SK海力士 (000660.KS) 深度研究 — 商品DRAM引擎與倍數之爭

The commodity-DRAM engine, priced as a structural HBM story商品 DRAM 才是這輪暴利的引擎；共識已把 2028 畫成高原，剩下的戰場只是倍數

Key considerations主要考量

(1) The engine behind 2026's windfall is the price of conventional DRAM, not the structure of HBM. Conventional DRAM contract prices rose +93~98% in 1Q26 while HBM contract prices fell; DDR5 profitability actually overtook HBM3E in 2026, exactly as TrendForce called it back in October 2025 — and the company itself slowed its own HBM4 conversion line to chase the fatter commodity margin instead. Pricing the most cyclical variable in the industry as if it were a permanent structural story is the biggest mismatch of this whole cycle.

(1) 2026 這波暴利，引擎其實是 conventional DRAM 的「價格」，不是 HBM 的「結構」。1Q26 conventional DRAM 合約價漲 +93~98%，同期 HBM 合約價卻在跌；DDR5 的獲利性 2026 年真的超車了 HBM3E，正好應驗 TrendForce 早在 2025 年 10 月的預測 — 而公司自己也放慢了 HBM4 轉線，轉去追更肥的商品毛利。把產業裡最循環性的變數，當成一個永久的結構故事來定價，是這輪循環裡最大的錯位。

(2) The supply response to all this has already been paid for. Between the company's own $26.5B ADS raise, ₩30.9tn (≈$22B) of newly-committed capital spending, and Samsung/Micron/CXMT all expanding at the same time, the 2027-29 supply wave isn't a hypothetical risk scenario anymore — it's a funded construction schedule with dates attached.

(2) 供給端的反應已經全額付款了。自家 $26.5B 增資、新增 ₩30.9兆（≈$22B）資本承諾，加上三星、美光、CXMT 同時擴產 — 2027-29 的供給波已經不是假設情境，是一張標好日期、已經簽約的工程進度表。

(3) The bear argument half-cashed itself in on 07-29: the deceleration is real, and the beat streak this page once leaned on ended at 10 straight — 2Q26 OP ₩60.54tn (≈$43.4B) missed the ₩64.16tn consensus by -5.6%. What has moved since is capital allocation, not the operating model. ✗→ 08-21 On 08-19 the board approved a ₩40tn (≈$28.7B) buyback with full cancellation, running 08-20 to 11-19 — the largest share retirement ever executed by a Korean listed company, roughly 3.3% of shares outstanding — and lifted the 2025-27 shareholder-return floor from "within 50%" of cumulative free cash flow to "50% or more" ✓ disclosure + Korean press ×5公告＋韓媒×5. The stock rose +12.73% to ₩1,691,000 on 08-20, carrying the KOSPI up 5.89% with it. This page raises 2027E EPS ₩300,000→₩309,000 and the target ₩1,800,000→₩1,850,000 — the entire increase is the smaller denominator. Not one won of operating profit was added ◆.

(3b) Which is why the view does not improve. A +2.8% target against a +7.9% price leaves +9.4% of implied upside, and the framework keeps saying range-bound — the Buy line now sits at ₩1,608,700 and the close is 5.1% above it. Two things deserve saying plainly rather than being cheered. First, the announcement raised the price the company itself must now pay: at ₩1,691,000 rather than the ₩1,662,000 reference, ₩40tn retires ~23.7M shares instead of ~24.07M. Second, this is a buyback of 3.3% of the equity at ≈4.7× book, funded from peak-cycle cash, by a company whose own capex schedule — and this page's bear case — describes a 2027-29 supply wave arriving into it. Retiring stock near a cyclical top is how memory balance sheets have been damaged before. The arithmetic is accretive; whether the timing is, is the same question the 07-29 capital raise asked from the other direction ◆.

(3) 空方論據在 07-29 兌現了一半：減速是真的，本頁曾倚靠的 beat 紀錄停在連 10 季 — 2Q26 OP ₩60.54兆（≈$43.4B）低於共識 ₩64.16兆達 -5.6%。但這之後動的是資本配置，不是營運模型。✗→ 08-21 08-19 董事會通過 ₩40兆（≈$28.7B）庫藏股取得並全數註銷，執行期 08-20 至 11-19 — 這是韓國上市公司史上最大規模的股份註銷，約占流通股數 3.3% — 同時把 2025-27 年股東回饋下限，從累計自由現金流的「50% 以內」上調為「50% 以上」✓ 公告＋韓媒×5。股價 08-20 大漲 +12.73% 至 ₩1,691,000，並帶動 KOSPI 單日 +5.89%。本頁因此將 2027E EPS 自 ₩300,000 上調至 ₩309,000、目標價自 ₩1,800,000 上調至 ₩1,850,000 — 上調的全部來自分母變小。營業利益一塊錢都沒有加 ◆。

(3b) 也正因如此，觀點沒有變好。目標價 +2.8%、股價 +7.9%，剩下的隱含空間是 +9.4%，框架依然說「區間」 — 買進線現在落在 ₩1,608,700，收盤高出它 5.1%。有兩件事該平鋪直敘地說，而不是跟著歡呼。第一，這則公告抬高了公司自己接下來要付的價格：以 ₩1,691,000 而非參考價 ₩1,662,000 計算，₩40兆買回的是約 23.7M 股，而不是約 24.07M 股。第二，這是在約 4.7 倍股價淨值比、用循環高點的現金，買回 3.3% 的股權 — 而這家公司自己的 capex 時程、以及本頁的空方情境，講的都是 2027-29 那道正在逼近的供給波。在循環頂附近註銷自家股票，正是記憶體業資產負債表過去受過傷的方式。算術上它是增益的；時點上是不是，跟 07-29 那次增資從反方向問的，是同一個問題 ◆。

(4) Two research items sit alongside the buyback and neither is a 2027 earnings input. The High Bandwidth Flash (HBF) standard published with Sandisk(SNDK.O) through the Open Compute Project has no customer and no volume before ~2030, with Samsung and Kioxia contesting the same format. And on 08-20 the company published a co-packaged optics (CPO) roadmap in Nature Electronics — moving the memory interface itself onto light. It is a peer-reviewed research roadmap with academic co-authors, no product, and no date. Both are optionality, and optionality is not a catalyst ◆. Full treatment in the HBF section and the CPO section.

(4) 庫藏股之外還有兩則研究面消息，兩則都不是 2027 的獲利輸入項。與 Sandisk(SNDK.O) 透過 Open Compute Project 發布的 High Bandwidth Flash（HBF，高頻寬快閃記憶體）標準，沒有客戶、2030 年前沒有量，三星與 Kioxia 還在爭同一個格式。而 08-20 公司在 Nature Electronics 發表了 CPO（Co-Packaged Optics，共同封裝光學）技術藍圖 — 把記憶體介面本身搬上光。那是一份有學界共同作者的同儕審查研究藍圖：沒有產品，也沒有時程。兩者都是選擇權，而選擇權不是催化劑 ◆。完整處理見 HBF 章節與 CPO 章節。

TP:目標價： This analysis models 2026/2027 EPS at ₩343,000 (incl. one-offs) / ₩300,000; 6-month target price ₩1,800,000 (prior: ₩1,960,000), based on 2027E EPS ₩300,000 × 6× PER. The cut is 100% earnings-driven — 2Q26 actuals rolled in and the 2H26/2027 path trimmed ~-7~8% — while the 6× multiple is unchanged: the AI-era forward PER band runs roughly 4–10×, and a below-median 6× was granted because this analysis expected the market to price the 2028 rollover 6-12 months early. The 07-29 tape suggests the market moved that schedule up by a full year. Revised at the 08-21 refresh: the ₩40tn cancellation removes ~3.3% of the share count, so 2027E EPS rises to ₩309,000 and the target to ₩1,850,000 (₩309,000 × 6× = ₩1,854,000, rounded). The 6× multiple is again untouched, and so is every line of the operating model — revenue, operating profit and margin for 2026E and 2027E are all carried forward unchanged. This is a denominator revision, and it is the only kind of upgrade a buyback can honestly produce. 本研究預估 2026、2027 年 EPS 為 ₩343,000（含一次性）、₩300,000，目標價 ₩1,800,000（6 個月，07-29 自 ₩1,960,000 下修），基於 2027E EPS ₩300,000 × 6 倍 PER。當時下修 100% 來自獲利面 — 2Q26 實績滾入、2H26/2027 路徑下修約 -7~8% — 6 倍 PER 不變：AI 世代 forward PER 區間約 4–10 倍，給予中值以下的 6 倍，係因本研究原就預估市場會提前 6–12 個月定價 2028 獲利轉平；07-29 的走勢顯示，市場把這張時程表整整提前了一年。08-21 更新調整：₩40兆註銷會移除約 3.3% 的股數，因此 2027E EPS 上調至 ₩309,000、目標價上調至 ₩1,850,000（₩309,000 × 6 倍＝ ₩1,854,000，取整）。6 倍 PER 一樣沒有動，營運模型也一行都沒有動 — 2026E 與 2027E 的營收、營業利益、營益率全數延用。這是一次分母的修正，而這也是一次庫藏股在誠實前提下唯一能產生的上調。

Risk factors: China's annual VEU licensing review · a Samsung HBM4 same-price war · the 2027 supply wave · won/KOSPI concentration risk.

風險因子：中國 VEU 年審 · 三星 HBM4 同價戰 · 2027 供給波 · 韓元/KOSPI 集中度。

Sources for this section (consolidated here instead of inline): price & valuation = KRX closes via Korean press ×5 + computed (08-20) · buyback = board disclosure 08-19 via Korean press ×5 · CPO = Nature Electronics via company newsroom (08-20) · 2Q26 actuals = official release + sedaily (07-29) · contract prices = TrendForce (03-31/06-01/07-03) · HBF = OCP spec via SK hynix + Sandisk releases (08-03) · consensus = FnGuide (07-10) / S&P Global via stockanalysis (07-22), both pre-print · financials = SEC 424B4 (07-10) · ADR = IBKR (08-20) · estimates, TP and view = this analysis ◆. Per-number ledger: see the Sources section.本節出處（統一註記於此，不逐句插標）：價格與估值＝KRX 收盤價（韓媒×5）＋本研究計算（08-20）· 庫藏股＝08-19 董事會公告，韓媒×5 · CPO＝Nature Electronics，公司 newsroom（08-20）· 2Q26 實績＝官方 release＋sedaily（07-29）· 合約價＝TrendForce（03-31/06-01/07-03）· HBF＝OCP 規格，SK hynix 與 Sandisk release（08-03）· 共識＝FnGuide（07-10）／S&P Global via stockanalysis（07-22，皆為財報前）· 財務＝SEC 424B4（07-10）· ADR＝IBKR（08-20）· 預估、目標價與觀點＝本研究判斷 ◆。逐數字查證表見〈資料來源〉章。

The full case — business & tech · moat · bull vs bear · variant · management · competition (click to expand)完整論證 — 事業與技術・護城河・多空對決・變異認知・經營層・競爭（點開展開）

The argument in three paragraphs論點全文 — 三段講完

Start with what's actually moving the profit line: plain old commodity DRAM. Contract prices for conventional DRAM jumped +93~98% quarter-on-quarter in 1Q26 and roughly +58~63% again in 2Q26 (TrendForce) — while HBM contract prices, the part everyone talks about, actually fell over the same stretch. The 2Q26 print (07-29) then made the deceleration official: the company's blended DRAM ASP rose ~+30% QoQ, down from 1Q's +60-70% pace. The number carrying this whole profit boom is the one that historically snaps back fastest — and it has now visibly started slowing.

先講清楚真正在推動獲利的是什麼：是最普通的商品 DRAM。1Q26 conventional DRAM 合約價季增 +93~98%，2Q26 再漲約 +58~63%（TrendForce）— 反而是大家都在談的 HBM，同一段時間合約價是下跌的。而 07-29 的 2Q26 財報，把「減速」變成了官方數字：公司混合 DRAM ASP 季增約 +30%，明顯低於 1Q 的 +60-70%。撐起這整波獲利的這個數字，恰好是歷史上最容易均值回歸的那個 — 而它已經肉眼可見地慢下來了。

Sell-side has already quietly drawn the ceiling: pre-print consensus operating profit sits at ₩386tn (≈$277B) for 2027 and ₩383tn (≈$275B) for 2028 — basically flat. Even the analysts who are bullish won't extrapolate growth past next year. This analysis goes further and models 2027E OP at ₩275tn (≈$197B; revised down from ₩300tn on 07-29), another 29% below that.

賣方自己其實已經悄悄畫出天花板了：財報前共識 2027 年營業利益 ₩386兆（≈$277B）、2028 年 ₩383兆（≈$275B） — 幾乎是持平的兩個數字。連看多的分析師都不敢把成長線繼續往 2028 拉。本研究比共識更保守，2027E OP 估在 ₩275兆（≈$197B，07-29 自 ₩300兆下修），再低 29%。

And the stock's two valuation languages have been re-priced repeatedly inside a month — down, then back up, then up again on the buyback. At the 08-20 close the stock sits -43.4% off the June peak and trades at 5.5× this analysis's 2027E EPS; PBR, still in the 8× range in late July, is ≈4.7× — partly the price, partly the ₩63tn (≈$45.4B) Kioxia gain landing in book value.

而這檔股票的兩套估值語言，在一個月內被反覆重新定價 — 先往下、再走回來，然後因庫藏股再往上一段。08-20 收盤距 6 月高點 -43.4%，相當於本研究 2027E EPS 的 5.5 倍；7 月下旬還在 8× 檔的 PBR 現為約 4.7× — 一半是股價、一半是 ₩63兆（≈$45.4B）Kioxia 利益進了帳面淨值。

The analyst reset this page said hadn't started has now started — and it split instead of converging. In the 24 hours after the print, seven houses cut, two raised and nine held; the range widened to ₩1,480,000–₩4,700,000, a high-to-low ratio of 3.2× on the same company in the same week ✓ g-enews roster글로벌이코노믹 名單 07-30. The average across that roster of 18 is ≈₩3,296,000 ⚠ computed from one article's roster, not the full 37-analyst universe依單一報導名單計算，非完整 37 位樣本 — still +110% above the tape. That dispersion is the thesis in one number: when a fight is about the multiple rather than the direction, the street cannot converge, because there is no multiple everyone agrees a plateau deserves.

本頁原本說「根本還沒開始」的分析師重估，現在開始了 — 而且是分裂，不是收斂。財報後 24 小時內，7 家下修、2 家上修、9 家維持；目標價區間拉開到 ₩1,480,000–₩4,700,000，同一家公司、同一週，最高是最低的 3.2 倍 ✓ 글로벌이코노믹 名單 07-30。這份 18 家名單的平均值約 ₩3,296,000 ⚠ 依單一報導名單計算，非完整 37 位樣本 — 仍高於現價 +110%。這個離散度本身就是論點的濃縮版：當爭的是倍數而不是方向，街上就沒辦法收斂 — 因為沒有人同意「高原」該值幾倍。

Business & technology事業與技術

Finding: DRAM is 77.3% of revenue, and the mix has shifted 10pp toward it in two years. The 1Q26 424B4 breakdown shows DRAM 77.3% (₩40.66tn), NAND 22.0% (₩11.57tn), other 0.7% ✓ 424B4. FY25 was DRAM 77.1%/NAND 21.3%; FY24 was 67.6%/29.1% ✓ 424B4 — the swing toward DRAM is driven by HBM and server DDR5, not NAND.

發現：DRAM 佔營收 77.3%，且兩年內組合已向 DRAM 位移 10 個百分點。1Q26 424B4 揭露 DRAM 77.3%（₩40.66兆）、NAND 22.0%（₩11.57兆）、其他 0.7% ✓ 424B4。FY25 為 DRAM 77.1%/NAND 21.3%；FY24 為 67.6%/29.1% ✓ 424B4 — 向 DRAM 位移主因是 HBM 與伺服器 DDR5，不是 NAND。

Segment mix (T5)產品組合（T5）

Segment段別	FY25 revFY25 營收	% of total占比	FY26E	% of total占比	YoY
DRAM	₩74.90兆 ✓	77.1%	₩272兆 ◆	78.4% ◆	+263% ◆
NAND	₩20.69兆 ✓	21.3%	₩73兆 ◆	21.0% ◆	+253% ◆
Other其他	₩1.55兆 ✓	1.6%	₩2兆 ◆	0.6% ◆	—

Value chain價值鏈

Upstream: equipment/materials (ASML EUV, Applied Materials, etc.) → SK hynix (IDM: design + manufacturing + packaging) → customers: NVIDIA(NVDA.O) at 14.8% of 1Q26 revenue (₩7.78tn) ⚠ reported ratio, secondary轉述比例, an undisclosed North American hyperscaler crossing 10% for the first time at 12.4% ⚠, plus CSPs and server OEMs broadly. Pull HBM out of the chain and GPUs cannot ship — but pricing power sits inside long-term agreements, not a spot auction.

上游：設備/材料（ASML EUV、應材等）→ SK海力士（IDM：設計+製造+封裝）→ 客戶：NVIDIA(NVDA.O) 1Q26 佔營收 14.8%（₩7.78兆）⚠ 轉述比例，某北美超大規模客戶首度 >10%，佔 12.4% ⚠，加上廣泛的 CSP 與伺服器 OEM。HBM 抽掉，GPU 就無法出貨 — 但議價力在長約裡，不在現貨拍賣。

Unit economics單位經濟

Contract price leads, spot is a sideshow. 1Q26 companywide OPM 72% ✓; DRAM ASP rose mid-60%s QoQ ✓ TrendForce. HBM4 12-layer parts are reported above $600/unit, yet HBM contract price fell in 1Q26 — suppressing ASP growth even as headline unit prices climb ⚠✓.

合約價為主、現貨為輔。1Q26 全公司 OPM 72% ✓；DRAM ASP 單季 +mid-60% QoQ ✓ TrendForce。HBM4 12層報導單價 >$600/顆，但 HBM 合約價 1Q26 反而下跌 — 壓抑 ASP 成長,即使頭條單價在漲 ⚠✓。

HBM cross-section — the physical stack next to the GPUHBM 剖面 — GPU 旁的堆疊結構

✓ SK hynix newsroom + EETimes MR-MUF's thermal bumps run ~4× denser than TC-NCF, the differentiator behind SK hynix's HBM lead. HBM4E 12-layer samples shipped 2026-06-18 at 16Gbps/pin, 48GB/stack.MR-MUF 熱阻尼凸塊密度較 TC-NCF 多約 4 倍，是 SK海力士 HBM 領先的差異化來源。HBM4E 12 層樣品已於 2026-06-18 出樣，規格 16Gbps/pin、48GB/stack。

Demand transduction chain — what actually turns into volume需求傳導鏈 — 什麼會放量

Agentic-AI token consumption is exploding → inference compute deployment (GPU+ASIC) → three channels absorb it (HBM4 content per GPU, high-capacity server DDR5, eSSD) → CSPs lock in long-term agreements out of shortage fear → contract prices spike → operating margin runs 72%→77%. Structural vs cyclical verdict: the volume expansion is structural, but 2026's profit explosion is driven mainly by price, not volume — conventional DRAM contract price rose +93~98% QoQ in 1Q26 and +58~63% in 2Q26E ✓ TrendForce — and price is the most mean-reverting variable in this industry.

