasing, and paired the record with a deliberately moderated Q3 guide. That candor is a credibility marker: a team papering over pull-forward would have let the record speak for itself.

1. A Record That Beat a Cautious Setup

"The past quarter began with concerns about a slowdown in demand due to trade tensions and overall economic uncertainty. However, demand for AI memory continued to grow… This was further supported by customers' preemptive purchasing to prepare for external risks, resulting in a more favorable environment than initially expected." — Song Hyun-jong, President & Head of Corporate Center

The quarter opened under a cloud (trade tensions, macro uncertainty) and closed at an all-time high. The swing factor was AI memory demand plus preemptive customer buying. Both DRAM and NAND beat guidance and pricing improved, producing the +26% sequential revenue jump.

Assessment: The beat is real, but the honest framing matters: management itself attributes a meaningful slice to customers pulling demand forward. That sets up the central analytical question of the quarter, addressed in topic 6.

2. HBM Demand: Structural and Sticky

"In this rapidly growing AI market, HBM is now positioned as a key product that quickly affects performance enhancement… we believe there is no doubt about growth potential of HBM demand." — SK hynix management, Q2 2025 earnings call

Management tied HBM's trajectory to the broadening of AI from training into inference, reasoning, and agents, each step lifting bandwidth needs, and to sovereign AI as a new long-term, government-funded demand pillar. The argument is that HBM demand is not a single-cycle phenomenon but a structural feature of how AI systems are built.

Assessment: This is the spine of the bull case and it is credible: HBM's role in determining AI system performance makes it strategically scarce, and scarcity plus co-development creates the stickiness management describes. We give this full weight.

3. The Plan to Double HBM, and HBM4 on Schedule

"Our plan to double HBM sales year over year remains unchanged… In March, we supplied the industry's first samples of HBM4 to customers. We are currently working closely with partners to optimize system-level performance." — Song Hyun-jong, President & Head of Corporate Center

The 2025 HBM-doubling target was reaffirmed, not hedged, and HBM4 reached customer sampling first in the industry. Management is sequencing the HBM3E-to-HBM4 transition while protecting profitability through pricing.

Assessment: A reaffirmed doubling target plus first-mover HBM4 samples is the cleanest evidence that the franchise is executing on schedule. The transition risk (yield, cost, customer qualification) is the standard one for a leading-edge node, and SK hynix has navigated each HBM generation from a leadership position since HBM2E.

4. HBM4 Economics: Cost Up, Price Up

"HBM4 exhibits major technological changes, including increased IO counts… new designs for improved power efficiency, and the adoption of logic processes in the base die… we are striving to reflect this cost increase in our pricing strategy for HBM4… while maintaining current profitability levels." — SK hynix management, Q2 2025 earnings call

Responding directly to the concern that HBM4's higher bill of materials could compress margins, management said it intends to pass the cost through in pricing and hold current profitability, while still setting prices at a level that "stimulates the AI market."

Assessment: The ability to push price increases through, in a market where the leader is supply-constrained and co-developing with customers, is exactly the pricing power the thesis requires. We model HBM4 as margin-neutral-to-accretive rather than dilutive, contingent on qualification timing.

5. Conventional DRAM: The DDR4 Squeeze Is Temporary

"The recent spike in DDR4 prices has been primarily driven by short-term supply concerns following announcements from some suppliers to discontinue DDR4 production… we view this as a temporary demand concentration rather than a structural shift." — SK hynix management, Q2 2025 earnings call

Legacy DDR4 prices spiked as suppliers signaled end-of-life and capacity shifted to HBM and DDR5. Management benefits from this in the near term (and noted it actually helps the China fabs, which serve legacy demand) but framed it honestly as transitional, with DDR4 falling from double-digit to single-digit revenue share and being phased out for the mass market.

Assessment: The right read: a real but temporary tailwind, not a reason to extrapolate legacy pricing. The more durable point is that capacity reallocation to HBM is tightening conventional DRAM broadly, which supports the whole DRAM complex.

6. The Pull-Forward Question and the H2 Setup

"Initially, customers intended to conservatively maintain inventory levels… but with growing uncertainty around tariff policies, they shifted strategies towards securing an appropriate level of inventory… we believe the likelihood of sharp fluctuation in supply and demand remains low." — SK hynix management, Q2 2025 earnings call

This is the crux. Management acknowledged that Q2 shipments ran well above guidance partly because tariff uncertainty pushed customers to build inventory, then argued the second half is defensible: system-build demand also rose so customer inventory did not balloon, memory suppliers' own inventories are greatly reduced, and end-of-life moves on legacy products tightened supply further. Q3 guidance nonetheless moderates (DRAM +low-to-mid single digit; NAND limited).

