| Oct 27, 2026 | +15.48% | Q3 FY2026 | Read transcript briefingQuarter in one view- Q3 FY2026: third consecutive record, and the first KRW100trn+ quarter. Revenue KRW104.6trn (+32% QoQ, +328% YoY). OP KRW83.3trn (+38% QoQ, +632% YoY); OPM 80% (+3pp QoQ) — both all-time highs. D&A KRW4.3trn; EBITDA KRW87.7trn (84% margin).
- Earnings-quality distortion flipped sign. Non-op swung to -KRW32.6trn (FX net loss 11.6trn on won strength; investment-asset valuation losses 22.9trn), vs +62.2trn in Q2. Pre-tax KRW50.7trn; net KRW39.8trn (38% margin) — net income is now *less than half* of operating profit. The Kioxia-driven inflation is gone; a large unexplained valuation loss replaced it.
- Guides: DRAM hit, NAND missed modestly. DRAM bits +~10% QoQ (as guided); DRAM ASP +high-20% QoQ (vs +~30% in Q2 — still decelerating in absolute terms, though management called it "similar"). NAND bits decreased slightly vs the +LSD guide — management framed it as deliberate mix discipline toward 321-layer TLC eSSD; NAND ASP +mid-20%.
- Q4 guide is the most aggressive of the cycle: DRAM bits +mid-20% QoQ (1cnm expansion + M15X ramp); NAND bits +mid-single-digit. HBM4 to overtake HBM3E in Q4 in both volume and revenue share.
- Shareholder returns finally landed, at scale: ~KRW100trn FY2026 program (≥30trn dividends, ≥60trn buyback/cancellation); >50% FCF payout pledge; KRW20,000/share Q3 dividend including a KRW10,000 Kioxia special.
- Balance sheet: cash 114.5trn (+26.5trn QoQ); debt 15.8trn (-2.8trn); net cash 98.7trn — essentially at the >100trn target; D/E 5%.
- Demand visibility claims escalated sharply: 14 LTAs signed (from ~10); >50% of 2027-28 capacity under LTA; DRAM/HBM "effectively sold out" through 2029 including volumes under discussion; 2027 HBM price/volume negotiations finalized with major customers.
What management is focused on- Pre-empting the AI-bubble debate with customer financials. The CEO and CFO both front-loaded demand durability. Against the monetization question (Citi), management cited hyperscaler cloud revenue +30-80% YoY, cloud order backlogs up 2-5x YoY, rising paying users, and a broadening investor base (sovereign AI, non-tech enterprises, on-prem) that makes demand resilient to any single buyer pausing. New quantified scaffolding: CSP capex +60% in 2027, secured power capacity +100% YoY, AI data-center infrastructure market +40% CAGR to 2030 (all attributed to market research firms, not company data).
- "Sold out through 2029" as the centerpiece. LTA count rose to 14; >50% of 2027-28 capacity committed; management said committed plus under-discussion volumes make DRAM and HBM effectively sold out for three years. Customers are reportedly requesting *extensions* beyond the typical 5-year term. This is the strongest version yet of the no-oversupply argument — and the most falsifiable.
- Defending the capex escalation with a "shell-first, ramp-later" doctrine. Against BofA's overproduction question: new fabs (Yongin, M17) will accelerate shell construction but open cleanrooms in phases; equipment move-in and ramp adjust to demand; investment screens include customer financial health, funding capability, and supply-chain readiness. "More fabs does not necessarily mean more supply" is the new mantra, justified by HBM/process complexity consuming wafer capacity.
- Neutralizing the HBM stack-lowering threat. Management reframed customer spec adjustments (12-high HBM4 → 8-high HBM4E) as proof of shortage, not weakness: lower stacks mean more cubes, more base dies, more back-end processing — and higher per-die capacity offsets much of the bit impact. Claimed customer volume requirements are more than 2x current supply capability.
- Delivering the long-deferred shareholder return, and reframing investment gains as a returns channel. The Kioxia gain is explicitly being recycled into special dividends ("return most of the gains"), with the Solidigm IPO cancellation explained as market-driven and partly motivated by preserving shareholder-return capacity.
Key numbers and quarter mechanics- P&L: revenue 104.6trn (+32% QoQ, +328% YoY); OP 83.3trn (80% OPM, +3pp QoQ); D&A 4.3trn; EBITDA 87.7trn (84%); non-op -32.6trn (FX -11.6trn, investment-asset valuation -22.9trn); pre-tax 50.7trn; net 39.8trn (38% margin).
- DRAM: bits +~10% QoQ (in line); ASP +high-20% QoQ. Server DDR5 shipments +40% QoQ; HBM4 full-scale mass production and shipments began in Q3. ASP trajectory: +mid-60% (Q1) → +~30% (Q2) → +high-20s (Q3) — deceleration continues but is now explicitly mix-driven (HBM4 priced above HBM3E).