Agentic AI（代理式 AI）token 消耗爆炸 → 推論算力部署（GPU+ASIC）→ 三通道吸收：每 GPU HBM4 搭載量、伺服器高容量 DDR5、eSSD → CSP 因恐懼短缺而鎖長約 → 合約價飆漲 → 營益率 72%→77%。結構 vs 循環判定：放量是結構的，但 2026 獲利爆發的主力是「價格」而非「量」— conventional DRAM 合約價 1Q26 QoQ +93~98%、2Q26E +58~63% ✓ TrendForce — 而價格是這個產業裡最會均值回歸的變數。

Volume growth (top: HBM4/DDR5/eSSD, ✓ sourced per the grounding table) is real, but the swing factor behind 2026's margin is the price box in amber — the most cyclical link in the chain.放量（上排：HBM4/DDR5/eSSD，均有來源，詳資料來源章節）是真的，但撐起 2026 獲利的擺動因子是琥珀色的價格區塊 — 整條鏈裡最循環性的一環。

編輯示意圖：無數token符號如洪流般湧向一排排伺服器機櫃，象徵AI推論需求對記憶體的壓力 — AI-generated illustration of token demand flooding into server racks.AI 生成示意圖：token 需求洪流湧向伺服器機櫃。

✓ newsroom 2026-07-22 · IEDM 2025官方newsroom 2026-07-22·IEDM 2025Latest research: CTI — a new partition for NAND scaling最新研究：CTI — NAND 微縮的新隔板

NAND scaling has shifted from "stack higher" to "shrink the cell at the same stack height, then pack more layers" — but tighter cell spacing causes electrical interference: electrons drift between neighboring cells, shifting the threshold voltage (Vth) and causing read errors. SK hynix's NAND stores charge in a CTN (charge trap nitride) layer, which has traditionally run as one continuous film — the company's own metaphor: "a long table with no partitions, where books (electrons) slide into the neighboring seat." CTI adds the partition.

NAND 微縮已從「疊更高」轉向「同高度內縮小 cell、加密層數」，但 cell 間距縮小 → 電場互擾、電子游移 → 閾值電壓（Vth）漂移、讀取錯誤。SK 的 NAND 用 CTN（charge trap nitride，把電子存在絕緣氮化層）— 傳統上 CTN 是連續層，官方比喻：「沒有隔板的長桌，書（電子）會滑到鄰座」。CTI＝加隔板。

The industry has tried for 10+ years without reaching production: traditional isolation etches pockets, which widens the channel hole and cuts pore density per unit area, plus uneven etching. SK hynix instead deposits an oxide barrier and then backfills the CTN — no etching, no pore-widening — achieving, for the first time industry-wide, isolation on a 176-layer production-grade device: cell-to-cell interference down 30%+, charge retention up 45%+, and cell size shrunk a further 10%+ (all company-stated research results) ✓ 官方newsroom 07-22·IEDM 2025.

業界研究 10 餘年未能量產：傳統 CTI 靠蝕刻挖口袋 → channel hole 變寬 → 單位面積孔數變少、密度受損，加上蝕刻不均。SK 改以沉積氧化屏障後回填 CTN — 不蝕刻、不撐孔 — 業界首次在 176 層生產級元件實現隔離：cell 間干擾 -30%以上、電荷保持力 +45%以上、cell 尺寸再縮 >10%（皆公司自述之研究成果）✓ 官方newsroom 07-22·IEDM 2025。

This is a research milestone on the path toward production, not a shipping product; the three result figures above are the company's own self-reported research data, not independently audited.屬研究里程碑（邁向量產），非出貨中產品；成果數據為公司自述。

In other words: retention and interference improvement is exactly the property QLC — 4 bits/cell, 16 Vth states, the narrowest margin between states — needs most → this directly supports Solidigm/eSSD's AI-storage story and the durability of its cost/GB curve.

換言之：保持力/干擾改善正是 QLC（4 bit/cell、16 個 Vth 態、分布最窄）最需要的特性 → 直接支撐 Solidigm/eSSD 的 AI 儲存故事與 cost/GB 曲線延續。

✓ SK hynix newsroom "Inside the Research" Ep.1 (2026-07-22) Schematic, not to lithographic scale — three stacked cells shown for clarity.示意圖，非微影比例尺 — 為求清楚僅畫出三層堆疊示意。資料來源：SK hynix newsroom「Inside the Research」Ep.1（2026-07-22）。

HBF — the first standard, published 2026-08-03HBF — 第一份標準，2026-08-03 發布

The problem it exists to solve. An AI inference server has two kinds of memory and a canyon between them. HBM is fast — roughly 2TB/s of bandwidth sitting right beside the GPU — but a stack only holds tens of gigabytes, and it is expensive per bit. An enterprise SSD holds tens of terabytes cheaply, but talks to the processor across a storage interface at a few GB/s, hundreds of times slower. Modern inference workloads (long-context models, KV caches, retrieval over large corpora) want something that does not exist: HBM-class bandwidth at SSD-class capacity. Everything the model cannot fit in HBM has to be fetched across that canyon, and the fetch — not the arithmetic — is what the GPU waits on. That wait is what the industry calls the memory wall.

它要解決的問題。一台 AI 推論伺服器有兩種記憶體，中間隔著一道峽谷。HBM 很快 — 就貼在 GPU 旁邊，頻寬約 2TB/s — 但一疊只裝得下幾十 GB，而且每一個 bit 都貴。企業級 SSD 便宜地裝下幾十 TB，卻要走儲存介面跟處理器對話，速度只有幾 GB/s，慢上百倍。現在的推論工作負載（長上下文模型、KV cache、大語料檢索）想要一個目前不存在的東西：HBM 等級的頻寬，配上 SSD 等級的容量。凡是塞不進 HBM 的東西，都得跨過那道峽谷去搬 — 而 GPU 等的是「搬運」，不是「算」。這個等待，業界叫做記憶體牆（memory wall）。

What HBF does. High Bandwidth Flash takes the physical trick that made HBM work — stack the dies vertically, drill through them, and put the whole tower next to the processor instead of across a bus — and applies it to NAND instead of DRAM. Because NAND stores far more bits per die, a stack lands at up to 512GB where an HBM stack lands at tens of GB; because it is non-volatile, it needs no refresh power at all, which DRAM burns continuously just to remember. The trade is latency: NAND reads are inherently slower than DRAM reads, so HBF is not an HBM replacement — it is a new tier underneath HBM and above the SSD.

HBF 做的事。High Bandwidth Flash（高頻寬快閃記憶體）把當初讓 HBM 成立的那個物理招式 — 把晶片垂直堆起來、打洞穿過去、整座塔放在處理器旁邊而不是隔著匯流排 — 拿來用在 NAND 上，而不是 DRAM。因為 NAND 每顆 die 能存的 bit 多得多，一疊可以做到 512GB，而 HBM 一疊只有幾十 GB；又因為它是非揮發性的，完全不需要 refresh 電力 — 而 DRAM 光是為了「記得」就得持續燒這份電。代價是延遲：NAND 的讀取天生比 DRAM 慢，所以 HBF 不是拿來取代 HBM 的 — 它是 HBM 下面、SSD 上面的一個新階層。

✓ OCP spec via SK hynix + Sandisk releasesOCP 規格，引自 SK hynix 與 Sandisk release 2026-08-03 Schematic; the left panel is a positioning map, not a plotted dataset. HBM4 reference figures are generation-typical, not a spec quotation.示意圖；左圖為定位示意，非實際數據作圖。HBM4 的對照數字為該世代的一般值，非規格書引述。

What was actually published. The specification covers the xPU-to-HBF host interface and its electrical characteristics, reliability and packaging guidance for the die stack, and a software guide for read and write operations — three grades spanning roughly 0.4TB/s to 3.0TB/s, 8-high and 16-high stacks up to 512GB, over UCIe as the chiplet interconnect ✓ OCP spec, both partners' releasesOCP 規格，雙方 release. The consortium — SK hynix, Sandisk(SNDK.O), Google(GOOGL.O) and Tenstorrent — formed in 2026-02 and published inside the Open Compute Project rather than keeping the format proprietary ✓. Two SK hynix executives keynoted FMS 2026 on 08-04 (Santa Clara, 08-04~06), where the company also showed its 10th-generation 375-layer 4D NAND for the first time ✓.

實際發布的內容。這份規格涵蓋 xPU 對 HBF 的主機介面與電氣特性、die stack 的可靠度與封裝指引、以及讀寫作業的軟體指南 — 三個等級橫跨約 0.4TB/s 至 3.0TB/s，8 層與 16 層堆疊、單疊最高 512GB，以 UCIe 作為 chiplet 互連 ✓ OCP 規格，雙方 release。這個聯盟 — SK hynix、Sandisk(SNDK.O)、Google(GOOGL.O)、Tenstorrent — 2026-02 成立，並選擇在 Open Compute Project 內公開，而不是把格式留成專有規格 ✓。兩位 SK 高層於 08-04 在 FMS 2026（Santa Clara，08-04~06）發表 keynote，公司也在該場首度展出第十代 375 層 4D NAND ✓。

What this is not: a 2027 number它不是什麼：不是 2027 年的數字 ◆

A roadmap does exist, but it predates this spec and it is a technical one: Sandisk has targeted first HBF samples in 2H CY2026 and samples of the first AI-inference devices carrying HBF in early 2027, dating from the original August 2025 partnership ✓ TrendForce + trade pressTrendForce＋產業媒體. What the 08-03 release does not contain is the commercial half — no named customer, no price, no volume commitment — and both the partners' February framing and Korean coverage of this week's announcement put full commercialisation at around 2030 ✓ SK newsroom 2026-02 + Seoul Economic Daily 08-04官方 newsroom 2026-02＋首爾經濟 08-04. Samples in 2026, revenue that matters at the end of the decade: that gap is the whole point.

時程表是存在的，但它早於這份規格，而且是技術時程：Sandisk 的目標是 2026 下半年出第一批 HBF 樣品、2027 年初出第一批搭載 HBF 的 AI 推論裝置樣品 — 這個時程來自 2025 年 8 月最初的合作案 ✓ TrendForce＋產業媒體。08-03 的發布真正沒有的是商業那一半 — 沒有具名客戶、沒有價格、沒有量的承諾 — 而雙方 2 月的說法與本週韓媒報導，都把全面商業化放在約 2030 年 ✓ 官方 newsroom 2026-02＋首爾經濟 08-04。2026 年出樣、2030 年才有像樣的營收：這中間的落差，就是重點所在。

And SK hynix does not own this format. Sandisk signed HBF standardisation MOUs with both SK hynix and Samsung, and Samsung — the NAND share leader — began conceptual design and early HBF development back in 2025-10; Samsung and Sandisk are reported to target integration into NVIDIA, AMD and Google products in late 2027 to early 2028. Kioxia, meanwhile, demonstrated a 5TB high-speed HBF prototype in August 2025, a year before this specification existed ✓ TrendForce ×3 / DigiTimes, reportedTrendForce×3／DigiTimes，轉述 ⚠ competitor plans are press-reported, not company-confirmed競爭者計畫屬媒體轉述，非公司證實. Publishing through the OCP is therefore best read as a land-grab in a contested field, not as a moat: you open a standard when you want it adopted before somebody else's becomes the default ◆.

而且 SK 海力士並不擁有這個格式。Sandisk 在 HBF 標準化上，跟 SK 海力士和三星「都」簽了 MOU；而 NAND 份額龍頭三星早在 2025-10 就已展開 HBF 的概念設計與早期開發，並據報導與 Sandisk 一同瞄準 2027 年底至 2028 年初把 HBF 導入 NVIDIA、AMD、Google 的產品。另一邊，Kioxia 早在 2025 年 8 月就展示過 5TB 高速 HBF 原型 — 比這份規格早了整整一年 ✓ TrendForce×3／DigiTimes，轉述 ⚠ 競爭者計畫屬媒體轉述，非公司證實。因此，選擇透過 OCP 公開，比較合理的讀法是「在一塊競爭中的地上先插旗」，而不是護城河：你會把標準開放出來，是因為你想在別人的版本變成預設之前先被採用 ◆。

The only dated NAND item in this week's announcement is a different one: mass production of high-capacity eSSDs built on the 375-layer 4D NAND, guided for early 2027 ✓ — a cost-curve story for the existing eSSD business, not HBF.

本週發布中唯一有日期的 NAND 項目是另一件事：以 375 層 4D NAND 打造的高容量 eSSD，指引 2027 年初量產 ✓ — 那是現有 eSSD 事業的成本曲線故事，不是 HBF。

Where it does count: NAND is the weaker leg of this company at 22.0% of revenue, and its structural problem has always been that bits get cheaper faster than demand absorbs them. HBF is an attempt to sell NAND bits into a bandwidth-priced socket instead of a capacity-priced one. If it lands, it re-rates the segment that this page treats as the commodity half of the business. That is worth a line in the optionality column and nothing in the 2027E model.

它真正算數的地方：NAND 是這家公司比較弱的那條腿（占營收 22.0%），而它的結構問題一直是 bit 變便宜的速度快過需求吃掉它的速度。HBF 是一次嘗試 — 把 NAND 的 bit 賣進一個「按頻寬計價」的位置，而不是「按容量計價」的位置。真的成了，被重估的會是本頁一直當成商品那一半的部門。這值得在「選擇權」欄位寫一行，在 2027E 模型裡寫零。

One structural note ◆. SK hynix booked ₩63.27tn (≈$45.4B) of gains in 2Q26 exiting its Kioxia stake — and is now co-authoring the HBF standard with Sandisk, which has been Kioxia's NAND manufacturing joint-venture partner for 25 years (Yokkaichi and Kitakami, extended through 2034 in 2026-01) ✓ Sandisk/Kioxia releasesSandisk／Kioxia release 2026-01. Reading a strategy into that sequencing is interpretation, not disclosure: the Kioxia sale was a financial exit from an investment consortium, and the Sandisk work is a standards alliance. They are different instruments. But the direction is consistent — the company converted a passive stake in a NAND rival into cash, and spent its NAND attention on defining the next format instead. The sharper edge is that Kioxia is not a bystander here: it demonstrated its own 5TB HBF prototype in 2025-08 ✓ reported轉述. So SK hynix sold out of a company that is now a direct competitor in the format SK is trying to standardise — with that company’s own manufacturing partner as SK’s co-author. Whatever the intent, the alignments in NAND are not tidy, and no one should model them as if they were.

一個結構性註記 ◆。SK 海力士在 2Q26 認列 ₩63.27兆（≈$45.4B）利益、出清 Kioxia 持股 — 而現在，它正和 Sandisk 共同撰寫 HBF 標準，而 Sandisk 已經當了 Kioxia 25 年的 NAND 製造合資夥伴（四日市與北上兩廠，2026-01 延長至 2034 年）✓ Sandisk／Kioxia release 2026-01。從這個先後順序讀出一套戰略，屬於詮釋、不是公司揭露：Kioxia 那筆是從投資聯盟財務性退出，Sandisk 這邊是標準聯盟，兩者是不同的工具。但方向是一致的 — 公司把一個 NAND 對手的被動持股換成現金，然後把 NAND 的注意力花在定義下一個格式上。更尖銳的一點是：Kioxia 在這件事上不是旁觀者 — 它早在 2025-08 就展示過自家的 5TB HBF 原型 ✓ 轉述。也就是說，SK 海力士賣掉的那家公司，如今是它想標準化的這個格式裡的直接競爭者 — 而那家公司的製造夥伴，正是 SK 這份規格的共同作者。不論本意為何，NAND 這一塊的結盟關係並不乾淨，任何人都不該把它當成乾淨的來建模。

CPO — putting the memory interface on light, published 2026-08-20CPO — 把記憶體介面搬上光，2026-08-20 發表

On 08-20 the company published a co-packaged optics (CPO) roadmap in Nature Electronics, co-authored with the University of Virginia, UIUC, Nanyang Technological University, MIT and Yonsei ✓ company newsroom + The Elec + Seoul Economic Daily, 08-20公司 newsroom＋The Elec＋首爾經濟，08-20.

08-20 公司在 Nature Electronics 發表了一份 CPO（Co-Packaged Optics，共同封裝光學）技術藍圖，共同作者包括維吉尼亞大學、UIUC、南洋理工、MIT 與延世大學 ✓ 公司 newsroom＋The Elec＋首爾經濟，08-20。

The problem it names. Compute throughput has been rising about 3× every two years; the interconnect bandwidth feeding it rises only about 1.4× over the same period. Those two rates compound apart, and the gap is what the field calls the bandwidth wall — a point where adding GPUs stops adding useful work because the wires between them cannot carry the data fast enough. It is a systems problem, not a chip problem, which is the paper's actual thesis: the competition is moving from chips to systems.

它指出的問題。算力大約每兩年成長 3 倍；餵給它的互連頻寬同期只成長約 1.4 倍。這兩個速率會複利地愈拉愈開，而這個差距就是業界所說的頻寬牆 — 到了某個點，再加 GPU 也不會增加有用的產出，因為它們之間的線路來不及把資料送過去。這是系統的問題，不是晶片的問題 — 而這正是這篇論文真正的主張：競爭正在從晶片轉移到系統。

What CPO does about it. Today a processor talks to distant memory over copper: electrical signals down long traces, where power burned and signal lost both grow with distance. CPO puts the optical transceiver inside the same package as the processor, so the link leaves the package as light rather than as electricity. Light does not degrade over board-level distance the way a copper trace does, and it costs far less energy per bit to move. The paper's targets are explicit: more than 100 Tb/s per node, under 1 picojoule per bit, and chip-to-chip latency under 10 nanoseconds.

CPO 怎麼處理它。今天處理器要跟遠端記憶體溝通，走的是銅：電訊號沿著長線路傳送，而耗掉的功率與損失的訊號都隨距離增加。CPO 把光收發器放進與處理器同一個封裝內，因此連結離開封裝時是光、不是電。光不會像銅線那樣在板級距離上衰減，而且每搬一個 bit 所耗的能量低得多。論文的目標寫得很明確：每個節點超過 100 Tb/s、每 bit 低於 1 皮焦耳、晶片間延遲低於 10 奈秒。

Left: the two growth rates the paper starts from — compute roughly 3× every two years against interconnect bandwidth roughly 1.4×, drawn on a log scale so the compounding gap (shaded) is the bandwidth wall itself. Right: the three engineering targets the roadmap sets. Bottom: today a processor reaches distant memory over a long copper trace, where both power draw and signal loss scale with distance; under CPO the transceiver sits inside the package and the link leaves as light, letting several accelerators share one pooled memory resource. The bottom row is the paper's end-state architecture, not a product.左：論文的起點是兩條成長率 — 算力約每兩年 3 倍，對上互連頻寬約 1.4 倍，畫在對數尺度上，複利拉開的那塊陰影就是頻寬牆本身。右：藍圖設定的三個工程目標。下：今天處理器要碰到遠端記憶體，得走一條長銅線，耗電與訊號損耗都隨距離放大；在 CPO 之下，收發器坐進封裝內、連結以光離開，讓多個加速器可以共用同一座記憶體資源池。下面這一排是論文的終局架構，不是產品。

What this is not: a 2027 number either它也不是什麼：一樣不是 2027 年的數字 ◆

This is a peer-reviewed roadmap with five universities on the author list. That is a real signal about research depth, and it sits naturally beside the CTI work and the 2026 IEEE Corporate Innovation Award — this company publishes. It is not a signal about revenue. There is no product, no customer, no sampling date and no volume commitment attached to it, and CPO in the industry at large is still at the stage of moving optics next to switch ASICs, not next to memory. The roadmap's own sequencing — 2D packaging, then 2.5D interposers, then 3D heterogeneous integration — describes years of work, and the paper does not date any of it.