Assessment: The honest swing factor in the thesis. We take management's "correction unlikely" at partial credence: the supply-side argument (lean supplier inventory, capacity going to HBM) is genuinely supportive, but pull-forward by definition borrows from the future, and a tariff de-escalation that removes the urgency could soften H2 ordering. This is the single most important thing to watch, and the reason the Outperform is a considered call rather than a layup.

7. CapEx Steps Up, Anchored to HBM Visibility

"Our total investment for this year will increase from the original plan, with the majority of the additional spending allocated to HBM-related equipment. The final investment scale… will be confirmed once negotiations with major customers are concluded." — SK hynix management, Q2 2025 earnings call

2025 investment is being raised versus the original plan because management now has demand visibility on 2026 HBM and judges some proactive spend necessary. M15X opens in Q4 2025 (DRAM, including next-generation HBM from 2026) and Yongin Phase 1 completes in Q2 2027. Management repeatedly stressed CapEx discipline and that the increase is demand-visibility-led.

Assessment: A CapEx step-up underwritten by contracted HBM visibility is a higher-quality investment than a speculative capacity build. The bear watch is the perennial memory risk that today's HBM-driven build becomes tomorrow's conventional-DRAM oversupply, but the demand-anchored, customer-negotiated framing is the right discipline at this point in the cycle.

8. China Fabs and US Export Controls

"There has been no change to our validated end user or VEU status that we obtained in October 2023… we will continue operating our China fabs according to our existing plan." — SK hynix management, Q2 2025 earnings call

On the recurring geopolitical risk, management reported its VEU status unchanged and, notably, argued the recent environment actually benefits the China fabs: as capacity shifts to HBM and DDR5, legacy DDR4/LPDDR4 shortages have emerged, and the China fabs are well-suited to supply that durable legacy demand. A separate question on a major GPU customer resuming China-bound AI-chip shipments drew a measured "we are well-positioned to respond" on the associated HBM.

Assessment: Geopolitics remains a genuine tail risk for any memory maker with China exposure, but the near-term read is benign-to-favorable, and management's framing of the China fabs as a legacy-supply asset is sensible. We keep this as a monitored risk, not an active drag.

9. The Balance Sheet Keeps Improving

"Cash and cash equivalents stood at KRW 17 trillion, up KRW 2.7 trillion from the previous quarter… net debt [fell] by KRW 4.1 trillion to KRW 4.9 trillion." — Song Hyun-jong, President & Head of Corporate Center

Record cash generation funded both a rising investment plan and continued de-leveraging: cash up to ₩17tn, net debt down to ₩4.9tn, net-debt-to-equity at 6% (a 5-point improvement). EBITDA of ₩12.6tn at a 57% margin is the engine.

Assessment: A memory maker that can self-fund a HBM-driven CapEx step-up while cutting net debt toward zero is in a materially stronger financial position than in prior cycles. This balance-sheet trajectory is an under-appreciated support and a reason the downside is better cushioned than memory's history would suggest.

10. The Outlook: Inference, Agents, and Sovereign AI

"Big tech companies have begun full-fledged competition by unveiling AI agents with enhanced reasoning models… This evolution is expected to further drive demand for high-performance, high-density memory. Additionally, ongoing investments by governments… for sovereign AI are likely to become a new long-term driver." — Song Hyun-jong, President & Head of Corporate Center

Management's demand framework rests on AI broadening from training into reasoning and agents (each more memory-intensive), general-purpose server replacement on new CPUs, AI features lifting memory content in PCs and phones, and sovereign AI as a structural new pillar.

Assessment: Directionally aligned with what hyperscaler and foundry commentary corroborates. The framework is sound; the near-term swing factor remains the pace at which this structural demand offsets any payback from Q2's pull-forward.