- NAND: bits down slightly QoQ (below the +LSD guide — first bit-guide miss in at least five quarters, attributed to deliberate high-value mix adjustment); ASP +mid-20% QoQ (vs +mid-50s in Q2 — sharp deceleration).
- Q4 guide: DRAM bits +mid-20% QoQ; NAND bits +mid-single-digit QoQ. No ASP guidance, as usual.
- 2027 demand view: DRAM bit growth +mid-20%, NAND +high-teens — "achievable" growth similar to 2026, i.e., supply-capped, not demand-capped. Supply shortfall "projected to continue through 2030."
- Balance sheet/capex: cash 114.5trn; debt 15.8trn; net cash 98.7trn; D/E 5%. 2027 capex to "increase significantly" vs 2026's high-40trn — no absolute figure given.
- Shareholder return mechanics: FY2026 total ~100trn = ≥30trn dividends + ≥60trn buyback/cancellation (incl. the 40trn August program, completed last week, cancellation pending). Q3 actions: 14trn dividend (KRW20,000/share = 10,000 regular + 10,000 Kioxia special) + 20trn buyback, of which 10trn for employee comp is excluded from the shareholder-return tally — net Q3 return ~24trn. Remainder of the Kioxia special comes in the year-end dividend.
- Persistent gaps: still no DRAM/NAND revenue split, no HBM revenue or mix quantification, no segment margins, no LTA pricing floors/caps or deposit magnitudes, and no explanation of the KRW22.9trn investment-asset valuation loss (composition, which assets, whether Kioxia-related marks reversed).
Product and launch scorecard- HBM4: the strongest quarter of execution evidence yet. Full-scale mass production and shipments in Q3; 12-high shipments on schedule per customer agreements; yield "already exceeded its planned yield" — target revised up; claimed industry-best fab *and* full-flow back-end yields. Q4 crossover: HBM4 to surpass HBM3E in both shipment and revenue share, becoming the flagship. M15X ramping to full capacity by year-end. Still no absolute HBM revenue, share, or customer-count disclosure.
- 2027 HBM negotiations: finalized, not just "progressing smoothly." Volume and price concluded with major customers, "faithfully reflecting" HBM4's added value. Notably, management expects the HBM-vs-conventional profitability gap to narrow meaningfully — readable two ways (conventional pricing catching up, and/or HBM4 pricing not fully keeping pace); direction of HBM pricing itself was not stated.
- HBM4E: sample shipments completed in Q2 (consistent with last quarter's early-sampling claim); development/qualification on customer timelines; 8-high variant explicitly positioned as the stack-lowered option with higher per-die capacity offsetting bit impact.
- Custom HBM: elevated to a named strategic thread — "growing importance," projects underway with major customers, "first supplier to key customers" claim. No specifics.
- SOCAMM / server LPDDR: Q4 plan to expand 1cnm DDR5 and "AI server LPDDR5 SOCAMM" sales — note the transcript says LPDDR5 SOCAMM, not the SOCAMM2 branding used previously; worth checking whether this is translation or a product-naming change.
- 321-layer NAND: to exceed 50% of domestic capacity in Q4 (prior target was end-2026 — effectively on/ahead of schedule); eSSD supply expanding.
- NAND AI-storage roadmap got concrete: SALT-KV (software tiering of KV cache across HBM/DRAM/SSD by reuse value) and CMM-hybrid (CXL-based, full KV cache on SSD with DRAM preloading) named as in-development technologies; system demonstrations on real AI workloads underway with customers/platform partners; HBF repositioned as the future tier between HBM and SSD. This is the most detailed NAND architecture disclosure to date.
- 4F²/3D DRAM (new disclosure): strategy is to extend 6F² via EUV scaling and process maturity, then transition to 4F² or 3D DRAM "at the right time" — a deliberate fast-follower posture versus competitors pushing new cell structures earlier.
- Quiet/dropped threads: iHBM (last quarter's reveal) not mentioned; LPDDR6, CXL as a product line, PQC21, 245TB eSSD, GDDR7, OpenAI LOI, tariffs, China/VEU all still absent. Solidigm's eSSD momentum not quantified this quarter (no repeat of the ">3x QoQ" datapoint).
Sell-side read-through- Tone: congratulatory, zero challenge. Every questioner opened with congratulations; no one pressed on the NAND bit-guide miss, the KRW22.9trn valuation loss, the ASP deceleration, or the HBM/conventional margin-gap comment. The sell-side has moved from testing the story (Daiwa's ASP challenge in Q2) to underwriting it.
- Peter Lee (Citi): AI monetization risk — got the most evidence-based demand answer yet (hyperscaler cloud growth 30-80%, backlogs 2-5x, paying-user growth, investor-base diversification). Still entirely customer-cited, no SK hynix-specific data.