這是一份有五所大學掛名的同儕審查藍圖。那是關於研究深度的真實訊號，也跟 CTI 研究、2026 IEEE Corporate Innovation Award 放在一起很自然 — 這家公司是會發論文的。但它不是關於營收的訊號。它沒有產品、沒有客戶、沒有出樣日、沒有量的承諾，而且整個產業的 CPO 目前還停在「把光學搬到交換器 ASIC 旁邊」的階段，還沒到記憶體旁邊。藍圖自己的順序 — 2D 封裝、再到 2.5D 中介層、再到 3D 異質整合 — 講的是好幾年的工程，而論文一個日期都沒給。

Where it would count, if it lands: pooled memory shared across accelerators attacks the same structural problem HBM solves by stacking — how to put more bandwidth in front of a GPU — but it attacks it at the rack level rather than the package level. A memory maker that defines that interface sells into a bigger socket. Treat it exactly as this page treats HBF: a line in the optionality column, and nothing in the 2027E model. Two optionality items in one month is not a trend; it is a research department doing its job ◆.

如果真的成了，它算數的地方：跨加速器共享的記憶體池，攻擊的是 HBM 用堆疊解決的同一個結構問題 — 怎麼把更多頻寬送到 GPU 面前 — 但它是在機櫃層級攻擊，不是在封裝層級。能定義那個介面的記憶體廠，賣進去的是更大的位置。請完全比照本頁對 HBF 的處理：在「選擇權」欄位寫一行，在 2027E 模型裡寫零。一個月內出現兩個選擇權項目不是趨勢，那是研發部門在做它的工作 ◆。

Moat護城河

Moat 1 — HBM stacking know-how, locked in with NVIDIA護城河一 — HBM 堆疊 know-how，靠 NVIDIA 共同開發鎖住

Mechanism: proprietary MR-MUF stacking yield and thermal performance, formalized through a multi-year NVIDIA(NVDA.O) co-development agreement (announced 2026-06-08, value undisclosed) — the pairing that underwrites 56.4% 1Q26 HBM share (IDC, via 424B4) ✓.

機制：MR-MUF 堆疊良率與散熱表現的專有技術，透過與 NVIDIA(NVDA.O) 多年期共同開發協議（2026-06-08 宣布，金額未揭露）正式綁定 — 這正是 1Q26 HBM 份額 56.4%（IDC，引自 424B4）✓ 的支撐。

What it takes to break it: a rival must match the thermal/yield performance at production scale and win an equivalent co-design lock-in with NVIDIA — shipping one qualifying part isn't enough.

要打破它需要什麼：對手得同時做到量產級的散熱／良率表現，而且要贏得跟 NVIDIA 同等級的共同設計綁定 — 光是出貨一顆合格產品還不夠。

Erosion tripwire: Samsung's HBM4 is seeking price parity with SK (vs a ~30% discount in the HBM3E era) ✓ TrendForce — parity would mean the premium this moat commands is gone even while share holds.

侵蝕觸發點：三星 HBM4 正尋求與 SK 同價（HBM3E 時代低約 30%）✓ TrendForce — 一旦同價，即便份額未失，這道護城河能收的溢價也已經不在了。

Moat 2 — the oligopoly protects the industry, not the ticker護城河二 — 寡占結構保護的是產業，不是這檔股票

Mechanism: a three-player DRAM/HBM oligopoly with capex-scale barriers to entry; management told the 1Q26 call "customer demand already exceeds planned capacity for the next three years." That protects industry-wide pricing power, but not necessarily SK's own share — Samsung's HBM4 reached mass production in 2026-02 (>$1B in 4 months) and Micron's HBM3E+HBM4 is booked through 2027 ✓ TrendForce, both inside the same three-player structure.

機制：三寡占的 DRAM/HBM 產業結構，進入門檻是資本規模；管理層在 1Q26 法說會表示「客戶需求已超出未來三年的規劃產能」。這保護的是整個產業的定價權，不必然是 SK 自己的份額 — 三星 HBM4 已在 2026-02 量產（4 個月營收破 $1B）、美光 HBM3E+HBM4 訂到 2027 ✓ TrendForce，兩者都站在同一個三寡占結構裡。

What it takes to break it: a fourth player reaching qualified, at-scale volume — not just capital, but years of yield-learning inside a structure the incumbents already occupy.

要打破它需要什麼：要有第四家玩家做到合格且量產的規模 — 不只是資本，還得在既有三家已站穩的結構裡，走完多年的良率學習曲線。

Erosion tripwire: CXMT (長鑫存儲, newly listed in 2026) raised ₩14tn for next-gen DRAM/HBM ✓ ebn — this moat only erodes once CXMT actually crosses from commodity DRAM into qualified HBM volume.

侵蝕觸發點：CXMT（長鑫存儲，2026 甫上市）IPO 募 ₩14兆投入次世代 DRAM/HBM ✓ ebn — 這道護城河，只有在 CXMT 真正從商品 DRAM 跨進合格 HBM 量產時，才會開始被侵蝕。

Bull case多方論點

The shortage is intact — just no longer accelerating短缺仍在，只是不再加速

✗→ 07-29 2Q26 actual: OP ₩60.54tn (+61% QoQ), OPM 76.3% — a record, but below the ₩64.16tn consensus ✓ official官方. Prices are still rising (company blended DRAM ASP +~30% QoQ, NAND +mid-50% ✓ call, reported法說轉述), NAND still has "almost zero new capacity in 2026" ✓ TrendForce 05-25, and management guides 2026 industry DRAM demand +mid-20% YoY.

✗→ 07-29 2Q26 實績：OP ₩60.54兆（+61% QoQ）、OPM 76.3% — 仍創新高，但低於共識 ₩64.16兆 ✓ 官方。價格還在漲（官方 DRAM 混合 ASP +約30% QoQ、NAND +mid-50% ✓ 法說轉述），NAND「2026 幾乎零新增產能」✓ TrendForce 05-25，管理層指引 2026 產業 DRAM 需求 +mid-20% YoY。

Three-year visibility三年能見度

2026 DRAM+NAND fully sold out (official 3Q25 language) + customer demand already exceeds planned capacity for the next three years (1Q26 call) + a reported 3-year Microsoft DDR5 LTA ✓⚠. Partially in the price.

2026 DRAM+NAND 全量售罄（官方 3Q25 語言）+ 客戶需求超出未來三年規劃產能（1Q26 法說）+ 微軟三年 DDR5 長約（轉述）✓⚠。部分在價內。

HBM4 dominanceHBM4 主導

NVIDIA(NVDA.O)'s Rubin generation share reportedly ~2/3 for SK, above the originally expected 50%+ ✓ TrendForce 01-28, reported轉述; HBM4E samples already shipped. In the price.

NVIDIA(NVDA.O) Rubin 世代 SK 份額 ~2/3（報導估計，高於原預期 50%+）✓ TrendForce 01-28,轉述；HBM4E 已出樣。在價內。

What's genuinely underappreciated真正被低估的

1. Book-value compounding — and it's no longer hypothetical: the Kioxia gains booked in 2Q26 lifted BVPS to ≈₩363,000 (computed ⚠), so PBR came down from the 8× range to ≈4.3× (≈3.9× at the 07-29 low) ✗→ 08-03; on this analysis's path, FY27-end BVPS reaches ~₩860,000 ◆ — today's price is ~1.8× two-years-out book. 2. The net-cash fortress is now an official number, not a pro-forma: ₩69.4tn (≈$49.7B) net cash (₩88tn (≈$63.1B) cash, ₩18.6tn (≈$13.3B) debt) at end-2Q26 ✓ official官方 07-29 — ammunition for the next cycle trough. 3. If agentic-AI inference token economics still have runway, as a Bank of Korea report suggests ✓ daum/BOK, reported轉述 07-19, the 2028-plateau assumption is actually too conservative.

1. 帳面價值複利 — 而且已不是假設：2Q26 Kioxia 利益入帳後 BVPS 升至約 ₩363,000（計算 ⚠），PBR 從 8× 檔降到約 4.3×（07-29 低點時約 3.9×）✗→ 08-03；沿本研究路徑，FY27 末 BVPS 約 ₩860,000 ◆ — 現價約為兩年後帳面的 1.8 倍。2. 淨現金堡壘已是官方實報數字、不再是 pro-forma：2Q26 末淨現金 ₩69.4兆（≈$49.7B）（現金 ₩88兆（≈$63.1B）、負債 ₩18.6兆（≈$13.3B））✓ 官方 07-29 — 下輪景氣谷底的併購/加碼火藥。3. 代理式 AI 推論的 token 經濟若如韓國央行(BOK)報告所言仍有續航 ✓ daum/BOK轉述 07-19，2028 平原假設反而太保守。

Bear case空方論點

Structured as fear → company claim → independent evidence → this analysis's judgment, held with equal weight against the bull case above.

結構：恐懼 → 公司說法 → 獨立證據 → 本研究判斷，與上方多方論點同等權重。

Fear: cycle-top恐懼：高點論

고점론 (cycle-top). From the 2026-06-26 intraday peak of ₩2,987,000 the stock is -43.4% ✗→ 08-21; -8.4% on 7/2, another leg on 7/16 (CXMT + AI-capex worries) ✓ sedaily/ebn, then -20.4% across 07-28/29 around the 2Q26 print ✓ computed計算.

고점론（cycle-top）。自 2026-06-26 盤中高點 ₩2,987,000 至今 -43.4% ✗→ 08-21；7/2 單日 -8.4%、7/16 CXMT+AI capex 疑慮再殺一波 ✓ sedaily/ebn，07-28/29 財報前後兩日再 -20.4% ✓ 計算。

Then the tape turned on it. 07-31 closed limit-up +29.95% at ₩1,718,000 — SK hynix's first daily limit since 2009, and its first ever under the 30% band introduced in 2015 — after Microsoft and Amazon results on 07-30 removed the AI-capex fear and the Philadelphia semiconductor index rose 8.19%; foreign investors bought ₩5tn of Korean equities that session ✓ Korean press ×3韓媒×3. 08-03 then gave back -8.80% to ₩1,567,000 on profit-taking, and on 08-20 the ₩40tn buyback added another +12.73% to ₩1,691,000 ✗→ 08-21. A ±30% round trip in three sessions with no new fundamental data — followed by a 12.7% capital-allocation day — is a statement about positioning, not about the cycle ◆.

然後盤面反過來打它。07-31 以 +29.95% 上限漲停收在 ₩1,718,000 — SK 海力士 2009 年以來首次漲停，也是 2015 年放寬到 30% 之後的第一次 — 起因是 07-30 微軟與亞馬遜財報消除了 AI capex 疑慮、費城半導體指數大漲 8.19%，當日外資買超韓股約 ₩5兆 ✓ 韓媒×3。08-03 再因獲利了結吐回 -8.80%，收 ₩1,567,000；08-20 再因 ₩40兆庫藏股大漲 +12.73%，收 ₩1,691,000 ✗→ 08-21。三個交易日、沒有任何新的基本面數據，卻走完 ±30% 的來回 — 這說的是籌碼，不是循環 ◆。

Company's claim公司說法

Three-year visibility, sold out, "the favorable pricing environment continues" ✓ 1Q26 release.

三年能見度、售罄、「有利價格環境延續」✓ 1Q26 release。

Independent evidence獨立證據

HBM contract price already fell in 1Q26, hitting SK's ASP ✓ TrendForce
The company itself slowed its HBM4 conversion line — full-year shipment target cut from 4.5 to 4B Gb — to chase commodity margin instead: management voting with its feet on where the excess profit really is ✓ TrendForce 06-23, reported轉述
The supply wave is already funded: M15X tape-out 1Q26, Yongin fab-1 cleanroom 1Q27 (₩9.4tn), P&T7 late-2027 (₩19tn), Indiana 2H28 (₩5.9tn), a Cheongju NAND plant 2029 ⚠ single-source單源; unrecognized capital commitments jumped from ₩6.7tn to ₩30.9tn quarter-over-quarter ✓ 424B4; plus Samsung expansion, Micron running full, and CXMT's ₩14tn raise
Issuing $26.5B of new shares at a 9.4× peak PBR is a textbook cycle-top signal — management selling stock when its own shares are most expensive
Foreign ownership has been a full-year net seller (53.8%→49.9%) ✓ ajunews

HBM 合約價 1Q26 已下跌，衝擊 SK ASP ✓ TrendForce
公司自己放慢 HBM4 轉線 — 全年出貨目標自 4.5 下修至 4B Gb — 轉去追商品毛利：管理層用腳投票告訴你超額利潤在哪裡 ✓ TrendForce 06-23,轉述
供給波已被融資完成：M15X 1Q26 投片、Yongin 一廠 1Q27 無塵室（₩9.4兆）、P&T7 末-2027（₩19兆）、Indiana 2H28（₩5.9兆）、清州 NAND 廠 2029 ⚠ 單源；未認列資本承諾單季自 ₩6.7兆跳升至 ₩30.9兆 ✓ 424B4；加上三星擴產、美光滿載、CXMT ₩14兆增資
在 9.4× 峰值 PBR 上增發 $26.5B 新股 — 教科書級的循環頂部訊號（管理層在自己股票最貴時賣股票）
外資全年淨減持（53.8%→49.9%）✓ ajunews

This analysis's judgment & tail risks本研究判斷與尾部風險

2027 "plateau" is more probable than "cliff" — and the LTA cushion is no longer an assumption: the company finalized long-term agreements with ~10 customers in 2Q26 ✓ official官方 07-29. But pre-print consensus 2027E OP of ₩386tn (≈$277B) runs far too high — this analysis's 2027E OP is ₩275tn (≈$197B; -28.8% vs that consensus, revised 07-29). The cliff risk has largely migrated to 2028.

2027「高原」比「懸崖」機率高 — 而且長約緩衝已不是假設：公司 2Q26 與約 10 家客戶簽訂長約 ✓ 官方 07-29。但財報前共識 2027E OP ₩386兆（≈$277B）仍明顯過高 — 本研究 2027E OP ₩275兆（≈$197B，較該共識 -28.8%，07-29 下修）。懸崖風險主要遷移到 2028。

Tail risks: China's annual VEU licensing regime (Wuxi DRAM/Dalian NAND equipment licenses renewing yearly starting 2026) ✓ CNBC 2025-12-30; the scope of Section 232 semiconductor tariffs remains unsettled ⚠ conflicting sources來源衝突; won/KOSPI concentration (SK + Samsung = 54.8% of KOSPI market cap) ✓ g-enews 06-25.

尾部風險：中國 VEU 年審制（無錫 DRAM/大連 NAND 設備許可 2026 起逐年續）✓ CNBC 2025-12-30；Section 232 半導體關稅範圍未定 ⚠ 來源衝突；韓元、KOSPI 集中度（SK+三星＝54.8% KOSPI 市值）✓ g-enews 06-25。

The 2026-29 supply wave2026-29 供給波

The 2027-29 capacity additions land squarely inside the window consensus has drawn as a profit plateau — a shaded coincidence this analysis reads as evidence the wave is already priced to matter.2027-29 新產能落地時點，恰好落在共識自己畫出的獲利高原窗口內 — 這個重疊，本研究視為供給波確實有分量的佐證。

Variant perception變異認知

Consensus reads this stock as an HBM/AI structural growth name where "the PBR era is over," to be re-rated on P/E or SOTP instead — sell-side has broadly dropped PBR as a framework ✓ MTN 02-03, and iM Securities uses a 7.5× PBR to derive a ₩4.2M target for Mirae Asset ✓ newspim 06. What the market overlooks:

共識把本股讀成 HBM/AI 結構性成長股，「PBR 時代結束」，改用 P/E 或 SOTP 重估 — 券商紛紛棄用 PBR 框架 ✓ MTN 02-03，iM 證券用 7.5× PBR 推算未來資產目標價 ₩4.2M ✓ newspim 06月。市場忽略的：

This is not a fight over 2026-27 earnings direction — it's about which slice of the P&L (commodity price vs HBM structure) the market is using to justify the multiple, and whether the fully-funded 2027-29 supply wave is priced at all.這不是在爭 2026-27 的獲利方向 — 而是市場用哪一塊損益表（商品價格 vs HBM 結構）替倍數找理由，以及已全額付款的 2027-29 供給波，是否根本沒被定價。

The structure of the argument is unchanged: both tails remain fatter than one-line framings admit — bull 20% at ₩3,460,000, bear 35% at ₩1,020,000 ◆. What keeps changing is where the price sits between them. On 07-29 it fell through the pre-committed entry zone and the view went to Buy; by the 08-03 close of ₩1,567,000 it had climbed back above the Buy line and the view went back to range-bound; by the 08-20 close of ₩1,691,000 it sits 5.1% above a Buy line that has itself moved up to ₩1,608,700 ✗→ 08-21. Across the first two steps nothing in the estimate moved at all; across the third only the share count did.

That is worth naming plainly, because it is the practical shape of this whole thesis. If the disagreement with consensus is about the multiple rather than the direction — and the sell-side's own post-print range of ₩1.48M to ₩4.70M says it is — then the stock will keep crossing its own decision lines on flows, a chairman's share purchase and hyperscaler earnings, while the 2027 earnings question sits untouched until a contract-price print or a quarterly report actually answers it. The label follows the price. The position should follow the pre-committed levels — and the levels only move when the estimate does ◆.