Guidance & Outlook

SK hynix guides bit shipments and frames pricing qualitatively rather than giving revenue or margin targets. For Q3 2025:

The moderated Q3 shipment guide is the natural counterweight to a Q2 that ran hot, and it is consistent with management's pull-forward candor. The constructive signal is that the HBM-doubling target is unaffected, which is what protects the profit pool even as conventional bit growth pauses. Product cadence into H2: M15X opens in Q4 (HBM from 2026), LPDDR-based server modules and 24Gb GDDR7 within the year, and 321-layer client and enterprise SSDs by year-end.

Analyst Q&A Highlights

The Long-Term Shape of HBM Demand

The opening question asked how SK hynix sees HBM demand beyond 2025 and what the key momentum drivers are. Management tied it to the broadening of AI workloads and argued the growth runway is structural even as the early-stage hyper-growth rate normalizes.

Q: "Driven by strong demand for AI, HBM is expected for continuous and rapid growth. How do you forecast to meet the long-term demand for HBM beyond 2025? And what will be important momentums to drive HBM demand?"
— Han Dong-hee, SK Securities

A: "As you can see from continuous CAPEX expansion by big tech… AI demand is expanding from training to inferencing, and inferencing has further divided… into areas like reasoning and agents… In the future, the HBM market may not sustain the early-stage explosive growth rate, but AI technology will rapidly advance and spread… we expect HBM will continue to enjoy strong growth."
— SK hynix management

Assessment: The honest version of the bull case. Management is not promising the early hyper-growth rate forever; it is arguing for a long, durable growth curve underpinned by the broadening of AI workloads. That is the more defensible and more investable claim.

2026 HBM Supply Negotiations

An analyst pressed on how 2026 HBM supply agreements with major customers are progressing. Management declined customer specifics but characterized negotiations as on-plan, with product mix and pricing under active discussion.

Q: "How are negotiations progressing for the 2026 HBM supply with major customers?"
— SK Kim, Daiwa Securities

A: "We cannot provide customer details, but negotiations are proceeding as planned. As customer projects become increasingly diverse, we are engaged in close negotiations for product mix and pricing… offering optimal supply conditions for a long-term partnership."
— SK hynix management

Assessment: "On plan" with mix and pricing under negotiation is what you want to hear from a supply-constrained leader a year ahead. The visibility it implies is also what is underwriting the CapEx step-up.

How Much Demand Was Pulled Forward

The most important question of the call asked how much tariff uncertainty pulled demand into Q2 and what that implies for the second half. Management conceded the pull-forward but argued an inventory correction is unlikely.

Q: "To what extent did the uncertainties affect pulling demand in Q2, and how do you assess the resulting increase in customer inventory and its impact on demand for the second half?"
— Ryu Young-ho, NH Investment & Securities

A: "Our Q2 shipment volume significantly exceeded the initial guidance, and this was indeed driven by purchasing demand… customers… shifted strategies towards securing an appropriate level of inventory… System build activity also increased… so customer inventory levels did not rise to a significant level. Moreover, memory suppliers have already reduced their inventories greatly."
— SK hynix management

Assessment: We credit the supply-side half of this answer (lean supplier inventory, capacity migrating to HBM) more than the demand-side reassurance. The pull-forward is real; the question is whether structural AI demand fills the gap before any payback shows up. The moderated Q3 guide says management is planning prudently.

What Drove the NAND Shipment Surge

An analyst asked which customer segment drove NAND shipments so far above guidance. The answer was a combination of enterprise SSD, tariff pull-forward, and a China mobile bump.

Q: "Your NAND shipment growth in this quarter significantly outperformed prior guidance. Can you please elaborate on which customer segment drove this growth?"
— Jay Kwon, JPMorgan

A: "Demand for enterprise SSDs increased as hyperscale customers expanded their AI investments, while pulling demand for individual components also rose due to tariff-related uncertainty. Additionally, mobile set build demand in the China region grew due to promotional events… our inventory levels have now normalized."
— SK hynix management

Assessment: Two of the three drivers (pull-forward, China promotions) are transient; only enterprise SSD is structural. Consistent with our read of NAND as the lower-quality leg this quarter.

HBM4 Cost and Profitability

An analyst raised the concern that HBM4's higher cost could erode profitability. Management said it plans to price the cost through.

Q: "As HBM4 entails a greater cost increase compared to previous HBM products, there are some concerns about potential profitability decline. What is SK hynix's perspective on this?"
— Kim Seung-woo, Meritz Securities

A: "HBM4 exhibits major technological changes… we are striving to reflect this cost increase in our pricing strategy for HBM4, and we aim to… establish optimal pricing with customers while maintaining current profitability levels."
— SK hynix management

Assessment: Pricing power, stated plainly. A leader passing through cost in a supply-short market is the signature of the moat. We model HBM4 as broadly margin-neutral.