- Sunwoo Kim (Meritz): dividend/buyback split rationale — answer: sustainable dividend benchmarked to peers and yield, specials for investment gains, buybacks because "our share price today is undervalued." Quarterly buyback sizing left to Board discretion — flexibility retained.
- Giuni Lee (Goldman): LTA implications — extracted the call's biggest claim: DRAM/HBM effectively sold out through 2029 (committed + under-discussion volumes), with >50% LTA coverage of 2027-28 capacity to be maintained "over the medium to longer term."
- Simon Woo (BofA): the only cycle-risk question (overproduction) — got the phased-cleanroom/shell-first doctrine and the demand-screening framework. A genuine answer on mechanism, but the flexibility pledge is untested.
- Jay Kwon (JPM): HBM4/4E/custom update — produced the yield-above-plan disclosure and the Q4 crossover confirmation.
- Dong-hee Han (SK Securities): HBM stack-lowering — the most commercially sensitive question; management's answer (demand >2x supply, more cubes/base dies, profitability "far above the peak in past cycles") is confident but concedes customers are actively re-spec'ing HBM, which is new information about supply tightness and customer behavior.
- Young-gun Kim (Mirae): 4F² — a technology-positioning question reflecting competitive anxiety about next-gen DRAM; answer was deliberately non-committal on timing.
- Young-ho Ryu (NH): KV-cache tiering and NAND demand — produced the SALT-KV/CMM-hybrid disclosures.
- Notable absences: no question on the non-op loss composition; no NAND miss question; no Solidigm IPO-cancellation follow-up (funding implications for Solidigm's fab expansion unaddressed); no tariff/China/OpenAI questions for the third-plus consecutive quarter.
Management credibility- Positive: the shareholder-return promise was kept, big. After two-plus quarters of "within the year" deferrals, management delivered a ~100trn program with completed (not promised) execution on the 40trn August buyback, a Board-resolved 14trn dividend, and a >50% FCF framework. The Kioxia special dividend also honors the stated principle of recycling strategic-investment gains. This is the quarter's largest credibility deposit.
- Positive: HBM4 delivery exceeded its own plan. Yield above target (target revised up), Q4 crossover confirmed, 2027 negotiations finalized — the H2 ramp claim that was "falsifiable" after Q2 has so far been validated.
- Positive: LTA trajectory is consistent and compounding — ~10 customers (Q2) → 14 (Q3), with the new >50%-of-2027/28-capacity quantification answering last quarter's coverage-ratio refusal.
- Negative: first bit-guide miss in the track record. NAND guided +LSD, delivered a slight decline. Management's mix-discipline explanation is plausible (321-layer eSSD focus) but unaccompanied by eSSD revenue data this quarter — the one quarter they stopped quantifying eSSD is the quarter NAND missed.
- Negative: the KRW22.9trn investment-asset valuation loss was disclosed as a line item and never explained. After Q2's +63.3trn gain, the swing in non-op items across two quarters is ~96trn. Reported net income remains nearly useless for trend analysis, and management volunteered nothing.
- Mixed: "profitability gap between HBM and conventional DRAM to narrow meaningfully" is a striking disclosure buried in prepared remarks — it implicitly concedes conventional DRAM is the marginal profit driver into 2027, which cuts against the HBM-premium narrative management has cultivated. No analyst picked it up.
- Watch: "effectively sold out through 2029" includes volumes merely "under discussion" ("highly likely to come through"). Folding uncommitted demand into a sold-out claim is aggressive framing; the hard number remains >50% of 2027-28 capacity.
- Watch: Solidigm IPO cancellation is reasonably explained (market volatility), but it reverses the financing plan for Solidigm's manufacturing expansion with no stated alternative — a loose end.
What changed versus the prior quarter- Earnings quality inverted: non-op +62.2trn → -32.6trn; net margin 118% → 38% vs OPM 76% → 80%. Operating profit is now the dominant and clean number; net income is depressed by FX and unexplained valuation losses.
- Shareholder returns went from promise to policy: ~100trn FY2026 program, >50% FCF pledge, Kioxia special dividend — versus "within the year" deferrals gated by ADR constraints. The ADR excuse disappeared entirely.
- Demand-visibility claims escalated an order of magnitude: ~10 LTAs → 14; coverage ratio refused → >50% of 2027-28 capacity disclosed; new "sold out through 2029" framing; customers requesting term extensions.
- 2027 HBM negotiations concluded (volume and price) — versus "underway, progressing smoothly" in Q2. New and underexplained: HBM/conventional profitability gap to "narrow meaningfully."
- Volume acceleration replaced price as the growth engine: Q4 DRAM bit guide +mid-20% (vs +~10% for Q3) on 1cnm + M15X; DRAM ASP growth still decelerating (+30% → +high-20s); NAND ASP decelerated hard (+mid-50s → +mid-20s).
