SK hynix (Trans): Revenue topped KRW 100tn; full-year shareholder returns of KRW 100tn

Below is Dolphin Research's Trans of $SK Hynix(SKHY.US) 2026 Q3 earnings call

I. Core takeaways

1. Shareholder returns

a. FY2026 shareholder returns approx. KRW 100 tn: cash dividends no less than KRW 30 tn, buybacks and cancellation no less than KRW 60 tn (incl. KRW 40 tn announced in Aug, repurchased last week and to be cancelled shortly), delivering the pledge to return 50%+ of FCF.b. BOD resolution today: Q3 DPS of KRW 20,000, totaling KRW 14 tn, of which KRW 10,000 is performance-linked and KRW 10,000 is a special dividend from Kioxia investment gains. Most Kioxia gains will be returned via specials, with the remainder included in the year-end payout.c. This quarter's buyback of KRW 20 tn: KRW 10 tn for shareholder returns and fully cancelled, and KRW 10 tn for management and employee comp, not counted as shareholder returns; excluding the latter, Q3 shareholder returns are approx. KRW 24 tn.d. Ongoing quarterly buybacks thereafter, with size set by the BOD based on share price, market and cash flow.

2. Outlook and capex

a. Q4: DRAM bit shipments to grow by the low-20s percent QoQ; NAND bit shipments to grow by the mid-single digits QoQ.b. In 2027, constrained supply implies achievable demand growth similar to 2026: DRAM up low-20s percent, NAND up high-teens percent; revenue and profit to rise materially in absolute terms.c. 2027 capex to increase sharply vs. 2026 to lift advanced node mix, ramp the first Yongin fab and fund infra at home and abroad, with pace and size tied to demand visibility and returns.

3. Key Q3 financials

a. Top line: revenue of KRW 104.6 tn (+32% QoQ, +328% YoY), first-ever quarter above KRW 100 tn.b. Profit: OP of KRW 83.3 tn (+38% QoQ, +632% YoY), OPM 80% (+300bps QoQ); EBITDA of KRW 87.7 tn, margin 84%.c. Volume/price: DRAM bit shipments approx. +10% QoQ, server DDR5 shipments up 40%+ QoQ, ASP up north of 20%; NAND bit shipments down slightly QoQ, ASP up low-20s percent.d. Non-op: net FX loss of KRW 11.6 tn and investment valuation loss of KRW 22.9 tn, total non-op net expense KRW 32.6 tn; PBT KRW 50.7 tn, NP KRW 39.8 tn, NPM 38%; D&A KRW 4.3 tn.

4. Balance sheet and strategic investments

a. Cash KRW 114.5 tn (+KRW 26.5 tn QoQ), interest-bearing debt down KRW 2.8 tn to KRW 15.8 tn, net cash KRW 98.7 tn; debt-to-equity 5% (-200bps QoQ).b. Solidigm has terminated pre-IPO and IPO evaluations amid financial market volatility and a pullback in semiconductor stocks; expansion financing will proceed to balance both companies' shareholder value.

II. Detail from the call

2.1 Management remarks

1. Market view

a. Third-party estimates point to global CSP capex growth of 60%+ by 2027, with locked-in power capacity up 100%+. AI data center infra CAGR is approx. 40% through 2030.b. AI investment is broadening from a few tech giants to more enterprises and sovereigns; even with pockets of adjustment, overall momentum should remain strong.c. Near term, enterprise and consumer AI drive demand; over the medium to long term, autonomous driving, robotics and other physical AI take the baton. PC and smartphone units may decline on memory price hikes and tight supply, but will be more than offset by AI infra demand.d. Advanced node complexity and fab lead times mean supply will trail demand, and this constraint is expected to persist through 2030.

2. DRAM and HBM

a. HBM4 entered full mass production and shipment in Q3, with 12-high HBM4 delivered per agreed cadence with key customers; M15x ramping faster and to reach full run-rate by year-end.b. In Q4, HBM4 will surpass HBM3e by share of HBM shipments and revenue, becoming the flagship.c. Volume and pricing for 2027 HBM have been agreed with key customers, fully reflecting HBM4's product value, and the profitability gap vs. conventional DRAM will narrow meaningfully.d. For conventional DRAM, rapidly increase 1c-nm server mix in Q4, expanding 1c DDR5 and LPDDR5 SOCAMM for AI servers.

3. NAND

a. In Q4, lift 321-layer to 50%+ of domestic capacity, expanding supply of 321-layer enterprise SSDs.b. Over the medium to long term, scale high-performance, high-capacity eSSD for KV cache tiering, and through HBF plus system optimization, position NAND as a core solution for AI inference.

4. Long-term agreements (LTA)

a. Signed LTAs with 14 core customers including major global CSPs, typically five years, with some under extension discussions.b. Pricing structures differ by customer and product, generally reflecting market trends within agreed bands; safeguards such as deposits help ensure planned purchases become firm orders.c. LTA-locked volume already accounts for 50%+ of total capacity in 2027 and 2028.