CapEx: How Much, and Where

An analyst asked how much 2025 investment will rise and whether the increase goes to equipment or infrastructure. Management said the increase is HBM-equipment-weighted and demand-anchored.

Q: "You mentioned that this year's investment will exceed initial plans. To what extent will it increase, and will most of the additional investment be allocated to equipment or infrastructure?"
— Peter Lee, Citigroup

A: "Our total investment for this year will increase from the original plan, with the majority of the additional spending allocated to HBM-related equipment. The final investment scale… will be confirmed once negotiations with major customers are concluded… we will continue to comply with CAPEX discipline."
— SK hynix management

Assessment: Equipment-weighted, contract-anchored, discipline-flagged. The right answer. The size will be set by customer commitments, which keeps the spend tied to visible demand.

Navigating Rising HBM Competition

The closing question asked how the HBM leader intends to navigate intensifying competition. Management leaned on the stickiness of HBM relationships and a culture-based moat.

Q: "As an HBM leading supplier, what is your strategy moving forward to navigate… intensifying competition and uncertainty…?"
— Chae Min-seok, Korea Investment & Securities

A: "As long as we are in the memory business, competition is something that we have to take… the demand for HBM has become very sticky, and leading players now have greater advantage… Such competitive [advantages] in corporate culture cannot easily be emulated… SK hynix will continue to stay top of mind of HBM customers."
— SK hynix management

Assessment: "Very sticky" plus a culture-and-execution moat is the right framing, and it is borne out by SK hynix's sustained HBM leadership since HBM2E. Competition (Samsung, Micron) is the genuine long-term risk to monitor, but nothing on this call suggested the lead is narrowing.

Market Reaction

All price levels are FMP-verified KRX closing prices; the benchmark is the KOSPI (^KS11). SK hynix discloses before the Korean market open, so the reaction is the print-day session.

• Pre-print close (2025-07-22): ₩268,500.
• Reaction (2025-07-23) close: ₩269,000, +0.2% on the day, essentially flat on a record print, while the KOSPI rose +0.4%.
• Volume: ~3.1M shares versus a ~4.2M 30-day average (0.7x), a quiet, low-conviction session.
• YTD setup: Into the print the shares were up ~+54% year-to-date (from a 2024 year-end close near ₩173,900), already pricing a good deal of the AI-memory recovery.

A flat reaction to an all-time-record quarter is the tell. After a ~54% YTD run, the buy-side had positioned for a strong print, and a record that came partly from pull-forward, with a deliberately moderated Q3 guide, gave momentum traders a reason to take profits rather than chase. The stock lagging a rising KOSPI on results day reinforces the read: this was sentiment and positioning, not a fundamental verdict.

We read the muted reaction as the opportunity rather than the warning. The operating result is a clean record, the HBM franchise is doubling, and the price did not extend on the news. For an initiation, a fundamentally strong quarter that the market met with a shrug is a more attractive entry than one met with a melt-up.

Street Perspective

Debate: How Much of the Record Was Borrowed From the Second Half?

Bull view: Management itself flagged the pull-forward and still guided HBM to double, with supplier inventories lean and capacity migrating to HBM. Structural AI demand (inference, agents, sovereign AI) fills the gap; the moderated Q3 guide is prudence, not weakness.

Bear view: A record built partly on tariff-driven inventory builds borrows from H2 by definition. If trade tensions ease and the urgency fades, Q3-Q4 ordering softens, conventional DRAM/NAND ASPs roll, and the "record" looks like a cycle peak in hindsight.

Our take: Split, leaning constructive. We credit the lean-supply argument and the HBM-doubling anchor, but treat H2 conventional bit demand as the swing variable. The HBM profit pool is what protects earnings even if conventional volume pauses, which is why the thesis rests on HBM, not on the headline shipment numbers.

Debate: Is the HBM Moat Durable, or Will Samsung and Micron Close It?

Bull view: SK hynix has led HBM since HBM2E, shipped the first HBM4 samples in March, and frames its edge as an un-copyable bundle of product, yield, supply reliability, and culture. HBM demand is "very sticky" and co-developed multiple years out.