論證的骨架沒有變：兩條尾巴仍比一句話式的框架更肥 — 多方 20% 權重 ₩3,460,000、空方 35% 權重 ₩1,020,000 ◆。一直在變的，是價格站在兩者之間的哪個位置。07-29 它跌穿預先承諾的進場區，觀點轉為「買進」；到 08-03 收盤 ₩1,567,000，它又爬回買進線之上，觀點退回「區間」；到 08-20 收盤 ₩1,691,000，它高於一條自己也往上移到 ₩1,608,700 的買進線 5.1% ✗→ 08-21。前兩步之間估計值一動也沒動；第三步只動了股數。

這件事值得直說，因為它就是整個論點在現實裡的形狀。如果與共識的分歧在倍數、不在方向 — 而賣方自己財報後 ₩1.48M 到 ₩4.70M 的區間正是這麼說的 — 那麼這檔股票就會不斷因為資金流、會長買股、超大規模業者的財報而來回穿越它自己的決策線，而 2027 年的獲利問題會原封不動地擱著，直到某一份合約價數據或某一季財報真的回答它。標籤跟著價格走；部位應該跟著預先承諾的價位走 — 而價位只有在估計值動的時候才動 ◆。

Numbers, valuation & scenarios數字、估值與情境

₩tn except per-share (₩)₩兆，每股例外(₩)	FY25A	FY26E ◆	FY27E ◆
Revenue營業收入	97.15 ≈$70B	323 ≈$232B	360 ≈$258B
Operating profit營業利益	47.21 ≈$34B	240 ≈$172B	275 ≈$197B
Operating margin營益率	48.6%	74.3%	76.4%
EPS (₩)	62,044	343,000*	309,000 post-cancellation註銷後
This analysis vs consensus (OP)本研究 vs 共識（營業利益）	—	-9.5%	-28.8%

FY26E = 1H26 actuals + a trimmed 2H26 path (3Q OP ₩70tn / 4Q ₩72tn ◆); FY27E = this analysis's own assumption. *FY26E EPS is reported-basis and includes one-offs (1Q's ₩17.1tn financial gain, 2Q's ₩63.3tn pretax Kioxia-related investment gains); ex-one-off core is ≈₩272,000 — the TP uses the clean 2027E path. Consensus in the comparison row is PRE-print: FnGuide 07-10 (2026E OP ₩265.2tn / 2027E OP ₩386.1tn) ✓; it will reset downward after the 2Q miss. All ≈US$ figures on this page use ₩1,395/$ (08-20), rounded — re-rendered from ₩1,430/$ as the won firmed a further ~2.5%.FY26E＝上半年實績＋下修後的 2H26 路徑（3Q OP ₩70兆／4Q ₩72兆 ◆）；FY27E＝本研究自行假設。*FY26E EPS 為財報口徑、含一次性項目（1Q 金融收益 ₩17.1兆、2Q Kioxia 相關投資利益稅前 ₩63.3兆）；剔除一次性後核心約 ₩272,000 — 目標價使用的是乾淨的 2027E 路徑。對照列共識為財報前數字：FnGuide 07-10（2026E OP ₩265.2兆／2027E OP ₩386.1兆）✓，2Q miss 後將下修重置。本頁所有 ≈US$ 皆以 ₩1,395/$（08-20）換算、四捨五入 — 韓元再升值約 2.5%，已自 ₩1,430/$ 全面重算。

2Q26 scorecard — marking our own homework (T4)2Q26 對帳單 — 自己改自己的考卷（T4）

2Q26	Actual實際數 ✓ official官方 07-29	This analysis (prior)本研究前估	Diff差異	Consensus市場共識	Diff差異
Revenue營收	₩79.32兆 ≈$56.9B	₩83.9兆 ≈$60.2B	-5.5%	₩83.93兆 ≈$60.2B	-5.5%
Operating profit營業利益	₩60.54兆 ≈$43.4B	₩64.2兆 ≈$46.0B	-5.7%	₩64.16兆 ≈$46.0B	-5.6%
OPM	76.3%	76.5%	-0.2pp	76.5%	-0.2pp

This analysis anchored its 2Q on the consensus number and missed by the same ~5.6% consensus did: LTAs covering roughly half of sales capped how much of the spot-price surge actually reached the P&L — a mechanism neither the street nor this model priced. Net profit ₩93.92tn (≈$67.3B) (118% net margin) is not comparable to any estimate: it includes ₩63.27tn (≈$45.4B) of one-off investment gains (Kioxia stake disposal + SPC revaluation) ✓ sedaily 07-29; core ex-one-off net is ≈₩45.5tn (≈$32.6B) (computed ⚠).本研究 2Q 錨定共識，因此和共識一起錯了約 -5.6%：約佔營收一半的長約（LTA）封住了現貨飆價流進損益表的比例 — 這個機制，街上和本模型當時都沒有定價。淨利 ₩93.92兆（≈$67.3B）（淨利率 118%）與任何估計都不可比：內含 ₩63.27兆（≈$45.4B）一次性投資利益（Kioxia 持股處分＋SPC 評價）✓ sedaily 07-29；剔除後核心淨利約 ₩45.5兆（≈$32.6B）（計算 ⚠）。

Estimate revisions (T3) — what changed and why預估修正對照（T3）— 改了什麼、為什麼

	Revised調整後 ◆ 08-21	Prior調整前 ◆ 07-29	Diff差異	Consensus (pre-print)共識(財報前)	Diff差異
2026E rev	₩323兆	₩323兆	0.0%	₩358.3兆	-9.9%
2026E OP	₩240兆	₩240兆	0.0%	₩265.2兆	-9.5%
2027E rev	₩360兆	₩360兆	0.0%	—	—
2027E OP	₩275兆	₩275兆	0.0%	₩386.1兆	-28.8%
2027E share count2027E 股數	705.2M	729.0M	-3.3%	—	—
2027E EPS	₩309,000	₩300,000	+3.0%	—	—
TP	₩1,850,000	₩1,800,000	+2.8%	₩3,408,502*	—

08-21 revision: denominator only. Every operating line above is carried forward from 07-29 unchanged — same revenue, same operating profit, same margin, same 6× multiple. What changed is the share count: the ₩40tn buyback retires ~23.7M shares (~3.3%), taking the 2027E divisor from 729.0M to 705.2M. Net of the ~₩0.95tn of after-tax interest income given up by spending the cash, 2027E EPS goes ₩300,000 → ₩309,000 (+3.0%) and the TP ₩1,800,000 → ₩1,850,000 (+2.8%). The revision is contingent on execution: the buyback runs to 11-19 and cancellation follows within one to two weeks, so the accretion is a board resolution, not yet a fact. *Sell-side average TP is the pre-print figure (37 analysts, 07-22); the post-print roster of 18 houses averages ≈₩3,296,000 with a ₩1.48M–4.70M range (07-30 ⚠). The realized 3Q26 contract print — the number that would actually move the operating lines — is still not settled as of 08-21.08-21 修正：只動分母。上表每一條營運線都從 07-29 原封延用 — 同樣的營收、同樣的營業利益、同樣的營益率、同樣的 6 倍 PER。變的是股數：₩40兆庫藏股註銷約 23.7M 股（約 3.3%），2027E 的除數從 729.0M 降到 705.2M。扣掉花掉這筆現金所放棄的約 ₩0.95兆稅後利息收入後，2027E EPS 自 ₩300,000 升至 ₩309,000（+3.0%），目標價自 ₩1,800,000 升至 ₩1,850,000（+2.8%）。這項修正取決於執行：買回期到 11-19，註銷在其後一到兩週內完成，因此增益目前是一紙董事會決議，還不是事實。*賣方平均目標價為財報前數字（37 位，07-22）；財報後 18 家名單平均約 ₩3,296,000，區間 ₩1.48M–4.70M（07-30 ⚠）。真正會動到營運線的 3Q26 實際合約價，截至 08-21 仍未定案。

TP derivation目標價推導 ◆

This analysis models 2026/2027 EPS at ₩343,000 (incl. one-offs) / ₩309,000; the 6-month target price is ₩1,850,000 (prior: ₩1,800,000), based on 2027E EPS ₩309,000 × 6× PER (= ₩1,854,000, rounded). The +2.8% increase is entirely share-count-driven: the ₩40tn cancellation takes the 2027E divisor from 729.0M shares to 705.2M, net of ~₩0.95tn of forgone after-tax interest income. The 6× multiple is unchanged, and no operating estimate moved. A buyback cannot raise earnings power — it can only divide the same earnings among fewer holders, and that is exactly the size of this upgrade.

本研究預估 2026、2027 年 EPS 為 ₩343,000（含一次性）、₩309,000，目標價 ₩1,850,000（前次 ₩1,800,000），基於 2027E EPS ₩309,000 × 6 倍 PER（＝₩1,854,000，取整）。+2.8% 的上調全部來自股數：₩40兆註銷把 2027E 除數自 729.0M 股降到 705.2M 股，並已扣除約 ₩0.95兆放棄的稅後利息收入。6 倍 PER 不變，營運估計一項未動。庫藏股沒有辦法提高獲利能力 — 它只能把同樣的獲利分給更少的持有人，而這次上調的幅度，剛好就是那個大小。

◆ Cross-check: TP ₩1,850,000 ≈ 2.1× FY27E BVPS (≈₩880,000 post-cancellation — the ₩40tn leaves equity, but so do the shares it retires, so book per share is roughly flat); bear-case TP ₩1,020,000 ≈ 1.2× FY27E BVPS, close to the historical trough. The current price is ≈1.9× FY27E book, up from 1.8× at the 08-04 anchor.

◆ 交叉檢驗：TP ₩1,850,000 ≈ 2.1× FY27E BVPS（註銷後約 ₩880,000 — ₩40兆離開權益，但它買走的股數也一起消失，每股帳面因此大致持平）；bear TP ₩1,020,000 ≈ 1.2× FY27E BVPS，貼歷史谷底。現價約為 FY27E 帳面的 1.9 倍，較 08-04 錨點的 1.8 倍再墊高。

◆ Anchor reconciliation: the probability-weighted scenario value is ₩1,881,500 — +11.3% vs the ₩1,691,000 close, and +1.7% above the ₩1,850,000 target. The canonical TP stays the base-case formula; the weighted figure running slightly higher reflects the residual 20% bull tail, not a model inconsistency.

◆ 錨點調和：機率加權情境值為 ₩1,881,500 — 較收盤 ₩1,691,000 高 +11.3%，較目標價 ₩1,850,000 高 +1.7%。正式 TP 仍以基準情境公式為準；加權值略高，係因殘餘的 20% 多方尾部所致，非模型不一致。

◆ The band boundary, stated explicitly: this page's view bands are defined on 6-month absolute return, so the Buy threshold of +15% corresponds to a price of ₩1,850,000 ÷ 1.15 = ₩1,608,696. Above it the framework says range-bound; below it, Buy. The 08-20 close of ₩1,691,000 sits 5.1% above it. Note what the buyback did to this line: by raising the target it also raised the entry price, from ₩1,565,217 to ₩1,608,696 — but it raised the market price by more than twice as much, so the gap to Buy widened rather than closed. Last update this was a 0.11% boundary call; this one is not close.

◆ 把分界線講清楚：本頁的觀點區間定義在 6 個月絕對報酬上，因此 +15% 的買進門檻對應到的價格是 ₩1,850,000 ÷ 1.15 = ₩1,608,696。高於它，框架說「區間」；低於它，說「買進」。08-20 收盤 ₩1,691,000 高出它 5.1%。注意庫藏股對這條線做了什麼：它把目標價抬高，因此也把進場價抬高，自 ₩1,565,217 升到 ₩1,608,696 — 但它把市價抬高了兩倍不止，所以距離「買進」的差距是變寬、不是變窄。上次更新這是一個 0.11% 的邊界判斷；這一次一點都不邊界。

Top: the three scenarios by probability weight (base = the TP price; weights last reweighted 07-29 and unchanged since — the buyback changes the share count, not the memory cycle, so no probability moved). Bottom: the same three scenario prices plus the probability-weighted value (~₩1.88M, +11.3% vs the close), the 08-20 close (₩1,691,000) and the ₩1,608,700 line where the Buy band begins, on a shared linear 0.8M–3.6M value axis. Both the close and the Buy line moved right since the last update — but the close moved further, so the gap between them opened instead of closing.上：三個情境按機率權重呈現（基準情境＝目標價；權重自 07-29 重估後未再變動 — 庫藏股改變的是股數，不是記憶體循環，因此沒有一個機率被動過）。下：同樣三個情境價位，加上機率加權值（約₩1.88M，較收盤 +11.3%）、08-20 收盤（₩1,691,000），以及買進區起點 ₩1,608,700，畫在同一條 0.8M–3.6M 線性數值軸上。收盤與買進線自上次更新後都往右移 — 但收盤移得更多，所以兩者之間的距離是拉開、不是收攏。

Scenario情境	Prob.機率 ◆	Key assumptions關鍵假設	2027E path路徑	Scenario TP情境 TP
Bear cliffBear 懸崖	35% (30%→)	Contract prices turn down starting 4Q26, -30% through 2027; AI-capex digestion year; CXMT/Samsung supply shock. Weight raised 07-29: deceleration confirmed4Q26 起合約價轉跌、2027 全年 -30%；AI capex 消化年；CXMT/三星供給衝擊。07-29 上調權重：減速獲確認	rev ~₩300tn, OP ~₩180tn (≈$129B), EPS ~₩204,000 post-cancellationrev~₩300兆,OP~₩180兆（≈$129B）,註銷後EPS~₩204,000	204,000×5×=₩1,020,000 (-39.7%; ≈1.2× FY27E BVPS, near historical trough ~1×貼歷史谷底~1×)
Base plateauBase 高原	45%	LTAs cushion 1H27 — now contractual (11 customers signed as of 2Q26) — with gradual easing 2H27; 2027 OP 29% below pre-print consensus1H27 長約撐價 — 已成合約事實（2Q26 止簽約 11 家）— 2H27 緩跌；2027 OP 較財報前共識低 29%	rev ₩360tn, OP ₩275tn (≈$197B), EPS ₩309,000 post-cancellationrev₩360兆,OP₩275兆（≈$197B）,註銷後EPS₩309,000	309,000×6×=₩1,850,000 (+9.4%; =TP)
Bull continuationBull 續航	20% (25%→)	Token demand absorbs all new capacity; pre-print 2027 consensus lands as-is. Weight cut 07-29: LTAs covering ~half of sales cap the spot-price capture the bull case needstoken 需求消化全部新產能；財報前 2027 共識成真。07-29 下調權重：約占營收一半的長約，封住了 bull 情境需要的現貨價上檔	OP ₩386tn (≈$277B), EPS ~₩433,000 post-cancellationOP₩386兆（≈$277B）,註銷後EPS~₩433,000	433,000×8×=₩3,460,000 (+104.6%; ≈back above the June peak重回6月高點上方)

Probability-weighted value = 0.35×1,020,000 + 0.45×1,850,000 + 0.20×3,460,000 = ₩1,881,500 (+11.3% vs the 08-20 close). Close ₩1,691,000 ≈ base EPS × 5.5×. All three scenario EPS figures are restated onto the post-cancellation 705.2M share count; the probability weights themselves are unchanged.

機率加權值 = 0.35×1,020,000 + 0.45×1,850,000 + 0.20×3,460,000 = ₩1,881,500（較 08-20 收盤 +11.3%）。收盤 ₩1,691,000 ≈ base EPS × 5.5×。三個情境的 EPS 都已改用註銷後的 705.2M 股數重述；機率權重本身未動。

Asymmetry check不對稱性檢驗 ◆

✗→ 08-21 Re-run at ₩1,691,000 against the post-buyback scenario set, the asymmetry is the worst this page has recorded. To double (→₩3.38M) now requires 98% of the entire bull scenario (₩3.46M) — it needed 93% on 08-03 and 82% at the 07-29 low. To fall by half (→₩845,500) corresponds to 2027 OP of roughly ₩151tn (≈$108B) at 5× — about 16% below the bear scenario's ₩180tn, versus 21% below on 08-03 and 30% below at the 07-29 low. Both edges have now moved against the buyer three updates running. Note that the buyback nudged the bull ceiling up (₩3.36M→₩3.46M) and the bear floor up (₩1.00M→₩1.02M) — the cancellation lifts every scenario by the same ~3% — but the price rose 7.9% in the same step, so the compression is entirely a price effect. The stock remains an option on the 2027 memory contract price; the buyback made it a slightly larger option and a materially more expensive one.

✗→ 08-21 以 ₩1,691,000 對照庫藏股後的情境組重跑，不對稱性是本頁記錄以來最差的一次。翻倍（→₩3.38M）現在需要吃掉整個 bull 情境（₩3.46M）的 98% — 08-03 時要 93%，07-29 低點時只要 82%。跌一半（→₩845,500）對應 2027 OP 約 ₩151兆（≈$108B）× 5 倍 — 比 bear 情境的 ₩180兆低約 16%，08-03 時是低 21%、07-29 低點時是低 30%。兩邊的邊界已經連續三次更新都往買方不利的方向移動。要注意的是，庫藏股把多方天花板往上推了一點（₩3.36M→₩3.46M），空方地板也往上推了一點（₩1.00M→₩1.02M）— 註銷讓每個情境都同步抬高約 3% — 但同一步裡股價漲了 7.9%，所以壓縮完全是價格造成的。本股仍是一張關於 2027 記憶體合約價的期權；庫藏股把這張期權變大了一點，卻把它變貴了很多。