2.2 Q&A

Q: AI investment keeps rising but monetization is slower. Why do you believe the momentum will sustain?

A: Two pillars: AI investment is translating into business results, and the investor base is structurally diversifying. What appears as pacing changes is customers assessing feasibility and returns project by project. Hyperscalers are posting cloud revenue growth of 30%–80% YoY, global cloud backlogs are up 2–5x, and paid users keep rising; distributed buildouts, smaller data centers, on-site power and cooling upgrades are easing constraints.Early investment was concentrated in a few large tech names, so any single change caused volatility; now non-tech enterprises, sovereign AI and on-prem deployments are joining fast, and isolated slowdowns are being backfilled. We expect a period of structural undersupply in memory to persist.

Q: How do you set the split between dividends and buybacks within total shareholder returns?

A: We first set dividends at a sustainable level, and deploy the rest to buybacks for cancellation. We reference earnings improvement, historical dividend yield, global peer yields and a prudent cash buffer. Most Kioxia gains will be paid out as special dividends, with a portion prepaid before year-end closing and the remainder included in the annual dividend.Management believes the current share price under-reflects earnings and cash flow, and buyback-cancellations are effective in lifting per-share value; buybacks will be executed quarterly within the annual envelope.

Q: What is the LTA lock-in and sales visibility over the next few years?

A: On top of LTAs, additional demand under discussion means DRAM and HBM output is effectively sold out through 2029. The volumes under discussion are tied to specific customer procurement plans, with high likelihood of conversion. LTAs account for 50%+ of total capacity in 2027–2028, and we aim to keep a similar ratio longer term to secure stable profitability; non-LTA capacity stays flexible for market dynamics and next-gen products.

Q: With ongoing fab builds and expansions, how do you mitigate oversupply risk if demand undershoots?

A: We invest in phases, building shells early and pacing tool-in and ramps with demand. More capex does not equal more supply: HBM and other high-performance products consume more resources, bit growth from scaling is increasingly limited, and investment is largely to enable node migrations and higher value-add products.At Yongin and M17, we are accelerating the fab shells and opening cleanrooms in stages. Demand assessments go beyond customer capex guides to include their financial health, funding capacity and key supply-chain readiness, alongside our own net cash targets; we will adjust capex, production plans and resource allocation as conditions evolve.

Q: HBM4 supply progress, and plans for HBM4E and custom HBM?

A: HBM4 yields are ahead of plan and targets have been raised, with Q4 share of shipments and revenue set to exceed HBM3e. Both front-end and back-end yields remain best-in-class, with back-end metrics covering bumping, stacking and full KGSD test; HBM3E has also surpassed the planned mature yield.HBM4E sampling was completed in Q2, with development and qualifications proceeding to customer timelines; custom HBM projects are underway with key customers.

Q: Some customers are reducing HBM stack heights. What is the impact on shipments and profitability?

A: Our medium- to long-term HBM outlook is unchanged, and customer demand remains more than 2x our current supply. Reducing layers is a strategy to secure more components under tight supply; even shifting from 12-high HBM4 to 8-high HBM4E, higher per-die capacity offsets most of the impact. Delivering the same bits requires more components, lifting base-die and back-end demand, which is already in our plans.Profitability depends not only on stack height but also on deal terms, product mix, yields, back-end efficiency and cost. In 2027, with higher HBM4 mix and pricing, HBM revenue and profit should grow sharply YoY and margins should improve notably, staying well above prior cycle peaks.

Q: Peers are advancing new cell architectures. How is your 4F² DRAM preparation and tech direction?

A: We will extend the 6F² platform with EUV as far as possible before transitioning to 4F² or 3D DRAM. The most effective path now is to leverage a mass-production-proven platform for reliability and cost to bolster product competitiveness.New cell architectures must balance performance, yield, reliability and cost, and also complete design and customer quals, so R&D timing does not equal commercialization timing; we have a stepwise roadmap from extreme scaling to 4F² and then 3D DRAM, to be introduced when both technology and economics are ready.

Q: KV cache is growing fast and being reused across memory tiers. What does this mean for NAND demand?

A: Tiered memory architectures will create new demand for storing and reusing inference context on NAND. Agents and long context windows make it hard to fit the entire KV cache into HBM or DRAM; SSDs are already offloading KV cache in some systems, and will evolve into a shared context memory tier across servers and GPUs.We are developing SALT KV (software that allocates data across HBM, DRAM and SSD based on reuse value and storage cost) and CMM Hybrid (CXL-based, storing the full KV cache on SSD and prefetching into DRAM).We also validate eSSD performance against real inference workloads. Over the medium to long term, we plan to insert HBF as a new tier between HBM and SSD.

<End of text>

Risk disclosure and disclaimer:Dolphin Research Disclaimer and General Disclosure