- NAND missed its bit guide for the first time in the recent track record (slight decline vs +LSD), and the eSSD quantification that supported the Q2 NAND story was not repeated.
- Capex trajectory steepened: 2027 guided "significantly" above 2026's high-40trn, with a formalized shell-first/phased-cleanroom execution doctrine.
- HBM stack-lowering emerged as a new, management-acknowledged phenomenon — customers re-spec'ing 12-high to 8-high; framed as shortage evidence.
- Solidigm IPO cancelled — a strategic reversal attributed to market conditions.
- PC/mobile framing blended: Q2's pure "supply starvation" became "cost burdens from rising memory prices and supply shortages" — demand destruction is back in the language alongside shortage.
- New disclosures: SALT-KV, CMM-hybrid, 4F²/3D DRAM strategy, custom HBM emphasis, CSP capex +60%/power +100% 2027 forecasts, supply constraint "through 2030."
- Dropped/quiet: iHBM, SOCAMM2 branding, Solidigm eSSD metrics, LPDDR6, CXL (as product), tariffs, China/VEU, OpenAI LOI.
Bull case- The Q4 guide implies the cycle's biggest volume quarter yet: DRAM bits +mid-20% with HBM4 crossing over and M15X hitting full capacity — revenue and OP records are near-certain to extend even with flat ASPs.
- Contractual visibility is now quantified and large: >50% of 2027-28 capacity under LTA with deposits, 14 customers, extension requests, and finalized 2027 HBM pricing — the oversupply bear case requires believing 14 sophisticated customers all overcommitted simultaneously.
- HBM execution is best-in-class by management's own falsifiable metrics: yield above plan, back-end yield claimed industry-best, Q4 crossover, custom HBM pipeline — and customer demand claimed at >2x supply capability.
- Capital returns are now real and recurring: ~100trn FY2026, >50% FCF pledge, quarterly buyback cadence, special-dividend mechanism for investment gains — with net cash at 98.7trn and 84% EBITDA margins funding it without strain.
- NAND has an architecture-level AI story now: KV-cache tiering (SALT-KV, CMM-hybrid) and HBF create structural bit demand beyond the eSSD cycle, with 321-layer cost leadership past 50% of capacity.
- Supply discipline is structural, management argues: process complexity and HBM resource intensity cap bit growth industry-wide "through 2030," and the shell-first capex doctrine limits overshoot risk.
Bear case- Reported net income is still noise. A KRW22.9trn valuation loss with zero explanation, one quarter after a 63.3trn gain — two-quarter non-op swing of ~96trn. Whatever is being marked (residual Kioxia? other stakes?) is large, volatile, and undisclosed.
- NAND cracked first: bit-guide miss, ASP growth halved sequentially, and the eSSD quantification quietly dropped from the script. If AI-storage demand were as strong as the architecture slides suggest, the revenue datapoint should have been repeated.
- "Sold out through 2029" overstates the hard number. Committed LTA volume is >50% of 2027-28 capacity; the 2029 claim leans on volumes "under discussion." If even the committed half carries market-linked price ranges (as disclosed), LTAs protect volume, not price — downside participation is preserved.
- The HBM margin-gap comment may be a quiet concession. If conventional DRAM profitability is converging *up* toward HBM, fine; if HBM4 2027 pricing came in below the premium narrative, the finalized negotiations locked that in. Management did not say which, and no one asked.
- Stack-lowering is a real demand-leakage vector management chose to normalize: customers stretching limited HBM across more accelerators means fewer bits per system — the "more cubes, more base dies" offset is asserted, not quantified.
- Capex is compounding at peak margins with the returns tap now open: 2027 "significantly" above high-40trn while committing ~100trn to shareholders — any demand air pocket squeezes both simultaneously, and the Solidigm IPO failure removes one funding relief valve.
- PC/mobile demand destruction is back in the language ("cost burdens"), and the sell-side's unanimous congratulation means consensus is no longer stress-testing any of this.
Next-quarter watchlist- Q4 delivery vs the +mid-20% DRAM bit guide — the boldest guide of the cycle; M15X full-capacity-by-year-end and 1cnm ramp are the mechanics to verify. Any miss here breaks a multi-quarter clean streak at the worst possible time.
- HBM4 crossover confirmation: whether HBM4 did surpass HBM3E in Q4 volume and revenue, and any quantification of HBM revenue/mix (still never disclosed).
- The HBM/conventional margin-gap trajectory: Q4 segment commentary and 2027 pricing color that clarifies whether the narrowing is conventional-up or HBM-down.
- NAND recovery: Q4 bits +mid-single-digit guide vs Q3's miss; whether eSSD/≥30TB quantification returns; 321-layer crossing 50% of domestic capacity.