Catalysts (12mo, T8)催化劑（12個月，T8）

Timing時間	Event事件	Threshold → reaction門檻 → 反應 ◆	Status狀態
2026-08-20 → 11-19 (+1–2wk to cancel)(註銷再+1–2週)	₩40tn buyback + cancellation₩40兆庫藏股取得並註銷	Pre-committed bar: full ₩40tn executed and cancelled by ~December → the +3.0% EPS accretion and the ₩1,850,000 TP are confirmed as fact. Executed short of ₩40tn, or cancellation deferred → reverse the 08-21 TP revision back toward ₩1,800,000, since the entire upgrade is the share count. Track the monthly treasury-share disclosures rather than waiting for 11-19. Watch also whether the average execution price runs above the ₩1,662,000 reference — every won above it buys fewer shares and shrinks the accretion.預先承諾門檻：₩40兆全額執行並於約 12 月完成註銷 → +3.0% 的 EPS 增益與 ₩1,850,000 目標價確立為事實。若執行未達 ₩40兆、或註銷遞延 → 把 08-21 的目標價修正倒回 ₩1,800,000 方向，因為這次上調的全部就是股數。要追的是每月庫藏股取得公告，不是等 11-19。同時盯平均執行價是否高於 ₩1,662,000 參考價 — 每高一元就買到更少股數，增益就更小。	executing — 08-21 day 2執行中 — 08-21 第2天
2026-07-29 09:00 KST	2Q26 earnings + call2Q26 財報+法說	Pre-committed bar: OP ≥₩64tn = confirm / <₩60tn = bear migration. Actual: OP ₩60.54tn — in the gray zone between the two lines; 2H language did not weaken (capex raised to high-₩40tn, HBM4 ramping, demand-momentum intact). Reaction executed: 2026E/27E cut ~-7~8%, bear 30→35%, bull 25→20%, TP ₩1,960,000→₩1,800,000; view upgraded 區間→買進 on the price overshoot (see Update log).預先承諾門檻：OP ≥₩64兆＝確認／<₩60兆＝轉 bear。實際：OP ₩60.54兆 — 落在兩條線之間的灰色地帶；2H 語言未轉弱（capex 上修至 high-₩40兆、HBM4 放量、需求動能延續）。已執行反應：2026E/27E 下修約 -7~8%、bear 30→35%、bull 25→20%、TP ₩1,960,000→₩1,800,000；觀點因股價超跌自區間上調為買進（詳更新紀錄）。	resolved 07-29已結案 07-29
2026-07-03 (forecast)(預測) 2026-08 late (realized)下旬(實際)	TrendForce 3Q26 contract pricesTrendForce 3Q26 合約價	Pre-committed bar: conventional DRAM QoQ ≥+high-teens% → second entry tranche; flat/declining → stop adding, view downgrade review. The forecast leg is already on the record: +13–18% QoQ conventional DRAM, +10–15% NAND, published 07-03 — the top of the band only just touches the bar, so this is at-or-below it, not a confirmation. Reaction executed: second tranche not triggered, adding paused; 3Q26E ₩92tn/₩70tn left unrevised (already inside the implied range). The realized print still arbitrates — and as of 08-21 it has not settled: the latest available figures (TrendForce 07-09, restated 08-03) are still the same +13–18% forecast band, so nothing has arrived to move this row. The late-August window is now open and this stays the second thing to watch after the buyback execution.預先承諾門檻：conventional DRAM QoQ ≥+高十位數% → 進第二批；持平/轉跌 → 停止加碼、觀點下修檢討。預測腿已經在案：conventional DRAM +13–18% QoQ、NAND +10–15%，07-03 公布 — 區間上緣才勉強碰到門檻，因此屬於「打平或不足」，不算確認。已執行反應：第二批不觸發、加碼暫停；3Q26E ₩92兆/₩70兆維持不修（已落在隱含區間內）。實際合約價仍是最終仲裁者 — 而截至 08-21 它尚未定案：目前能取得的最新數字（TrendForce 07-09、08-03 重述）仍然是同一組 +13–18% 預測區間，沒有任何新資訊足以移動這一列。8 月下旬的視窗現已開啟，本項仍是庫藏股執行之後第二個要盯的。	forecast scored · realized still unsettled at 08-21, window now open預測已計分·實際08-21 仍未定案，視窗已開啟
~2026-10-28	3Q26 earnings — now the key data point3Q26 財報 — 下一個關鍵數據點	OP vs this analysis's revised path of ₩70tn: <₩65tn → trim review; ≥₩75tn (the old Kiwoom path) → model too conservative, TP upgrade review; OPM ≥75% sustainedOP 對照本研究新路徑 ₩70兆：<₩65兆 → 減碼檢討；≥₩75兆（原 Kiwoom 路徑）→ 模型過保守、TP 上修檢討；OPM ≥75% 續	scheduled已排定
~2026-10-07	ADS company lockup expiry (90 days)ADS 公司 lockup 90 天到期	informational only (company-side, no legacy shares)純資訊事件（公司端、無老股）	watch觀察
no date無日期	HBF commercialisation — a sampling date, a named customer, or a priceHBF 商品化 — 出樣日、具名客戶、或價格	The technical roadmap is public (Sandisk: first samples 2H26, AI-inference device samples early 2027) but the 08-03 spec carried no customer, no price and no volume commitment, and commercialisation is guided to ~2030. Watch for a named customer or a volume date — and equally for Samsung or Kioxia landing one first, which would settle whose format wins. Until one of those lands, this stays out of the model entirely.技術時程是公開的（Sandisk：2026 下半年首批樣品、2027 年初 AI 推論裝置樣品），但 08-03 的規格沒有客戶、沒有價格、沒有量的承諾，商業化指引在約 2030 年。要盯的是具名客戶或量產日期 — 同樣要盯三星或 Kioxia 是否先拿到，那會決定誰的格式勝出。在其中之一落地前，本項完全不進模型。	watch (no schedule)觀察(無時程)
2026-12	China VEU annual license renewal中國 VEU 年度換照	denial/tightening → tail risk realized, bear case under review拒發/加嚴 → 尾部風險實現、bear 檢討	speculative (annual regime)推測(年審制)
when disclosed公布時	TSMC CoWoS FY27 allocationTSMC CoWoS FY27 配分	the split becomes clear → upstream confirmation that ASIC volume is ramping分配明朗化 → ASIC 放量的上游確認	watch觀察
4Q26–1Q27	2027 LTA negotiations (HBM4/DDR5)2027 年度長約談判(HBM4/DDR5)	HBM4 NVIDIA share <50% or Samsung reaches price parity → moat-premium thesis dies, view under review; 2027 LTA coverage >80% at locked-in high prices → 2027E revised upHBM4 對 NVIDIA 份額 <50% 或三星達成同價 → 護城河溢價論死亡、觀點下修檢討；2027 LTA 覆蓋 >80% 且價格鎖高 → 2027E 上修	speculative (share data trackable quarterly)推測(季度份額數據可追蹤)
2027-01 late下旬	FY26 full-year + 4Q26 earnings, 2027 capex guidanceFY26 全年+4Q26 財報、2027 資本開支指引	a large capex upgrade → self-reinforcing confirmation of the supply wavecapex 大幅上修 → 供給波自我實現的再確認	scheduled已排定

Management & capital allocation經營層與資本配置

CEO Nohjung Kwak (곽노정) has led since 2022-03 (term through 2028-03) and also chairs SK Group's SUPEX semiconductor committee; the board chair is independent director Seung Beom Koh, since 2026-03 ✓ 424B4.

CEO Nohjung Kwak（곽노정）自 2022-03 上任（任期至 2028-03），並兼任 SK 集團 SUPEX 半導體委員會主席；董事會主席為獨立董事 Seung Beom Koh，自 2026-03 起 ✓ 424B4。

Capital allocation record資本配置紀錄

FY25 dividend ₩3,000/share (incl. ~₩1tn special dividend); 2026-02 retired 15.3M treasury shares (2.1% of shares); 2026-05 cleared a $1.7B exchangeable bond via a clean-up call; 2026-07 the $26.5B ADS raise, fully earmarked for capacity — ₩45.5tn Korea fab construction + ₩11.9tn EUV tools, delivery before 2027-12 ✓ 424B4. 2026-08-19: a ₩40tn (≈$28.7B) buyback with full cancellation, executing 08-20 to 11-19 — ~24.07M shares at the ₩1,662,000 reference price, ~3.3% of shares outstanding, and the largest share retirement ever executed by a Korean listed company. The board simultaneously lifted the 2025-27 shareholder-return floor from "within 50%" of cumulative free cash flow to "50% or more", and stated that intrinsic value was "not sufficiently reflected in the current share price" ✓ board disclosure 08-19 + Korean press ×508-19 董事會公告＋韓媒×5.

FY25 股息 ₩3,000/股（含特別股息 ~₩1兆）；2026-02 註銷 15.3M 庫藏股（占股本 2.1%）；2026-05 以 clean-up call 清償 $1.7B 交換債；2026-07 $26.5B ADS 增資，全數撥入產能 — ₩45.5兆韓國廠房興建 + ₩11.9兆 EUV 設備，2027-12 前交機 ✓ 424B4。2026-08-19：₩40兆（≈$28.7B）庫藏股取得並全數註銷，執行期 08-20 至 11-19 — 以參考價 ₩1,662,000 計約 24.07M 股，占流通股數約 3.3%，是韓國上市公司史上最大規模的股份註銷。董事會同時把 2025-27 年股東回饋下限，自累計自由現金流的「50% 以內」上調為「50% 以上」，並表示公司內在價值「未充分反映在目前股價上」✓ 08-19 董事會公告＋韓媒×5。

Reading the signal解讀

Shareholder returns (retirement + special dividend) and a top-of-cycle capital raise coexist — the raise's use of proceeds is honest (100% capex), but its timing is itself management's judgment on share price vs future capital needs. SK Square must maintain ≥20% ownership under Korea's Fair Trade Act holding-company rules, capping dilution from the raise at ~2.5% — limited economic dilution, but the signal still matters.

✗→ 08-21 The ₩40tn buyback, weighed rather than cheered. Set the two capital events side by side. In July the company raised ≈₩37tn of equity ($26.5B) to fund fabs and EUV tools. Five weeks later it committed ₩40tn to retiring equity. Against ~2.5% of dilution from the raise and ~3.3% of cancellation, the share count over Jul–Dec is close to a round trip. The two pools are not the same money — the ADS proceeds are contractually earmarked for capacity and the buyback comes from operating cash — but a company that had ₩40tn available to retire stock, and raised equity anyway six weeks earlier, is making a statement about sequencing that deserves to be read rather than skipped ◆.

Three things are true at once, and only the first is unambiguously good. (1) Mechanically it is accretive: ~3.3% fewer shares against unchanged earnings, worth ~+3.0% to 2027E EPS, and this page has taken that into the target. (2) The announcement raised the company's own cost of execution — at ₩1,691,000 instead of the ₩1,662,000 reference, ₩40tn buys ~23.7M shares instead of ~24.07M, so roughly a tenth of the intended retirement was consumed by the reaction to announcing it. (3) It retires stock at ≈4.7× book with peak-cycle cash, into the 2027-29 supply wave this page's own bear case describes and the company's own capex schedule funds. Memory balance sheets have been damaged by exactly this before. The July raise says management wanted capital for the downcycle; the August buyback says the stock is cheap today. Both cannot be the higher-return use of the same won ◆.

✗→ 08-04 The "no insider trades found" line is retired. On 07-30, SK Group chairman Chey Tae-won (최태원) disclosed an open-market purchase of 3,620 SK hynix common shares at ₩1,353,677, ₩4.90bn (≈$3.5M) in total ✓ filing via Korean press ×3公告，韓媒×3. Held against the bear case's cycle-top read of the ADS raise, this is the mirror image: the company sold stock to the market at the top, and its chairman bought stock from the market after a halving. Weigh it honestly — ₩4.9bn is a rounding error beside a ₩1,232tn market cap and beside his own holdings; it is a signalling instrument, not an economic one, and it landed the session before a limit-up ◆.

股東回饋（註銷庫藏股+特別股息）與頂部增資並存 — 增資用途誠實（全數為 capex），但時點本身就是管理層對「現在股價 vs 未來資本需求」的判斷。SK Square 須依韓國公平交易法控股規範維持 ≥20% 持股，增資稀釋上限被鎖在 ~2.5% — 經濟稀釋有限，但訊號意義仍在。

✗→ 08-21 ₩40兆庫藏股，用秤的、不用喊的。把兩件資本事件並排看。7 月公司增資約 ₩37兆（$26.5B）去蓋廠、買 EUV。五週後，它承諾用 ₩40兆去消滅股權。增資稀釋約 2.5%、註銷約 3.3%，7 到 12 月的股數幾乎是一趟來回。兩筆錢不是同一筆 — ADS 募資款依約專用於產能，庫藏股來自營運現金 — 但一家手上有 ₩40兆可以拿去註銷股票的公司，卻在六週前才剛增資，這件事在順序上是有態度的，值得讀，不該跳過 ◆。

有三件事同時成立，而且只有第一件是毫無疑義的好事。(1) 機械上它是增益的：股數少約 3.3%、獲利不變，對 2027E EPS 約 +3.0%，本頁已經把它算進目標價。(2) 這則公告抬高了公司自己的執行成本 — 以 ₩1,691,000 而非參考價 ₩1,662,000 計算，₩40兆買到的是約 23.7M 股而不是約 24.07M 股，等於原本要註銷的量，有大約十分之一被「宣布這件事」本身的反應吃掉了。(3) 它是在約 4.7 倍股價淨值比、用循環高點的現金註銷股票，而迎面而來的正是本頁空方情境所描述、公司自己 capex 時程所資助的 2027-29 供給波。記憶體業的資產負債表，過去就是這樣受傷的。7 月的增資說的是「管理層想為下行循環準備銀彈」；8 月的庫藏股說的是「今天的股價很便宜」。同一塊錢，這兩件事不可能同時是報酬率較高的用途 ◆。

✗→ 08-04 「未發現內部人交易」這句話正式作廢。07-30，SK 集團會長崔泰源（최태원）公告場內買進 SK 海力士普通股 3,620 股，均價 ₩1,353,677，合計 ₩49.0億（≈$3.5M）✓ 公告，韓媒×3。把它跟空方拿 ADS 增資當循環頂訊號的讀法擺在一起，正好是鏡像：公司在高點賣股票給市場，會長在腰斬後從市場買股票回來。但要誠實秤重 — ₩49億相對 ₩1,232兆市值、以及他自己的持股，都只是尾數；這是訊號工具，不是經濟行為，而且它出現在漲停的前一個交易日 ◆。

Competitive position競爭態勢

Segment分部	1Q26 share1Q26 份額	Note說明
DRAM	三星38.5% > SK28.8% > Micron美光22.4%	SK first overtook Samsung in 1Q25 (first time since 1992), retaken by Samsung in 4Q25SK 曾於 1Q25 首度超車三星（1992年來首次），4Q25 被奪回 ✓ TrendForce
HBM	SK 56.4% > Samsung三星 > Micron美光	IDC via 424B4; 3Q25 Counterpoint had it 57/22/21IDC引自424B4；3Q25 Counterpoint口徑為57/22/21 ✓
NAND	三星31.6% > SK(incl. Solidigm)集團(含Solidigm)17.6%	eSSD: SK group $4.64B, second placeeSSD:SK集團$4.64B居次 ✓ TrendForce

FY25 operating profit of ₩47.2tn surpassed Samsung's for the first time ✓ CNBC/TrendForce 2026-01. Geography: manufacturing concentrated in Korea (Icheon/Cheongju/Yongin) + China (Wuxi DRAM, Dalian NAND — the VEU review risk) + US packaging (Indiana, 2H28).

FY25 營業利益 ₩47.2兆首度超越三星 ✓ CNBC/TrendForce 2026-01。地理：製造集中韓國（利川/清州/龍仁）+ 中國（無錫 DRAM、大連 NAND — VEU 年審風險）+ 美國封裝（Indiana，2H28）。

R&D depth signal: the CTI research (§Business) is an IEDM 2025 industry first, and the company separately won the 2026 IEEE Corporate Innovation Award for HBM ✓ official官方 2026-04-25.

R&D 深度訊號：CTI 研究（詳事業章節）為 IEDM 2025 業界首例，公司並另外拿下 2026 IEEE Corporate Innovation Award（HBM）✓ 官方 2026-04-25。

Playbook操作框架

Playbook (not investment advice)操作框架（非投資建議）

If not holding未持有	◆ Do not chase. Wait for ₩1,608,700. The entry line moved up with the target — that is the framework working, not a concession — but the price moved up more than twice as far, so the close now sits 5.1% above the line rather than 0.11% above it. A ₩40tn buyback is a real reason for the stock to be worth more; it is not a reason to pay 5% through your own entry rule six sessions after the rule was written. Bid the line.不追價，等 ₩1,608,700。進場線隨目標價往上移 — 那是框架在運作，不是讓步 — 但股價往上移了兩倍不止，所以收盤現在高出這條線 5.1%，而不是像上次那樣只高出 0.11%。₩40兆庫藏股確實是這檔股票應該更值錢的理由；它不是「在自己的進場規則寫下六個交易日後、就穿價 5% 去買」的理由。掛在線上。
If holding已持有	◆ Hold. Adding stays paused — for the third update running, and again on price, not on thesis. Adding resumes at <₩1,608,700 (completing a first tranche); the second tranche still needs its own trigger: <₩1,250,000, or the realized TrendForce 3Q contract print landing at ≥+high-teens% QoQ. Trim tripwires are unchanged and unfired: the 3Q contract print coming in flat/down, or 3Q26 OP on 10-28 below ₩65tn. Nothing here instructs a sale, and the buyback is a mild argument against trimming — there is a ₩40tn bid working in the market through 11-19, which is a real, dated support this page did not have a month ago. But a support bid is a flow, not a valuation.續抱。加碼維持暫停 — 這是連續第三次更新如此，而且一樣是因為價格、不是因為論點。加碼恢復條件：<₩1,608,700（把第一批補完）；第二批仍需自己的觸發：<₩1,250,000，或實際 TrendForce 3Q 合約價落在 ≥+高十位數% QoQ。減碼絆線不變、也未觸發：3Q 合約價持平轉跌，或 10-28 的 3Q26 OP <₩65兆。這裡沒有任何一條指示賣出，而且庫藏股是一個「不要減碼」的溫和理由 — 到 11-19 為止，市場上有一張 ₩40兆的買單在運作，那是本頁一個月前還沒有的、有日期的真實支撐。但支撐買盤是資金流，不是估值。
Entry zone進場區	◆ Re-derived 08-21: < ₩1,608,700 (= ₩1,850,000 ÷ 1.15; ≈5.2× FY27E EPS, ≈4.4× PBR), up from ₩1,565,200. This is the first time the zone has re-derived itself automatically — it was tied to the framework on 08-04 precisely so that it would move when the target moved, and the ₩40tn cancellation moved the target. No judgment was applied to the level; it is arithmetic off the Buy band.08-21 重推：< ₩1,608,700（＝₩1,850,000 ÷ 1.15；≈5.2× FY27E EPS、≈4.4× PBR），自 ₩1,565,200 上移。這是進場區第一次自己重推 — 08-04 把它綁在框架上，為的就是「目標價一動它就跟著動」，而 ₩40兆註銷把目標價動了。這個價位沒有加入任何判斷；它就是買進區間的算術結果。
Why adding is paused為何暫停加碼	◆ Because the target rose +2.8% and the price rose +7.9%. The buyback is the first news since 07-29 that genuinely belongs in the model, and this page put it there — but it is worth ~3% of per-share value, and the market paid ~13% for it in a single session. Meanwhile the operating questions are all still open: the 2Q confirm bar (₩64tn) was missed, ASP deceleration is fact, and the realized 3Q contract print — the number that would move operating profit rather than share count — still has not settled as of 08-21. Buying here means paying a cycle-peak multiple for an accretion that is a board resolution rather than a completed cancellation.因為目標價漲 +2.8%，股價漲 +7.9%。庫藏股是 07-29 以來第一則真正該進模型的消息，本頁也確實把它放進去了 — 但它值的是每股價值約 3%，市場卻在一個交易日內為它付了約 13%。與此同時，營運面的問題全都還開著：2Q 的確認線（₩64兆）沒過、ASP 減速已成事實，而真正會動到營業利益（而不是股數）的 3Q 實際合約價，截至 08-21 仍未定案。在這裡買，等於用循環高點的倍數，去買一個目前還只是董事會決議、尚未完成註銷的增益。
Invalidation失效條件	◆ (1) The realized TrendForce 3Q26 contract print flat/down, or 3Q26 OP <₩60tn → base→bear migration, view downgrade review. (2) HBM share below 50% for two straight quarters, or Samsung reaches HBM4 price parity with NVIDIA → the moat-premium thesis dies, view under review. (3) New 08-21: the ₩40tn buyback executed short, or cancellation deferred past ~12月 → the 08-21 TP revision reverses toward ₩1,800,000, because the whole upgrade is the share count. (upside invalidation) (4) 3Q26 contract prices rise another +30% QoQ and 2027 LTA coverage locks >80% of pricing → this analysis's 2027E was cut too deep, TP upgrade review.(1) TrendForce 3Q26 實際合約價持平/轉跌，或 3Q26 OP <₩60兆 → base→bear 遷移、觀點下修檢討。(2) HBM 份額連兩季 <50% 或三星在 NVIDIA 達成 HBM4 同價 → 護城河溢價死亡、觀點下修。(3) 08-21 新增：₩40兆庫藏股執行未達額，或註銷遞延至約 12 月之後 → 08-21 的目標價修正倒回 ₩1,800,000 方向，因為這次上調的全部就是股數。(上行失效)(4) 3Q26 合約價再 +30% QoQ 且 2027 LTA >80% 鎖價 → 本研究 2027E 砍太深，TP 上修檢討。

For ADR holders (SKHY) — this page in ADS termsADR（SKHY）持有人 — 把本頁換算成 ADS 語言 ◆

1 ADS = 1/10 common share ✓ 424B4. Conversion: KRX price ÷ 10 ÷ FX (₩1,394.47/$ on 08-20; the won has firmed a further ~2.3% from the ₩1,427.53 used on 08-04, so every ≈US$ figure on this page was re-rendered again) ✓. The ₩1,691,000 KRX close = ≈$121.3 of intrinsic value per ADS; SKHY's 08-20 close of $163.08 therefore carries a ≈+34.5% premium ✓ IBKR + computed計算. Premium history: +4.9% at IPO pricing (07-10) → +21.6% on the two 07-28 closes → ≈+31% in the thin 07-29 overnight tape → +30.0% on the 08-03 closes → +34.5% on the 08-20 closes. The premium has broken through the 30% ceiling this page set for itself, and it did so while the KRX leg was rallying — the ADR rose only +4.43% on 08-20 against Seoul's +12.73%, so the premium did not widen because the ADR ran; it widened because it had already run ahead and gave back less than Seoul gained. It still has no anchor, and it can still go to zero.