Bear view: The HBM profit pool is the largest in memory, which guarantees Samsung and Micron throw everything at closing the gap. A competitor HBM4 qualification win at a key accelerator customer would erode the scarcity premium the whole thesis rests on.

Our take: The lead is real and, on this call, not narrowing. But HBM4 qualification news flow over the next several quarters is the single highest-stakes catalyst in either direction, and we weight it as the primary thing to monitor.

Debate: Is the Stock Cheap After a +54% YTD Run?

Bull view: The AI-memory cycle is mid-ramp, not peak, so earnings have room to grow; on forward earnings the shares trade well below the multiples NVDA/AVGO command, despite SK hynix being the indispensable HBM supplier. A clean record met with a flat reaction is the entry.

Bear view: Memory always looks cheap on the way up and the +54% YTD move has already captured the recovery. If pull-forward reverses, the "low" multiple is on earnings about to roll.

Our take: We side with the bull here, with discipline. Unlike a late-cycle peak, mid-2025 sits earlier in the AI-memory ramp, the print is clean (no non-operating inflation), and the balance sheet is de-levering. The valuation reflects memory's cyclical discount more than the HBM franchise's structural value, and that gap is the opportunity.

Valuation Framework

As of the print, 000660.KS traded near ₩268,500 (~$195) on roughly 728M shares, for a market capitalization of approximately ₩196tn (~$142B), with net debt down to ₩4.9tn. This is our initiation, so we frame the valuation rather than fold actuals into a pre-existing model.

Why the multiple understates the franchise. Memory carries a structural cyclical discount, and rightly so over a full cycle. But two things separate this setup from a typical memory peak: the cycle is mid-ramp (2025), not late, so forward earnings are still rising rather than rolling; and the print is clean (no non-operating gains propping up the headline). On forward earnings the shares sit well below the AI complex, despite SK hynix being the supply-constrained HBM leader doubling that business in 2025. The market is paying a memory-cyclical multiple for a franchise with a structural-growth core.

What would make it expensive: evidence that the pull-forward is large and H2 conventional demand rolls (cutting the forward earnings the multiple sits on), or an HBM4 competitive breakthrough that compresses the scarcity premium. We monitor both.

The base case is up double digits and the up/down skew is favorable: the bull (+38%) requires only that the franchise keep executing on a roadmap it is already delivering, while the bear (−26%) requires a demand reversal that management's lean-supply argument actively pushes against. That asymmetry, on a clean record print met with a flat tape, is what supports the Outperform.

Initiating Coverage: Bull / Bear Matrix

As our first published note on SK hynix, we set out the framework we will track each quarter.

Matrix assessment: Five bull points confirmed or on-track against one active near-term bear (pull-forward) and a set of latent/mild risks. The franchise quality is high and the print is clean. The rating turns on HBM continuing to execute (it is) and on the H2 pull-forward payback being manageable (the moderated guide and lean supply argue it is). That balance supports a constructive initiation.

Bottom Line: Initiating at Outperform

Rating decision: We initiate coverage of SK hynix (000660.KS) at Outperform. The case is straightforward: a structurally advantaged HBM franchise delivered an all-time-record, operationally clean quarter, reaffirmed a plan to double HBM in 2025, shipped the industry's first HBM4 samples, and de-levered the balance sheet, all while the AI-memory cycle remains mid-ramp and the valuation still reflects memory's cyclical discount rather than the franchise's structural value. The market met the record with a flat tape, which is the entry, not the warning.

The honest caveat, and the reason this is a considered call: part of the quarter was borrowed. Tariff-driven pull-forward inflated conventional DRAM and NAND shipments, and Q3 guidance moderates accordingly. We underwrite the Outperform on the HBM profit pool and the clean operating quality, not on the headline shipment numbers, and we will watch the H2 payback closely.

What would move us to Hold: clear evidence the pull-forward is large and H2 conventional demand and ASPs are rolling faster than HBM growth offsets; a stalled HBM4 qualification; or a sharp multiple re-rating that removes the valuation cushion.

What would move us to Underperform: a tariff de-escalation that triggers a genuine inventory correction; a Samsung or Micron HBM4 share breakthrough that contests SK hynix's scarcity premium; or signs the industry CapEx cycle is outrunning demand into 2026 oversupply.

Signposts for Q3 2025 earnings (late October 2025):