- The KRW22.9trn valuation loss: composition, recurrence risk, and remaining investment-asset exposure; whether Q4 non-op normalizes and net margin re-converges toward OPM.
- Year-end dividend and Kioxia special remainder: size of the "most of the gains" payout; whether FY2026 total return lands at the full ~100trn; Q1 2027 buyback cadence under the quarterly Board-resolution mechanism.
- LTA evidence: new signings beyond 14, any deposit visibility on the balance sheet, and whether "sold out through 2029" gets repeated, hardened, or quietly softened.
- 2027 capex number: the absolute figure (2026 was high-40trn; "significant increase" implies 50trn+); Yongin Fab 1 early-2027 cleanroom on schedule; M17 and post-Yongin cluster milestones; any overseas-site decision.
- Stack-lowering follow-through: whether 8-high HBM4E volumes materialize and how management reconciles bit impact with the "more cubes" offset claim.
- Solidigm funding plan: how manufacturing expansion gets financed post-IPO-cancellation, and whether eSSD momentum metrics return.
- Overdue items: tariff strategy (now three-plus quarters past the promised update), OpenAI LOI, China/VEU, iHBM status, LPDDR6, 245TB eSSD, GDDR7, and the SOCAMM vs SOCAMM2 naming question.
|
| Jul 28, 2026 | -11.35% | — | Read transcript briefingQuarter in one view- Q2 2026: the supercycle compounded again. Revenue KRW79.3trn (+51% QoQ, +257% YoY) — second consecutive record. OP KRW60.5trn (+61% QoQ, +557% YoY); OPM 76% (+5pp QoQ, all-time high). D&A KRW4.0trn; EBITDA KRW64.6trn (81% margin); pre-tax KRW122.7trn; net KRW93.9trn (118% net margin).
- Both guides hit; ASPs decelerated but stayed extreme. DRAM bits +high-single-digit QoQ (as guided); DRAM ASP +~30% QoQ (vs +mid-60% in Q1 — sharp deceleration, and Daiwa flagged it as below market expectations). NAND bits +mid-teens QoQ (as guided); NAND ASP +mid-50% QoQ (vs +mid-70%). eSSD revenue doubled QoQ; Solidigm ≥30TB eSSD revenue more than tripled QoQ.
- Earnings quality distortion went from large to dominant. Non-op +KRW62.2trn, including KRW63.3trn from sale/valuation of investment assets (Kioxia stake sale confirmed in Q&A) plus KRW1.1trn FX gains. Net profit (93.9trn) exceeds operating profit (60.5trn) by 55% — the majority of the headline print is one-off.
- Balance sheet: cash incl. ST investments KRW88trn (+33.6trn QoQ); debt KRW18.6trn (-0.7trn); net cash KRW69.4trn (from 35trn — 69% of the way to the >100trn target in one quarter); D/E 7% (from 12%).
- Capex finally quantified: 2026 capex "high KRW40 trillion range" — the first absolute figure in at least six quarters, driven by M15X pull-forward and Yongin Fab 1 (early 2027 cleanroom). New announcements: P&T7 (advanced packaging), M17 (new NAND fab), and a new domestic cluster beyond Yongin.
- ADR executed: listed on Nasdaq July 10; described as the largest-ever US IPO by a foreign company. Conversion limit 17,790,000 shares.
- Q3 2026 guide: DRAM bits +~10% QoQ; NAND bits +low-single-digit QoQ. No ASP guidance.
What management is focused on- Defending AI-capex durability against a new bear angle. The opening question (J.P. Morgan) raised data-center leasing and efficient models as slowdown risks. Management reframed both as utilization/monetization transitions, citing "explosive demand for recent high-efficiency AI models" and customer conversations supporting investment "beyond next year." Same structural-demand defense as prior quarters, now aimed at a different threat.
- LTA architecture disclosure — the biggest new information on the call. ~10 customers signed (including key customers); typical term ~5 years; non-uniform pricing mechanisms designed to damp volatility; deposits incorporated to enforce purchase commitments and improve demand visibility. LTA share of total sales deliberately undisclosed ("appropriate level"). This is the evidentiary backbone of the no-oversupply claim.
- HBM4 execution narrative restored. Mass production shipments began in Q2 for key customers; yield and quality "nearing HBM3E levels" (HBM3E described as mature); full ramp in H2. This directly answers last quarter's language-drift concern.
- Capacity expansion as competitive duty. M15X mass production pulled forward; Yongin Fab 1 early-2027 cleanroom followed by rapid capacity build-out; P&T7 packaging and M17 NAND announced; new domestic cluster beyond Yongin — a reversal of last quarter's flat "no plans" for fabs beyond Yongin. Overseas expansion left open ("no further decisions"), with site selection framed around power, water, talent, supply chain, customer access.