1 ADS＝1/10 普通股 ✓ 424B4。換算：KRX 股價 ÷10 ÷ 匯率（08-20 為 ₩1,394.47/$；韓元較 08-04 使用的 ₩1,427.53 再升值約 2.3%，因此本頁所有 ≈US$ 數字都再次重算）✓。KRX 收盤 ₩1,691,000＝每 ADS 內含價值 ≈$121.3；SKHY 08-20 收盤 $163.08 因此帶著 ≈+34.5% 溢價 ✓ IBKR＋計算。溢價觀測史：IPO 定價 +4.9%（07-10）→ 07-28 兩地收盤 +21.6% → 07-29 薄量夜盤 ≈+31% → 08-03 兩地收盤 +30.0% → 08-20 兩地收盤 +34.5%。溢價已經穿破本頁替自己設下的 30% 上限，而且是在 KRX 這一腿大漲的過程中穿破的 — ADR 08-20 只漲 +4.43%，首爾漲 +12.73%，所以溢價不是因為 ADR 衝上去而擴大，是因為它本來就跑在前面、這次少補了一段。它依然沒有錨，也依然可以歸零。

Level (KRX)價位（KRX）	= per ADS at 0% premium＝0% 溢價的 ADS	at a 20–25% premium band若維持 20–25% 溢價帶 ◆
TP目標價 ₩1,850,000	$133	$159–166
Entry zone (re-derived)進場區（重推） <₩1,608,700	<$115	<$139–144
Second tranche第二批 <₩1,250,000	<$90	<$108–112
Bear ₩1,020,000	$73	$88–91
Bull ₩3,460,000	$248	$298–310
Where SKHY actually closed 08-20SKHY 08-20 實際收盤	$163.08 = a 34.5% premium on ₩1,691,000 — above the ₩1,608,700 entry line in KRX terms and through the premium ceiling. It fails both legs by a wider margin than at any prior update. After hours it traded $168.46, pricing Seoul's open 08-21 session (intraday ₩1,758,000) at a ≈32% premium＝ ₩1,691,000 上加 34.5% 溢價 — 換算成 KRX 語言已高於 ₩1,608,700 進場線，同時穿破溢價上限。兩條腿都不合格，而且幅度是歷次更新以來最大。盤後成交 $168.46，等於用約 32% 溢價定價首爾 08-21 的盤中（₩1,758,000）

Premium math ◆ computed計算: buying at the $163.08 close and reaching the ₩1,850,000 target in 6 months returns +1.7% / -2.4% / -18.6% as the premium ends at 25% / 20% / zero. Read that row twice, then read it against the last update, where the same three numbers were +10.4% / +6.0% / -11.6%. The KRX-denominated call is +9.4%; bought through the ADR at a 34.5% premium, the best of the three outcomes is now under +2%, and the base outcome is negative. The premium is a second position you own on top of the thesis, and it is now unambiguously the larger of the two variables — a 5-point move in the premium outweighs the entire remaining upside to target.

溢價數學 ◆ 計算：以 $163.08 收盤價買進、6 個月後到達 ₩1,850,000 目標價，溢價收在 25%／20%／歸零 → 報酬 +1.7%／-2.4%／-18.6%。這一行要看兩遍，然後再對照上次更新 — 同樣這三個數字當時是 +10.4%／+6.0%／-11.6%。以 KRX 計價的這個判斷是 +9.4%；若透過帶 34.5% 溢價的 ADR 去買，三種結果裡最好的那個現在不到 +2%，而基準情況是負的。溢價是你在論點之外多持有的第二個部位 — 現在它毫無疑問是兩個變數裡比較大的那一個：溢價動 5 個百分點，就蓋過到目標價為止的全部剩餘空間。

Discipline, unchanged — and now breached by the market, not by the rule: premium ≤25% is the ADR-side entry condition, and 30% is a hard ceiling never to be paid through. At the 08-20 close the premium is 34.5%, i.e. past the ceiling rather than at it. The instruction that follows is not a judgment call: no ADR buying at any size while the premium is above 30%, regardless of what the KRX price does. If Seoul falls into the ₩1,608,700 entry zone next week and the premium is still 34%, the ADR still does not qualify — the KRX leg passing does not rescue a failed premium leg. Execute only in the regular US session (21:30 Taipei), after the ADR has digested the Seoul close: first tranche at ≤$139–144 and ≤25% premium, second tranche mirrors the KRX rule (the realized late-Aug TrendForce print, or <$108–112). For an existing ADR holder none of this instructs a sale; it instructs that the next dollar does not go in here.

紀律，維持不變 — 但這次是被市場穿破，不是被規則放寬：溢價 ≤25% 是 ADR 版的進場條件，30% 是絕不越過的硬上限。08-20 收盤的溢價是 34.5%，也就是已經越過上限，而不是踩在上限。接下來的指令不是判斷題：只要溢價高於 30%，ADR 一股都不買，不管 KRX 價格怎麼走。如果下週首爾跌進 ₩1,608,700 的進場區、而溢價仍在 34%，ADR 依然不合格 — KRX 那條腿過關，救不了溢價這條腿。執行只在美股常規盤（台北 21:30）、且在 ADR 消化完首爾收盤之後：第一批 ≤$139–144 且溢價 ≤25%；第二批同 KRX 規則（8 月下旬實際 TrendForce 數據，或 <$108–112）。對已持有 ADR 的人，以上沒有一條指示賣出；它指示的是：下一塊錢不投在這裡。

Jargon decoder名詞解碼

19 terms used across this page. Basis: the company's own 424B4 filing and newsroom disclosures, plus web-verified industry frameworks.

本文用到的 19 個術語。依據：公司自家 424B4 揭露 + newsroom 資訊 + 上網查證的產業框架。

19 terms — expand when you hit one you don't know19 個術語 — 遇到看不懂的名詞再點開

HBM High Bandwidth Memory: DRAM dies stacked vertically (8/12/16 layers), connected by TSV and sitting on a base logic die, packaged beside the GPU. Bandwidth is the bottleneck for AI training/inference. Why it matters: SK hynix's own MR-MUF stacking process holds a 56.4% 1Q26 share — but HBM contract prices fell in 1Q26 while commodity DRAM rose ~95%, a fact the AI narrative tends to skip. DRAM die 垂直堆疊（8/12/16 層），以 TSV 打通、置於 base die（基底邏輯晶片）上，與 GPU 同封裝。頻寬 = AI 訓練/推論瓶頸。為何重要：SK海力士自家 MR-MUF 堆疊製程握有 1Q26 56.4% 份額 — 但 1Q26 HBM 合約價下跌，同期商品 DRAM 卻漲近 95%，這正是 AI 敘事常忽略的事實。
DRAM Dynamic Random Access Memory: Volatile working memory for CPU/GPU — written on demand, erased on power-off. Server-grade uses RDIMM modules; high-capacity (128GB+) shipments doubled QoQ in 3Q25. Why it matters: 77.3% of 1Q26 revenue — this is the body of the business, not HBM. 揮發性記憶體，CPU/GPU 的工作記憶。一次寫入運算資料、斷電即失。伺服器用 RDIMM 模組；128GB+ 高容量款 3Q25 出貨 QoQ 倍增。為何重要：佔 1Q26 營收 77.3% — 這才是公司的本體，不是 HBM。
NAND NAND Flash: Non-volatile storage. AI data centers drive eSSD (enterprise SSD) and high-capacity QLC demand; subsidiary Solidigm leads the QLC eSSD line. Why it matters: 22.0% of 1Q26 revenue; the CTI research (§Business) targets exactly the cell-to-cell interference that QLC's narrow voltage margins are most sensitive to. 非揮發儲存。AI 資料中心帶動 eSSD（企業級 SSD）與 QLC 高容量需求；子公司 Solidigm 是 QLC eSSD 主力。為何重要：佔 1Q26 營收 22.0%；事業章節的 CTI 研究，正是針對 QLC 最敏感的 cell 間干擾問題。
DDR5/RDIMM: The current server DRAM generation and its registered-module form factor. Profitability overtook HBM3E in 2026, as TrendForce predicted in Oct-2025. Why it matters: this is the "conventional DRAM" whose contract price is the actual 2026 profit engine — not HBM. 目前的伺服器 DRAM 世代與其模組形式（registered DIMM）。獲利性 2026 年超車 HBM3E，如 TrendForce 於 2025 年 10 月的預測。為何重要：這正是「conventional DRAM」— 其合約價才是 2026 真正的獲利引擎，不是 HBM。
TSV Through-Silicon Via: Vertical electrical channels drilled through stacked dies, the physical mechanism that lets an HBM stack behave as one fast memory column instead of separate chips. Why it matters: TSV yield and density are core HBM manufacturing know-how — part of why HBM share is sticky. 貫穿堆疊晶片的垂直電氣通道，是讓 HBM 堆疊像單一高速記憶體柱、而非各自獨立晶片的實體機制。為何重要：TSV 良率與密度是 HBM 製造的核心 know-how — HBM 份額之所以黏著的原因之一。
Base die基底邏輯晶片 Base Die: The bottom logic die of an HBM stack, sitting nearest the GPU/interposer, that controls the DRAM dies stacked above it. Why it matters: the base die is where MR-MUF (below) is applied — the encapsulation choice that separates SK hynix's thermal performance from Samsung/Micron's. HBM 堆疊最底層的邏輯晶片，最靠近 GPU/中介層，負責控制上方堆疊的 DRAM die。為何重要：base die 正是下方 MR-MUF 製程施作的位置 — 這是 SK海力士散熱表現與三星/美光拉開差距的關鍵選擇。
MR-MUF vs TC-NCF: SK hynix's proprietary stacking process: a single mass reflow, then liquid epoxy fill (Mass Reflow-Molded Underfill) — vs Samsung/Micron's TC-NCF (thermo-compression, layer-by-layer film lamination). MR-MUF's thermal bumps run ~4× denser, improving heat dissipation. Why it matters: this is the single biggest technical differentiator behind SK's HBM lead. SK海力士獨門堆疊製程：一次回焊 + 液態環氧填充（Mass Reflow-Molded Underfill）— 相對三星/美光的 TC-NCF（熱壓合、薄膜逐層壓合）。MR-MUF 的熱阻尼凸塊密度多約 4 倍，散熱更佳。為何重要：這是 SK HBM 領先地位背後最大的單一技術差異化。
Contract vs spot price合約價 vs 現貨價: Memory is sold mostly on negotiated multi-month contract prices, with a smaller spot market as a sideshow. 1Q26 conventional DRAM contract prices rose +93~98% QoQ. Why it matters: the whole 2026 windfall is a contract-price story, not a spot-auction story — it moves on CSP negotiations, not headlines. 記憶體多以協商的多月期合約價銷售，現貨市場只是配角。1Q26 conventional DRAM 合約價 QoQ +93~98%。為何重要：2026 整個暴利故事是合約價故事，不是現貨拍賣故事 — 它隨 CSP 議約而動，不是隨頭條新聞而動。
LTA Long-Term Agreement: A multi-quarter/year supply-and-price lock between a memory maker and a customer (typically a CSP). Why it matters: LTA coverage above 80% into 2027 would lock in high prices and argue for raising this analysis's 2027E; below 50% HBM share retained via LTA is one of this analysis's own invalidation triggers. 記憶體廠與客戶（通常是 CSP）之間多季/多年期的供給與價格鎖定協議。為何重要：若 2027 LTA 覆蓋率超過 80%，代表高價已被鎖定，是本研究 2027E 上修的觸發條件之一；反之 LTA 內 HBM 份額跌破 50%，則是本研究自己設下的失效條件。
eSSD/QLC: Enterprise SSD, and Quad-Level Cell — 4 bits (16 voltage states) per cell, the densest/cheapest-per-GB NAND format, with the narrowest margin between voltage states. Why it matters: QLC is the format most exposed to the cell-interference problem the CTI research (§Business) targets — and Solidigm's AI-storage story rides on it. 企業級 SSD，以及 Quad-Level Cell — 每 cell 4 位元（16 個電壓態），密度最高、每 GB 成本最低的 NAND 格式，但電壓態之間的餘裕也最窄。為何重要：QLC 正是最容易受 cell 間干擾影響的格式（詳事業章節 CTI 研究）— Solidigm 的 AI 儲存故事就建立在它之上。
Bit growth位元成長率: Industry shorthand for total memory bits shipped, growing YoY — the volume half of revenue = price × volume. Why it matters: 2026's windfall is overwhelmingly a price story; this analysis separately tracks bit growth to make sure volume isn't quietly doing the work instead. 產業慣用語，指出貨位元總量的年增 — 是「營收＝價格×數量」中的數量那一半。為何重要：2026 暴利絕大部分是價格故事；本研究另外追蹤位元成長率，確保不是數量悄悄扛下大半功勞。
1c nm: The current (6th-generation, 10nm-class) DRAM process node, in mass production. Why it matters: node shrinks are how bit density (and margin) improve without new fabs — a lever behind the gross-margin path in the model. 目前量產中的 DRAM 製程節點（第 6 代，10nm 級）。為何重要：製程微縮是在不新建晶圓廠的前提下提升位元密度（與毛利）的手段 — 是財測模型毛利路徑背後的槓桿之一。
EUV Extreme Ultraviolet Lithography: The lithography technology required for the most advanced logic and memory nodes; ASML is the sole supplier. Why it matters: the July ADS raise earmarked ₩11.9tn specifically for EUV tools, delivery before 2027-12 — a direct funding line for the next node. 最先進邏輯與記憶體製程節點所需的曝光技術；ASML 為唯一供應商。為何重要：7 月 ADS 增資明確撥款 ₩11.9兆用於 EUV 設備，2027-12 前交機 — 直接對應下一代製程的資金來源。
VEU Validated End-User: A US export-control status letting a designated China facility keep receiving US-origin chipmaking equipment without a case-by-case license. Wuxi (DRAM) and Dalian (NAND) operate under it, now on an annual review cycle starting 2026. Why it matters: a denied or tightened renewal is one of this analysis's named tail risks. 美國出口管制下的一種資格認定，讓指定的中國廠區得以持續取得美國原產的晶片設備、免逐案申請許可。無錫（DRAM）與大連（NAND）廠依此運作，2026 年起改為年度換照制。為何重要：換照被拒或加嚴，是本研究列名的尾部風險之一。
ADS/ADR American Depositary Share/Receipt: A US-listed instrument representing foreign shares held by a depositary bank; SK hynix's ADS (1 ADS = 1/10 common share, ticker SKHY) began trading 2026-07-10 after a $26.5B raise — the largest-ever US listing by a foreign company. Why it matters: the raise is simultaneously a fortress-balance-sheet move and a top-of-cycle share-sale signal — both readings appear in the Variant section. 代表存託銀行持有之外國股份、在美國掛牌交易的憑證；SK海力士 ADS（1 ADS＝1/10 普通股，代號 SKHY）於 2026-07-10 掛牌，募資 $26.5B — 外國公司在美史上最大規模股票發行。為何重要：這筆增資同時是「堡壘資產負債表」與「頂部賣股訊號」— 兩種讀法都出現在變異認知章節。
OPM Operating Profit Margin: Operating profit ÷ revenue — the company's core profitability, unaffected by one-off financial items. 1Q26 companywide OPM hit 72%; 2Q26 came in at 76.3%. Why it matters: the model tracks OPM, not net income, because net income keeps getting distorted upward by one-offs — ₩17.06tn of financial gains in 1Q26, then ₩63.27tn of Kioxia-related investment gains in 2Q26 (net margin 118%, i.e., net profit above revenue). 營業利益 ÷ 營收 — 公司核心獲利能力，不受業外一次性項目影響。1Q26 全公司 OPM 達 72%；2Q26 實績 76.3%。為何重要：模型以 OPM 而非稅後純益為錨，因為淨利一再被一次性項目向上扭曲 — 1Q26 是 ₩17.06兆金融收益，2Q26 是 ₩63.27兆 Kioxia 相關投資利益（淨利率 118%，淨利比營收還高）。
CTI Charge Trap Nitride Isolation: A newly-disclosed NAND research breakthrough: segmenting the charge-storage layer per cell (instead of one continuous film) by depositing oxide barriers rather than etching them — avoiding the pore-widening that etch-based isolation causes. Why it matters: company-stated results (-30%+ interference, +45%+ retention, >10% smaller cells) are a research milestone, not a shipping product — see §Business for the full framing. 新揭露的 NAND 研究突破：把儲存電荷的絕緣層依 cell 分段隔離（而非一整條連續薄膜），方法是沉積氧化屏障而非蝕刻挖孔 — 避免蝕刻式隔離會造成的孔徑變寬問題。為何重要：公司自述成果（干擾 -30%+、保持力 +45%+、cell 縮 >10%）屬研究里程碑，非出貨中產品 — 完整框架見事業章節。
CTN Charge Trap Nitride: The insulating nitride layer SK hynix's NAND cells use to trap electrons representing a "1" or "0" — traditionally one continuous film running the length of the channel. Why it matters: a continuous CTN is the root cause CTI (above) is designed to fix — as cells shrink closer together, electrons drift between neighboring cells (Vth shift, read errors), one of NAND scaling's biggest physical bottlenecks. SK海力士 NAND cell 用來困住代表「1」或「0」電子的絕緣氮化層 — 傳統上是沿著 channel 連續的一整條薄膜。為何重要：連續 CTN 正是上方 CTI 要解決的根源問題 — cell 越縮越密，電子會在鄰近 cell 間游移（閾值電壓 Vth 漂移、讀取錯誤），是 NAND 微縮最大的物理瓶頸之一。
₩1兆 1 Trillion Won: The scale unit used throughout this page for revenue/profit figures (₩ = KRW). ₩1 trillion (兆) ≈ US$650-720M at recent FX (₩1,380-1,540/US$); this page annotates ≈US$ at ₩1,395/$ (08-20); the July ADS raise's $26.5B gross proceeds ≈ ₩40tn at the 424B4's reference rate. Per-share figures (EPS, dividend, book value, price) are in plain won, not trillions — always check which scale a number is on before comparing across tables on this page. 本頁營收/利潤數字採用的規模單位（₩＝韓元）。₩1兆以近期匯率（₩1,380-1,540/US$）約當 US$6.5-7.2億；本頁 ≈US$ 標註以 ₩1,395/$（08-20）換算；7 月 ADS 增資 $26.5B 毛額，以 424B4 參考匯率折算約 ₩40兆。每股數字（EPS、股息、帳面淨值、股價）皆為原始韓元，非兆為單位 — 比較本頁不同表格數字前，務必先確認單位刻度。

✓ SK hynix newsroom + EETimes Schematic, not to scale — layer counts and gaps are drawn for legibility. MR-MUF is SK hynix's stacking process; the industry alternative is TC-NCF (film placed and thermally compressed layer by layer), used by Samsung and Micron. See the HBM cross-section in §Business for how the stack sits beside the GPU.示意圖，非比例尺 — 層數與縫隙為求辨識而放大。MR-MUF 是 SK 海力士的堆疊製程；業界的另一條路線是 TC-NCF（逐層鋪膜、熱壓接合），三星與美光採用。整疊如何擺在 GPU 旁邊，見事業章節的 HBM 剖面圖。

Data provenance — every number, sourced資料來源 — 每個數字都有出處

130 claims: 113 ✓ verified · 12 ⚠ flagged · 5 ✗→ corrected — expand the ledger130 項查證：113 ✓ 已查證 · 12 ⚠ 未核實 · 5 ✗→ 已修正 — 點開展開逐項表

Claim主張	Verdict判定	Value · source · as-of數值 · 來源 · 日期
The 08-04 update · tape, HBF standard, sell-side reset08-04 更新 · 盤面、HBF 標準、賣方重置
98. Closes: 07-30 ₩1,322,000 (-5.64%) · 07-31 ₩1,718,000 (+29.95%, limit-up) · 08-03 ₩1,567,000 (-8.80%); 08-04 intraday ₩1,514,000 (-3.38%) at 10:23 KST收盤:07-30 ₩1,322,000(-5.64%)·07-31 ₩1,718,000(+29.95%,漲停)·08-03 ₩1,567,000(-8.80%);08-04 盤中₩1,514,000(-3.38%)於10:23 KST	✓ Korean press ×2 + limit-band arithmetic韓媒×2＋漲停幅度反推	topstarnews/ggilbo · 07-31, 08-03
99. Correction: 07-29 close is ₩1,401,000 (-9.61%), not the ₩1,377,000 this page published on 07-29; intraday low ₩1,246,000 (-19.61%)更正：07-29 收盤為 ₩1,401,000(-9.61%)，非本頁 07-29 所載的 ₩1,377,000；當日盤中低點 ₩1,246,000(-19.61%)	✗→✓	mt.co.kr + Newsis 07-29; cross-checked: 07-28 close 1,550,000 × (1−9.61%) = 1,401,000, and 07-30's 1,322,000 = 1,401,000 − 79,000. Root cause: stockanalysis labels an intraday quote "closing price" while the KRX session is open — reproducible on 08-04.mt.co.kr＋뉴시스 07-29；交叉驗算：07-28 收 1,550,000 ×(1−9.61%)=1,401,000，且 07-30 的 1,322,000 = 1,401,000 − 79,000。成因：stockanalysis 在 KRX 盤中會把即時報價標成「closing price」— 08-04 可複現。
100. Market cap ₩1,142.13tn at the 08-03 close市值 ₩1,142.13兆（08-03 收盤）	✓ computed, matches stockanalysis exactly計算，與 stockanalysis 完全吻合	1,567,000 × 728,865,500 shares1,567,000 × 728,865,500 股
101. USDKRW 1,427.53 (08-04, -0.13%); page annotates ≈US$ at a rounded ₩1,430/$USDKRW 1,427.53(08-04,-0.13%)；頁上 ≈US$ 以 ₩1,430/$ 標註	✓ TradingEconomics	08-04
102. SKHY 08-03 close $142.72; range since listing $124.80–194.80SKHY 08-03 收盤 $142.72；掛牌以來區間 $124.80–194.80	✓ IBKR	08-03/08-04
103. ADS premium +30.0% on the 08-03 closes; 07-29 premium restated ≈34% → ≈31% on the corrected KRX legADS 溢價 08-03 兩地收盤 +30.0%；07-29 溢價依更正後 KRX 腿自 ≈34% 改為 ≈31%	✓ computed (#98/#99/#101/#102)計算(#98/#99/#101/#102)	—
104. 07-31 was SK hynix's first daily limit-up since 2009 and its first under the 30% band adopted in 2015; drivers = MSFT/AMZN results 07-30, SOX +8.19%, foreign net buying ~₩5tn07-31 為 SK 海力士 2009 年以來首次漲停、2015 年放寬至 30% 後首見；動因＝07-30 微軟/亞馬遜財報、費半 +8.19%、外資淨買超約 ₩5兆	✓ three outlets agree三家media一致	seoul.co.kr/newsway/mt · 07-31
105. Chairman Chey Tae-won bought 3,620 shares at ₩1,353,677 on 07-30 = ₩4.90bn (≈$3.5M)崔泰源會長 07-30 買進 3,620 股 @₩1,353,677 ＝ ₩49.0億(≈$3.5M)	✓ filing, three outlets; arithmetic checks (3,620 × 1,353,677 = ₩4,900,310,740)公告，三家media；數字自洽(3,620×1,353,677=₩4,900,310,740)	insight/newspim/daum · 07-30
106. Post-print sell-side reset: 7 cuts, 2 raises, 9 held; range ₩1,480,000–4,700,000; roster average ≈₩3,296,000財報後賣方重置:7降/2升/9維持;區間₩1,480,000–4,700,000;名單平均約₩3,296,000	✓ roster名單 ⚠ average is this analysis's computation over one article's 18 houses, not the full 37-analyst universe平均為本研究就單一報導之18家計算，非完整37位樣本	g-enews · 07-30
107. TrendForce 3Q26: conventional DRAM contract +13–18% QoQ, NAND +10–15% QoQ; server DRAM +13–18% with LTAs capping increasesTrendForce 3Q26:conventional DRAM合約價+13–18% QoQ、NAND +10–15% QoQ;伺服器DRAM +13–18%，長約封住漲幅	✓ TrendForce ×2	trendforce.com · 07-03 / 07-09
108. HBF OCP spec: up to 512GB, 8-high & 16-high stacks, three grades ≈0.4–3.0TB/s, UCIe chiplet interconnect; covers host interface, electrical, reliability/packaging, software I/OHBF OCP規格:最高512GB、8層與16層堆疊、三等級約0.4–3.0TB/s、UCIe互連;涵蓋主機介面、電氣、可靠度/封裝、軟體I/O	✓ both partners' releases agree雙方release一致	SK hynix / Sandisk · 08-03
109. HBF is physically compatible with HBM4 (same pin layout, dimensions, power profile)HBF 與 HBM4 實體相容（腳位、尺寸、功耗規格相同）	✓ Sandisk release ⚠ stated on the Sandisk side only; not repeated in SK hynix's own release僅見於 Sandisk 一方，SK 自家 release 未複述	08-03
110. HBF consortium = SK hynix + Sandisk + Google + Tenstorrent, formed 2026-02; partners' own guidance puts demand pickup "around 2030"; no HBF sampling/production date disclosedHBF聯盟＝SK hynix+Sandisk+Google+Tenstorrent，2026-02成立;雙方指引需求起量「約2030年」;未揭露HBF出樣/量產日	✓ Feb + Aug releases2月與8月release	2026-02-25 / 08-03
111. 375-layer 4D NAND (10th gen, V10): first shown at FMS 2026, 2.5× performance-per-watt vs prior gen; eSSD mass production guided "early next year" (2027)375層4D NAND(第十代,V10):FMS 2026首度展出、每瓦效能較前代2.5倍;eSSD量產指引「明年初」(2027)	✓ official release官方release	08-03
112. FMS 2026: 08-04~06, Santa Clara Convention Center; SK hynix keynote 08-04 (EVP Kim Chun-sung, VP Kang Uk-song); 08-06 HBF panel with Sandisk VP Rajeev Nagabhirava and Google DeepMind's Xiaoyu MaFMS 2026:08-04~06,Santa Clara會議中心;SK 08-04 keynote(EVP 金春成、VP 姜旭松);08-06 HBF 座談，與 Sandisk VP Rajeev Nagabhirava、Google DeepMind Xiaoyu Ma 同台	✓ official release官方release	08-03
113. Sandisk–Kioxia NAND JV (Yokkaichi + Kitakami) extended through 2034-12-31; Sandisk committed US$1.17B of manufacturing payments 2026–29; 10th-gen 3D flash production started at Kitakami Fab2 on 07-03Sandisk–Kioxia NAND合資(四日市+北上)延長至2034-12-31;Sandisk承諾2026–29投入US$1.17B製造款;第十代3D flash於07-03在北上Fab2投產	✓ both companies' releases雙方公司release	2026-01-29/30 · 2026-07-02/03
114. HBF technical roadmap (predates the 08-03 spec): Sandisk targets first HBF samples 2H CY2026 and samples of the first AI-inference devices carrying HBF in early 2027 — from the original 2025-08 partnershipHBF 技術時程(早於 08-03 規格):Sandisk 目標 2026 下半年首批 HBF 樣品、2027 年初首批搭載 HBF 的 AI 推論裝置樣品 — 出自 2025-08 最初合作案	✓ TrendForce + trade pressTrendForce＋產業媒體	TrendForce 2026-01-29 · Tom's Hardware
115. Full HBF commercialisation guided to ~2030HBF 全面商業化指引約 2030 年	✓ two independent sources兩個獨立來源	SK newsroom 2026-02 ("around 2030") + Seoul Economic Daily 08-04官方 newsroom 2026-02（「約 2030」）＋首爾經濟 08-04
116. Sandisk holds HBF standardisation MOUs with Samsung as well as SK hynix; Samsung began conceptual design/early HBF development 2025-10; Samsung + Sandisk reported to target NVIDIA/AMD/Google integration late 2027–early 2028Sandisk 在 HBF 標準化上與三星同樣簽有 MOU；三星 2025-10 展開概念設計與早期開發；據報三星＋Sandisk 瞄準 2027 末–2028 初導入 NVIDIA/AMD/Google	✓ TrendForce + DigiTimesTrendForce＋DigiTimes ⚠ competitor plans are press-reported (orig. FN News), not company-confirmed競爭者計畫為媒體轉述(原始出處 FN News)，非公司證實	TrendForce 2025-11-11 / 2026-01-29 · DigiTimes 2025-10-07
117. Kioxia demonstrated a 5TB high-speed HBF prototype in 2025-08 — a year before this specificationKioxia 於 2025-08 展示 5TB 高速 HBF 原型 — 比本規格早一年	✓ reported轉述 ⚠ trade press, not a Kioxia release this analysis fetched產業媒體，非本研究取得之 Kioxia 官方稿	TrendForce / DigiTimes
118. FMS 2026 keynote title: "Orchestrating Efficient AI Infrastructure through Tiered Memory in the Era of Agentic AI" (08-04); official SK hynix newsroom page confirms every 08-03 spec figure and does not state HBM4 compatibility — which is why #109 keeps its ⚠FMS 2026 keynote 題目:「Orchestrating Efficient AI Infrastructure through Tiered Memory in the Era of Agentic AI」(08-04);官方 newsroom 頁確認 08-03 所有規格數字，且未載明與 HBM4 相容 — 這正是 #109 保留 ⚠ 的原因	✓ official newsroom官方 newsroom	news.skhynix.com/en/hbf-at-fms-2026/ · 08-04
2Q26 actuals & the 07-29 update · official release + call coverage2Q26 實績與 07-29 更新 · 官方release+法說報導
85. 2Q26 rev ₩79.3187tn / OP ₩60.5426tn / OPM 76%2Q26 營收₩79.3187兆/OP₩60.5426兆/OPM 76%	✓ official release官方release	news.skhynix.com/PRNewswire · 07-29
86. NI ₩93.9226tn (net margin 118%), pretax ₩122.71tn, incl. ₩63.27tn one-off investment gains (Kioxia stake disposal + SPC valuation)淨利₩93.9226兆(淨利率118%)、稅前₩122.71兆，含一次性投資利益₩63.27兆(Kioxia處分+SPC評價)	✓ official + sedaily breakdown官方+sedaily細拆	07-29
87. Cash ₩88tn (+33.6tn QoQ) / debt ₩18.6tn / net cash ₩69.4tn at end-2Q262Q26末現金₩88兆(+33.6兆QoQ)/負債₩18.6兆/淨現金₩69.4兆	✓ official release官方release	07-29
88. Blended DRAM ASP +~30% QoQ, NAND +mid-50% QoQ (call)DRAM混合ASP +約30% QoQ、NAND +mid-50% QoQ(法說)	✓ two call recaps agree兩家法說轉述一致	tradingkey/wallstengine · 07-29
89. 2026 capex guidance raised to the high-₩40tn range2026 capex 上修至 high-₩40兆檔	✓ two call recaps agree兩家法說轉述一致	tradingkey/BigGo · 07-29
90. LTAs finalized with ~10 customers incl. key strategic partners; "~half of total sales" coverage與約10家客戶(含主要策略夥伴)簽訂長約；「約佔營收一半」	✓ official (10 customers)官方(10家) ⚠ half-of-sales ratio single-source占比為單源轉述	官方/tradingkey · 07-29
91. HBM4 mass shipments began in 2Q, ramp through 2H; HBM4E samples completed 1HHBM4 2Q開始量產出貨、2H放量；HBM4E 1H完成出樣	✓ official release官方release	07-29
92. 07-29 close ₩1,401,000 (-9.61%); 07-28 close ₩1,550,000; two-day -20.4% (computed)07-29收盤₩1,401,000(-9.61%)；07-28收₩1,550,000；兩日-20.4%(計算)	✗→✓ restated 08-04, see #9908-04 更正，見 #99	was published as ₩1,377,000 (an intraday print)原載 ₩1,377,000（盤中價） · 07-29
93. First consensus-OP miss after 10 straight beats (60.54 vs FnGuide 64.16, -5.6%)連10季beat後首次miss(60.54 vs FnGuide共識64.16,-5.6%)	✓ computed vs #77/#78依#77/#78計算	07-29
94. 2026 industry demand guidance: DRAM +mid-20% YoY, NAND +high-teens% (call)2026產業需求指引:DRAM +mid-20% YoY、NAND +high-teens%(法說)	✓ two call recaps agree兩家法說轉述一致	wallstengine/BigGo · 07-29
95. USDKRW 1,447.9 (06:52 UTC); page annotates ≈US$ at a rounded ₩1,450/$USDKRW 1,447.9(06:52 UTC)；頁上 ≈US$ 以四捨五入的 ₩1,450/$ 標註	✓ Google Finance	07-29
96. SKHY ADR pre-market $127.19 (-2.29%); 07-28 US close $130.17SKHY ADR 盤前$127.19(-2.29%)；07-28 美股收盤$130.17	✓ IBKR snapshot快照	07-29
97. ADS premium = ADR ÷ (KRX÷10÷FX) − 1: ≈+33.7% pre-market / +21.6% on 07-28 closes / +4.9% at IPO pricingADS溢價=ADR÷(KRX÷10÷匯率)−1:盤前≈+33.7%/07-28兩地收盤+21.6%/IPO定價+4.9%	✓ computed (#51/#92/#95/#96)計算(#51/#92/#95/#96)	07-29
Price & valuation · IBKR snapshot + computed, 2026-07-22 (initial-coverage record; masthead carries the live numbers)價格與估值 · IBKR快照+計算 2026-07-22(初次評估紀錄;現值見頁首)
1. Close (+0.16%)收盤(+0.16%)	✓ IBKR	₩1,839,000 · 07-22
2. YTD +182.97%	✓ IBKR	stockanalysis same-day ~+181%, basis differs (651,000 vs 677,000 base, see #77)stockanalysis同日約+181%,口徑差異(基期651,000 vs 677,000,見#77)
3. 52wk ₩245,000–2,987,000, peak 2026-06-2652週₩245,000–2,987,000,高點2026-06-26	✓ IBKR + en.sedaily	07-02
4. Off peak -38.4%距峰-38.4%	✓ computed計算	—
5. 12mo +~576%12個月+~576%	✓ IBKR	open_52w 271,909; stockanalysis +571.56% close開盤基期271,909;stockanalysis口徑+571.56%相近
6. Market cap ₩1,340tn市值₩1,340兆	✓ computed計算	price × 728,865,500 shares (424B4)價格×728,865,500股(424B4)
7. ≈US$871B≈US$871B	✓ basis stated基準明示	424B4 FX 1,538.05 (07-09), not a live rate424B4匯率1,538.05(07-09),非即時匯率
8. PER TTM 17.4×	✓ computed計算	cross-check stockanalysis 07-22交叉核對stockanalysis 07-22
9. fwd PER 5.8×(consensus)/6.2×(this analysis)fwd PER 5.8×(共識)/6.2×(本研究)	✓◆	computed計算
10. PBR 7.9× (BVPS 231,950)PBR 7.9×(BVPS 231,950)	✓ computed from 424B4計算自424B4	—
11. Yield 0.16% (FY25 ₩3,000 basis)殖利率0.16%(FY25₩3,000基準)	✓ 424B4+computed計算	IBKR same basis 0.17%IBKR同口徑0.17%
12. Sell-side avg TP ₩3,408,502 (1.2M–5.3M, 37 analysts)賣方平均TP₩3,408,502(1.2M–5.3M,37位)	✓ opinion屬意見	stockanalysis/investing 07-22
Financials · SEC form 424B4, audited財務 · SEC 424B4,經審計數
13-17. FY23 rev 32.77/OP -7.73/NI -9.14/EPS -13,244; FY24 66.19/23.47/19.80/28,732	✓ 424B4	—
18-21. FY25 rev 97.15/OP 47.21(49%)/NI 42.95/EPS 62,044 (released 2026-01-28發布2026-01-28)	✓ 424B4+official release官方release	—
22-26. 2025 quarterly rev/OP: 1Q 17.64/7.44, 2Q 22.23/9.21, 3Q 24.45/11.38, 4Q 32.83/19.17 (sums to FY)2025各季rev/OP:1Q17.64/7.44,2Q22.23/9.21,3Q24.45/11.38,4Q32.83/19.17(加總=年度)	✓ official quarterly releases + 424B4官方季度release+424B4	—
27-30. 1Q26 rev 52.58(+60%QoQ,+198%YoY)/OP 37.61(72%)/NI 40.35/EPS 57,175	✓ official 04-22官方04-22+424B4	—
31. 1Q26 NI>OP: ₩17.06tn financial gain (pretax 51.6tn, tax 11.3tn)1Q26淨利>營益:金融收益₩17.06兆(稅前51.6兆,稅11.3兆)	✓ composition not broken out組成未細拆 ⚠	424B4
32. 2Q26 release date 2026-07-29 09:00 KST (not yet released)2Q26發布日2026-07-29 09:00 KST(未發布)	✓ 6-K	07-15 (TipRanks 07-22 was wrong, droppedTipRanks 07-22說法錯誤,棄用)
33-35. Segment: 1Q26 DRAM 77.3%/NAND 22.0%; FY25 77.1%/21.3%; FY24 67.6%/29.1%段別:1Q26 DRAM77.3%/NAND22.0%;FY25 77.1%/21.3%;FY24 67.6%/29.1%	✓ 424B4	—