- Shareholder returns still deferred. "Various options" under evaluation; specifics blocked by ADR-related regulatory constraints; communication promised "within the year" — same deadline as last quarter, now with a regulatory rationale.
Key numbers and quarter mechanics- P&L: revenue 79.3trn (+51% QoQ, +257% YoY); OP 60.5trn (76% OPM, +5pp QoQ); D&A 4.0trn; EBITDA 64.6trn (81%); non-op +62.2trn (investment-asset sale/valuation +63.3trn, FX +1.1trn); pre-tax 122.7trn; net 93.9trn (118% margin).
- Volumes/ASP: DRAM bits +HSD QoQ (in line); DRAM ASP +~30% QoQ — management attributed the below-expectation print to high-value product shipments pushed into H2 and mix effects, and guided H2 bit growth above H1 with HBM4 and 1cnm mix lifting blended ASP. NAND bits +mid-teens QoQ (in line, off Q1's low base); NAND ASP +mid-50% QoQ.
- Mix detail: sales centered on HBM3E and AI server DRAM; server LPDDR incl. SOCAMM2 "grew significantly"; eSSD revenue 2x QoQ; Solidigm ≥30TB eSSD revenue >3x QoQ.
- Demand outlook: 2026 DRAM demand +mid-20%, NAND +high-10s (bit growth); supply-demand balance unlikely to improve near-term due to process complexity and fab lead times.
- Q3 guide: DRAM bits +~10% QoQ (server-led); NAND bits +LSD QoQ — a notable NAND deceleration vs Q2's mid-teens.
- Balance sheet/capex: cash 88trn; debt 18.6trn; net cash 69.4trn; D/E 7%; 2026 capex high-40trn range.
- PC/mobile: "temporary sales adjustments due to difficulties in securing memory" — the framing shifted from demand destruction (Q1: cost-pressure shipment cuts) to supply starvation; recovery expected as shortage eases.
- Persistent gaps: no DRAM/NAND revenue split, no HBM revenue/mix, no segment margins, no LTA coverage ratio, no HBM pricing detail, no Kioxia gain mechanics.
Product and launch scorecard- HBM4: mass production shipments began Q2 for key customers; yield/quality "nearing HBM3E" levels; full ramp H2 2026. Clears last quarter's "getting ready to ramp" ambiguity — the strongest execution evidence on the call. Still no volume, share-of-HBM, or customer-breadth quantification.
- HBM4E: samples delivered to a major customer in H1 (ahead of the prior "H2 2026 samples" timeline); built on a mature, proven process; volume production targeted 2027.
- iHBM (new disclosure): in development for HBM5-era thermal management; cooling elements integrated in-package; >30% thermal-resistance reduction claimed. Pre-product but a concrete differentiator alongside hybrid bonding.
- SOCAMM2 (1cnm): full supply commenced in Q2; server LPDDR sales "grew significantly"; sample shipments planned to broaden the customer base beyond the initial platform.
- 321-layer NAND: became the largest share of NAND production in Q2; ~50% of domestic capacity by end-2026 reiterated ("as planned").
- NAND AI-storage strategy: portfolio framing across TLC eSSD, high-capacity QLC eSSD (data lake/HDD replacement), and SLC-mode high-performance SSDs; new tiers for KV-cache offload and near-GPU storage in development. Solidigm's ≥30TB momentum (>3x QoQ) is the first hard revenue evidence for the QLC/HDD-replacement thesis.
- 2027 HBM pricing: negotiations underway, "progressing smoothly"; management explicitly linked HBM pricing to sharply higher conventional DRAM prices plus resource/opportunity cost of HBM production — an analyst-readable signal that 2027 HBM pricing is being negotiated up, not down.
- Quiet/dropped threads: LPDDR6 flagship win not mentioned; CXL, HBF, PQC21 not mentioned; 245TB eSSD, GDDR7, LPDDR server module, OpenAI LOI, tariffs, China/VEU all still absent.
Sell-side read-through- Jay Kwon (J.P. Morgan): AI-capex slowdown risk (leasing, efficient models) — the cycle-longevity question has evolved from spot-price softness to hyperscaler capex behavior. Management's answer was emphatic and customer-cited but number-free.
- Rok-ho Kim (Hana): basis for long-term demand and oversupply risk — got the LTA-anchored demand-visibility defense and "phased, flexible" execution pledge. "Will not lead to oversupply right away" is a carefully qualified version of last quarter's stronger claim.
- Sunwoo Kim (Meritz): LTA terms — extracted the call's most valuable disclosures: ~10 customers, ~5-year typical terms, deposits, non-uniform pricing. Coverage ratio refused.
- SK Kim (Daiwa): the only challenging question — DRAM ASP +~30% "fell below market expectations." Management blamed H2-pushed high-value shipments and mix, and guided H2 bit growth above H1 with mix-driven ASP recovery. A plausible but unverifiable explanation; it concedes Q2 blended ASP was mix-depressed, not price-depressed.
- Dong-hee Han (SK Securities): HBM competitive moat vs fast-progressing rivals — produced the yield-nearing-HBM3E claim and the iHBM reveal. Competition anxiety is now explicit in the question set.
- Nicolas Gaudois (UBS): 2027 HBM pricing for HBM3/4/4E — got direction (constructive, conventional-DRAM-linked) but no terms, as expected.
- Sanjeev Rana (CLSA): overseas expansion (US/Japan) — door opened ("no domestic/overseas distinction in principle") but nothing decided.
- Hyung-geun Ryu (Daishin): NAND segmentation (QLC vs SLC-mode) — portfolio answer, no product specifics or customer names.
- Young-ho Ryu (NH): ADR fungibility — one-way conversion initially (ADR→stock free from July 30; stock→ADR limited by filings and the 17.79mn-share cap); no decision on expanding the program.
- Su-rim Lee (DS): capital allocation post-Kioxia/ADR — returns still "within the year," now with an ADR regulatory gag as the stated reason for no specifics.
- Notable absences: no question on the KRW63.3trn one-off gain dominating the print (third straight quarter of unexamined non-op swings); no HBM4 share/volume quantification demanded this time; no tariff, China, or OpenAI questions; no challenge on the NAND Q3 guide decelerating to LSD.
Management credibility- Positive: guidance hit on both axes again. DRAM +HSD, NAND +mid-teens — both as guided. The bit-guidance track record is now four-plus quarters clean.
- Positive: HBM4 language-drift resolved by delivery. Last quarter's "getting ready to ramp" became "mass production shipments began in Q2" with yields nearing HBM3E — the softening was scheduling, not execution. HBM4E samples also landed ahead of the stated H2 window.
- Positive: capex opacity ended. High-40trn 2026 figure after six quarters without a number; M15X pull-forward and Yongin timeline give it structure.
- Positive: LTAs moved from "structural options under review" to ~10 signed deals with disclosed architecture (5-year terms, deposits) — real progress on the prior quarter's key watch item, though coverage and pricing remain undisclosed.
- Mixed: ASP explanation under pressure. The below-consensus DRAM ASP was attributed to mix and shipment timing — consistent with the HBM4 H2 ramp story, but it means Q2's headline ASP strength was partly portfolio optics, and the H2 recovery claim is now falsifiable.
- Negative: headline earnings are now mostly non-operating. Net margin 118% vs OPM 76%; KRW63.3trn of one-off gains in a 93.9trn print. Management presented the net figure without flagging its composition beyond the line-item disclosure, and no analyst probed it.
- Negative: shareholder returns slipped again in substance. Same "within the year" promise as Q1, now with a regulatory constraint attached — the ADR process has become a reason for continued deferral while net cash piles toward 100trn.
- Watch: reversal on fabs beyond Yongin. Last quarter: "no plans." This quarter: new domestic cluster announced plus openness to overseas sites. The change is explained by demand, but it is a genuine strategic U-turn within one quarter.
What changed versus the prior quarter- ASP momentum decelerated hard while volumes re-accelerated: DRAM ASP +~30% (vs +mid-60%), NAND ASP +mid-50% (vs +mid-70%); DRAM bits +HSD and NAND bits +mid-teens (vs flat/-10%). The quarter was more volume-real than Q1 — and Q3 guides DRAM +~10%, the strongest bit guide of the cycle.
- HBM4 crossed from preparation to revenue: Q2 mass-production shipments, yields nearing HBM3E, H2 full ramp; HBM4E sampled early. Last quarter's central credibility question was answered.
- Capex went from "increase significantly" to high-40trn, with M15X pulled forward and three new capacity announcements (P&T7, M17, post-Yongin cluster). The investment supercycle is now matching the price supercycle.
- LTAs became concrete: ~10 customers signed, ~5-year terms, deposits, volatility-dampening pricing — vs "various approaches under review" in Q1.
- ADR completed: Nasdaq listing July 10, largest foreign-company US IPO; fungibility mechanics disclosed.
- Earnings quality deteriorated from caveat to distortion: non-op gains 14trn → 62.2trn; net margin 77% → 118% vs OPM 72% → 76%. The Kioxia sale is new information explaining the jump.
- PC/mobile narrative flipped: from demand destruction (cost-pressure shipment cuts) to supply starvation ("difficulties securing memory") — a friendlier framing that implies deferred, not destroyed, demand.
- New disclosures: iHBM (>30% thermal-resistance reduction, HBM5-era); 321-layer now the largest share of NAND output; 2026 demand growth forecasts (DRAM +mid-20%, NAND +high-10s); 2027 HBM pricing talks underway with conventional-price linkage.
- Dropped/quiet: LPDDR6, CXL, HBF, PQC21, tariffs, OpenAI LOI, China/VEU, 245TB eSSD, GDDR7.
Bull case- The cycle is now volume-led, not just price-led: Q3 DRAM bit guide of +~10% with HBM4 ramping and 1cnm conventional mix rising implies H2 revenue growth even if conventional ASP momentum flattens. Management explicitly guided H2 bit growth above H1 with mix-driven ASP improvement.
- HBM execution risk is de-risked: HBM4 in mass production at near-mature yields, HBM4E sampled early, 2027 pricing talks "progressing smoothly" with conventional-price tailwinds, and iHBM extending the roadmap into HBM5.
- Demand visibility is now contractual: ~10 five-year LTAs with deposits convert the no-oversupply argument from assertion to structure — and give management a rational basis for the high-40trn capex and post-Yongin cluster.
- NAND has a second engine: eSSD revenue doubling QoQ and Solidigm ≥30TB tripling is the first hard evidence that AI-inference storage (QLC HDD replacement, KV-cache offload) is a revenue line, not a slide.
- Balance sheet is a fortress: 69.4trn net cash, 7% D/E, 81% EBITDA margin — the 100trn net-cash target is reachable within quarters, and the ADR broadens the investor base into it.
- PC/mobile weakness reframed as supply starvation implies pent-up demand that returns when supply eases — upside to the +mid-20% DRAM demand forecast rather than downside.
Bear case- The headline print is mostly one-off. Strip the KRW63.3trn investment gains and Q2 net profit is roughly half the reported 93.9trn; net margin of 118% is not an earnings-power data point. Three straight quarters of large unexamined non-op items make reported net income nearly useless for trend analysis.
- DRAM ASP missed expectations, and the explanation (mix, H2-pushed shipments) concedes that blended pricing power is now hostage to HBM4 ramp execution. If the H2 mix recovery doesn't materialize, the ASP deceleration (+mid-60% → +30%) will look like the start of normalization, not a timing artifact.
- NAND momentum is already fading in the guide: Q3 bits +LSD after +mid-teens, with no ASP guide — the eSSD surge may be lumpy.
- Capex is escalating aggressively at peak margins: high-40trn in 2026 plus M15X acceleration, M17, P&T7, and a post-Yongin cluster — a strategic U-turn from "no fabs beyond Yongin" one quarter ago. If the LTA-anchored demand visibility proves softer than claimed, this is exactly how past oversupply cycles were built.
- Shareholder returns remain a promise, not a policy: two consecutive quarters of "within the year," now gated by ADR regulatory constraints, while cash accumulates and reinvestment is prioritized.
- Competitive anxiety is rising (explicit HBM-moat question), and management's answer rested on track record rather than quantified current share, qualification breadth, or 2027 volume commitments.
- Still unexamined: tariffs, China/VEU, OpenAI LOI, customer concentration, LTA pricing floors/caps, and the terms under which deposits could be forfeited.
Next-quarter watchlist- H2 mix-recovery claim: whether HBM4 ramp and 1cnm conventional shipments deliver the promised H2 bit growth above H1 and blended-ASP improvement — the direct test of the Q2 ASP-miss explanation.
- Q3 delivery vs guide: DRAM bits +~10%, NAND bits +LSD; whether NAND's deceleration is timing or eSSD lumpiness; any ASP commentary as 2027 contract talks progress.
- HBM4 quantification: share of HBM revenue, customer count, and whether 2027 HBM pricing concludes with disclosed direction; HBM4E development milestones toward 2027 volume production.
- LTA conversion and coverage: additional signings beyond ~10, any disclosure of sales coverage ratio, deposit magnitudes, or pricing-mechanism structure — and whether deposits show up in the balance sheet.
- Shareholder-return plan: the twice-promised "within the year" framework — size, recurrence, mix — now that the ADR regulatory window is the stated constraint; watch whether the constraint outlives the lockup period.
- Capex trajectory: whether high-40trn holds or rises; M15X ramp contribution; Yongin Fab 1 early-2027 cleanroom on schedule; any overseas-site decision following the opened door.
- Earnings quality: Q3 non-op composition post-Kioxia; whether net margin re-converges toward OPM; any remaining investment-asset stakes that could produce further one-offs.
- SOCAMM2 and Solidigm traction: customer-base expansion via samples; whether ≥30TB eSSD growth sustains after a 3x quarter; 321-layer progress toward ~50% of domestic capacity by year-end.
- PC/mobile: whether "supply starvation" framing holds or demand destruction re-emerges; any LPDDR6 flagship evidence.
- Overdue items: tariff strategy (two quarters past the promised update), OpenAI LOI, China/VEU, LPDDR6/CXL/HBF status, 245TB eSSD, GDDR7, LPDDR server module.
|
Corbanu