36. HBM >40% of DRAM revenue (disclosed 4Q24, no later %)HBM>40% DRAM營收(4Q24揭露,其後無明確%)	✓ time-stamped時效註明	official 4Q24 release官方4Q24 release
37. 1Q26 net cash ₩35tn (current assets 54.3tn vs borrowings 19.3tn)1Q26淨現金₩35兆(流動資產54.3兆vs借款19.3兆)	✓ 424B4+official官方	—
38. FY25 capex ₩27.52tn (cash-flow basis; media's 30.2tn figure conflicts, dropped)FY25 capex₩27.52兆(現金流基準;媒體30.2兆口徑衝突,棄用)	✓ 424B4	—
39. Unrecognized capital commitments ₩30.9tn (3/31) vs ₩6.7tn (12/31)未認列資本承諾₩30.9兆(3/31)vs₩6.7兆(12/31)	✓ 424B4	—
40. Equity ₩164.38tn, assets ₩222.83tn (3/31)權益₩164.38兆,資產₩222.83兆(3/31)	✓ 424B4	—
41-43. Shares: 728,002,365 → 15.3M retired → 712,702,365; outstanding 711,075,500 (07-09); post-ADS 728,865,500股數:728,002,365→2月註銷15.3M→712,702,365;流通711,075,500(07-09);ADS後728,865,500	✓ 424B4	—
44. Dividend policy FY25-27 fixed ₩1,500/yr + special upside; FY25 paid 3,000 (incl. 1,500 special); FY24 2,204, FY23 1,200股息政策FY25-27固定₩1,500/年+超額加碼;FY25實付3,000(含特別1,500);FY24 2,204,FY23 1,200	✓ 424B4	—
45. $1.7B exchangeable bond clean-up call (exercised 04-28, completed 05-28)$1.7B交換債clean-up call(04-28行使,05-28完成)	✓ 424B4	—
ADR (SKHY)
46. NASDAQ listing 2026-07-10, ticker SKHY (initial SKHYV, switch date uncertain)NASDAQ 2026-07-10掛牌,代碼SKHY(初期SKHYV⚠切換日不確定)	✓ EDGAR/424B4/Bloomberg	—
47. 1 ADS = 1/10 share; Citibank depositary1 ADS=1/10股;花旗存託	✓ 424B4	—
48. 177.9M ADS @ $149 = $26.5B gross, ~$26.2B net; 17.79M new shares (~2.5%)177.9M ADS@$149=毛額$26.5B,淨額~$26.2B;全新股17.79M股(~2.5%)	✓ 424B4	—
49. Largest-ever US listing by a foreign company (>Alibaba's $25B); ~7× oversubscribed外國公司在美最大股票發行(>Alibaba $25B);~7倍超額認購	✓ ⚠ oversubscription is single-source超購為單一消息源	Reuters/CNBC 07-10
50. Use of proceeds: ₩45.5tn Korea fab capex + ₩11.9tn EUV (by 2027-12)資金用途:₩45.5兆韓國廠房capex+₩11.9兆EUV(2027-12前)	✓ 424B4	—
51. IPO priced ~4.9% above 07-09 KOSPI close (₩2,186,000, FX 1538.05)IPO定價較07-09 KOSPI收盤(₩2,186,000,FX1538.05)溢價~4.9%	✓ computed from 424B4計算自424B4	—
52. 90-day company lockup; no greenshoe (Korean law); ~$7B cornerstone (Baillie Gifford/Coatue/SAP LP)90天公司lockup;無綠鞋(韓國法規);基石~$7B(Baillie Gifford/Coatue/SAP LP)	✓ 424B4	—
53. Luxembourg GDR "HYNSE" persists (1 GDS=1 share); no conversion link to OTC HXSCL盧森堡GDR「HYNSE」續存(1GDS=1股);與OTC HXSCL無轉換關係	✓ full-text search of 424B4424B4全文檢索	—
Industry & competition (named sources)產業與競爭(具名來源)
54. 1Q26 DRAM share: Samsung 38.5%/SK 28.8%/Micron 22.4% ($97B industry, +81% QoQ)1Q26 DRAM份額:三星38.5%/SK28.8%/美光22.4%($97B產業,+81%QoQ)	✓ TrendForce	06-01
55. SK first led DRAM in 1Q25 (first since 1992), retaken 4Q25SK於1Q25首度DRAM登頂(1992來首次),4Q25被奪回	✓ Counterpoint+TrendForce	2025-04 / 2026-02-26
56. HBM share 56.4% (1Q26, IDC)HBM份額56.4%(1Q26,IDC)	✓ cited via引自 424B4	—
57. NAND: 1Q26 Samsung 31.6%/SK group 17.6% ($7.53B)NAND:1Q26三星31.6%/SK集團17.6%($7.53B)	✓ TrendForce	05-25
58. FY25 OP first surpassed Samsung'sFY25 OP首度超越三星	✓ CNBC/TrendForce	2026-01-28
59. Conventional DRAM contract +93~98% QoQ 1Q26; +58~63% 2Q26Econventional DRAM合約價1Q26+93~98%QoQ;2Q26E+58~63%	✓ TrendForce	06-01/03-31
60. HBM contract price fell 1Q26, pressuring SK's ASPHBM合約價1Q26下跌,壓SK ASP	✓ TrendForce	06-01
61. DDR5 profitability overtook HBM3E in 2026 (predicted)DDR5獲利性2026超HBM3E(預測)	✓ TrendForce	2025-10-29
62. SK slowed HBM4 conversion; full-year target 4.5→4B Gb; HBM already >40% of revenueSK放慢HBM4轉線;全年目標4.5→4B Gb;HBM已>40%營收	✓ reported轉述韓媒	TrendForce 06-23
63. Samsung HBM4 mass production 2026-02, >$1B in 4 months, full-year target upgraded to 4B Gb, seeking price parity三星HBM4 2026-02量產,4個月>$1B,全年上修至4B Gb,尋求同價	✓ reported轉述	TrendForce 06-23/01-28
64. Micron HBM3E+HBM4 booked through 2027, HBM4 >$1B美光HBM3E+HBM4訂到2027,HBM4>$1B	✓ Micron FQ3'26	2026-06-24
65. NVIDIA Rubin: all three vendors certified (Jensen Huang's 06-05 Korea visit); SK share ~2/3 (estimated); multi-year agreement 06-08NVIDIA Rubin:三家全過認證(06-05黃仁勳訪韓);SK份額~2/3(估);多年協議06-08	✓ share is a reported estimate份額屬報導估計	Yahoo 06-05/TrendForce 01-28/thenextweb 06-08
66. HBM4 12-layer >$600/unit (reported)HBM4 12層>$600/顆(報導)	⚠ single-source, unverified against a second source轉述、未經雙源	TrendForce 01-28
67. NVIDIA = 14.8% of 1Q26 revenue (₩7.78tn); second >10% customer at 12.4% (identity undisclosed)NVIDIA=1Q26營收14.8%(₩7.78兆);第二家>10%客戶12.4%(身分未揭露)	✓ ratio比例 ⚠ NVIDIA attribution & 2nd customer identityNVIDIA認定與第二家身分	Korean quarterly report via BigGo韓國季報via BigGo 05-15
68. Microsoft 3-year DDR5 LTA微軟3年DDR5長約	⚠ single-source, reported單源轉述	TrendForce 06-23
69. 1Q26 eSSD market $18.46B (+86% QoQ), contract +~80%; SK group $4.64B1Q26 eSSD市場$18.46B(+86%QoQ),合約價+~80%;SK集團$4.64B	✓ TrendForce	06-11
70. Almost zero new NAND capacity in 20262026 NAND幾乎零新增產能	✓ TrendForce	05-25
71. CXMT IPO raised ₩14tn, triggered a 07-16 selloff in both DRAM leadersCXMT IPO募₩14兆,07-16引發雙雄大跌	✓ ebn	2026-07-16
72. VEU expiry, 2026 annual license regimeVEU到期,2026年審制許可	✓ CNBC	2025-12-30
73. Section 232 semiconductor tariffs 25% (effective 2026-01-15; scope for Korean memory unclear)Section232半導體關稅25%(2026-01-15生效;對韓記憶體適用範圍不明)	⚠ conflicting sources來源衝突	CHR/sedaily 2026-01
74. Plant timelines: M15X tape-out 1Q26; Yongin fab-1 cleanroom 1Q27 (₩9.4tn); P&T7 late-2027 (₩19tn); Indiana 2H28 (₩5.9tn); Cheongju NAND plant 2029 (part of a broader ₩170tn Chungcheong plan, single-source)廠房時程:M15X 1Q26投片;Yongin一廠1Q27無塵室(₩9.4兆);P&T7末-2027(₩19兆);Indiana 2H28(₩5.9兆);清州NAND 2029(₩170兆忠清計畫一環,單源)	✓ 424B4 ⚠ NAND plant single-sourceNAND廠單源	424B4+sedaily 07-02
75. HBM4 timeline: development complete 2025-09, shipped 4Q25; HBM4E 12-layer samples 06-18; MR-MUF vs TC-NCFHBM4時程:2025-09開發完成,4Q25出貨;HBM4E 12層06-18出樣;MR-MUF vs TC-NCF	✓ official releases官方releases+EETimes+StorageReview	—
82. CTI research: IEDM 2025 paper, author Sangwan Jin (Materials Development Technical Leader); "Inside the Research" Ep.1CTI研究:IEDM 2025論文,作者Sangwan Jin(Materials Development Technical Leader);「Inside the Research」Ep.1	✓ direct fetch直接抓取全文	news.skhynix.com/en/research-inside-ep1/ 2026-07-22
83. CTI results: interference -30%+, retention +45%+, cell shrink >10%, industry-first at 176-layer production grade (all company-stated)CTI成果:干擾-30%+、保持力+45%+、cell縮>10%、176層生產級首例(皆公司自述)	✓ 官方自述官方自述	same as above同上
84. 2026 IEEE Corporate Innovation Award (HBM), 2026-04-252026 IEEE Corporate Innovation Award(HBM),2026-04-25	✓ official newsroom (related-content list)官方newsroom(related content列表)	—
Consensus & market共識與市場
76. Consensus 2026E rev ₩358.33tn/EPS ₩318,226 (36 analysts); 2026E OP 265.2tn, 2027E OP 386.1tn, 2028E OP 383.1tn共識2026E rev₩358.33兆/EPS₩318,226(36位);2026E OP265.2兆,2027E OP386.1兆,2028E OP383.1兆	✓ stockanalysis(S&P Global)+FnGuide via fnnews	07-22/07-10
77. 2Q26E consensus rev 83.93tn/OP 64.16tn (OPM 76.5%)2Q26E共識rev83.93兆/OP64.16兆(OPM76.5%)	✓ FnGuide via ggilbo	07-20
78. 11 straight quarterly beats (if 2Q beats again)連續11季beat(若2Q再beat)	✓ framing表述	en.sedaily 07-02
79. PBR 6-year average 1.84×; sell-side has dropped PBR; iM 7.5× PBR TP 3.5M, Mirae Asset 4.2M, Kiwoom downgradePBR6年均值1.84×;券商棄用PBR;iM給7.5×PBR TP3.5M、未來資產4.2M、Kiwoom降評	✓ TPs are opinionTP屬意見	sedaily 01-19/MTN 02-03/newspim 06月
80. Foreign ownership 49.88% (07-15), 53.83% at year start; 1H26 foreign net sold ~₩150tn of KOSPI外資持股49.88%(07-15),年初53.83%;1H26外資淨賣KOSPI約₩150兆	✓ ajunews/newsis	07-16/07-05
81. SK+Samsung = 54.76% of KOSPI market cap (06-19); SK's own market cap first led KOSPI 06-22 (₩2,079.7tn); KOSPI broke 9,000 (06-25); USDKRW peaked 1,561.5 (06-08)SK+三星=KOSPI市值54.76%(06-19);06-22 SK市值₩2,079.7兆首度登頂KOSPI;KOSPI破9,000(06-25);USDKRW 06-08高見1,561.5	✓ g-enews/hankyung/betanews	06月
Judgment (not among the enumerated hard-number claims)判斷(不計入列舉硬數字)
TP formula (2027E×6×)目標價公式(2027E×6×)	◆	₩1,960,000
Scenario probabilities情境機率	◆	35/45/20 → weighted加權 ₩1,881,500 (weights last re-set 07-29 and unchanged; scenario EPS restated 08-21 onto the post-cancellation share count權重 07-29 重估後未動；情境 EPS 於 08-21 改用註銷後股數重述)
2026E/27E estimates, 2027E OP -28.8% vs pre-print consensus2026E/27E估計,2027E OP較財報前共識-28.8%	◆	revised 07-29, see T307-29修正,見T3
Playbook / entry zone / thresholds (decision layer)操作框架/進場區/門檻(decision layer)	◆	all zones/thresholds = this analysis's judgment · re-derived 07-29 (entry zone triggered, staged entry, tails 35/45/20)所有區間/門檻皆為本研究判斷 · 07-29 重推（進場區觸發、分批、權重35/45/20）
Corrected (model memory overridden by sources)已修正(模型記憶被來源修正)
"2Q26 earnings date = 7/23"「2Q26財報日=7/23」	✗→✓	corrected to 7/29 per the 6-K修正為 7/29(依6-K)
"NVIDIA ≈50%+ of revenue" (intuition)「NVIDIA佔營收~50%+」(直覺)	✗→✓	actually 14.8% — diluted by commodity DRAM revenue實為14.8% — 商品DRAM營收稀釋
"07-29 close = ₩1,377,000" (published 07-29)「07-29 收盤＝₩1,377,000」(07-29 發布)	✗→✓	corrected to ₩1,401,000 — the source labelled an intraday quote as the close; every 07-29 derived figure was recomputed (see #99)更正為 ₩1,401,000 — 來源把盤中報價標成收盤；所有 07-29 衍生數字均已重算（見 #99）
"No 2026 open-market insider trades found" (07-22)「2026 年未發現內部人公開市場交易」(07-22)	✗→✓	resolved 08-04: chairman Chey Tae-won bought ₩4.90bn on 07-30 (see #105)08-04 解除：崔泰源會長 07-30 買進 ₩49.0億（見 #105）
Dropped (couldn't verify → excluded)已剔除(無法查證→排除)
65% of revenue from the US65% 美國營收占比	⚠ dropped剔除	single-source, uncheckable單源、不可核
HBM5/DDR6 forward roadmap遠期roadmap	⚠ dropped剔除	single enthusiast-media source愛好者媒體單源
119. ₩40tn buyback + full cancellation, board-resolved 08-19, executing 08-20→11-19; ~24.07M shares at the ₩1,662,000 reference (08-18 close) = 3.30% of shares outstanding; largest share retirement by a Korean listed company₩40兆庫藏股取得並全數註銷，08-19 董事會決議，執行期 08-20→11-19；以參考價 ₩1,662,000（08-18 收盤）計約 24.07M 股＝流通股數 3.30%；韓國上市公司史上最大規模股份註銷	✓ disclosure + five outlets agree; arithmetic self-consistent (40tn ÷ 1,662,000 = 24.07M; 24.07M ÷ 728.87M = 3.30%)公告＋五家media一致；數字自洽(40兆÷1,662,000=24.07M;24.07M÷728.87M=3.30%)	khan/ajunews/mt/opinionnews/etnews · 08-19–20
120. Shareholder-return policy raised: 2025-27 cumulative FCF payout from "within 50%" to "50% or more"; company cites intrinsic value "not sufficiently reflected in the current share price"股東回饋政策上調：2025-27 累計 FCF 回饋自「50% 以內」改為「50% 以上」；公司表示內在價值「未充分反映於目前股價」	✓ board disclosure via Korean press董事會公告，韓媒	khan/mt · 08-19
121. Close 2026-08-20 ₩1,691,000 (+12.73%); prior close 08-19 ₩1,500,000收盤 2026-08-20 ₩1,691,000(+12.73%)；前一交易日 08-19 收盤 ₩1,500,000	✓ five outlets; cross-checked two ways: 1,500,000 × 1.1273 = 1,690,950 ≈ 1,691,000, and topstarnews' 08-20 intraday 1,640,000 = +9.33% off the same 1,500,000 base五家media；兩路交叉驗算:1,500,000×1.1273=1,690,950≈1,691,000，且 topstarnews 08-20 盤中 1,640,000 = 同一 1,500,000 基期 +9.33%	sedaily/ebn/hankyung/etnews/asiae · 08-20
122. 08-21 session still open at publication: intraday ₩1,758,000 (+3.96%) at 10:25 KST — not used as the masthead anchor發布時 08-21 盤中仍在交易：₩1,758,000(+3.96%)於 10:25 KST — 未用作 masthead 錨點	✓ stockanalysis, explicitly labelled intraday; per the 08-04 lesson the masthead anchors on the last completed sessionstockanalysis，明確標示盤中；依 08-04 教訓，masthead 一律錨定最後一個已完成交易日	08-21 10:25 KST
123. KOSPI 08-20 closed 6,852.58 (+5.89%); Samsung Electronics +9.49% to ₩271,000; foreign net buying ₩1.7–2.2tnKOSPI 08-20 收 6,852.58(+5.89%)；三星電子 +9.49% 至 ₩271,000；外資淨買超 ₩1.7–2.2兆	✓ four outlets agree on index and Samsung; foreign-flow figure varies by outlet, range given四家media於指數與三星一致；外資金額各家略異，故以區間表示	sedaily/hankyung/asiae/ebn · 08-20
124. Market cap ₩1,232.5tn at the 08-20 close (≈US$884B)市值 ₩1,232.5兆（08-20 收盤，≈US$884B）	✓ computed計算	1,691,000 × 728,865,500
125. USDKRW 1,394.47 (08-20); page annotates ≈US$ at a rounded ₩1,395/$; 08-21 spot 1,381.59USDKRW 1,394.47(08-20)；頁上 ≈US$ 以 ₩1,395/$ 標註；08-21 即期 1,381.59	✓ TradingEconomics	08-20 / 08-21
126. SKHY 08-20 close $163.08 (+4.43%); prior close $156.16; after-hours $168.46SKHY 08-20 收盤 $163.08(+4.43%)；前收 $156.16；盤後 $168.46	✓ stockanalysis close + IBKR realtime for the after-hours printstockanalysis 收盤＋IBKR 即時（盤後價）	08-20 16:00 EDT / 21:42 EDT
127. ADS premium +34.5% on the 08-20 closes — through the 30% ceiling; history 4.9% (07-10) → 21.6% (07-28) → ≈31% (07-29) → 30.0% (08-03) → 34.5% (08-20)ADS 溢價 08-20 兩地收盤 +34.5% — 已穿破 30% 上限；沿革 4.9%(07-10)→21.6%(07-28)→≈31%(07-29)→30.0%(08-03)→34.5%(08-20)	✓ computed (#121/#125/#126): 1,691,000 ÷ 10 ÷ 1,394.47 = $121.26; 163.08 ÷ 121.26 − 1 = 34.5%計算(#121/#125/#126):1,691,000÷10÷1,394.47=$121.26;163.08÷121.26−1=34.5%	—
128. 2027E EPS revised ₩300,000 → ₩309,000 and TP ₩1,800,000 → ₩1,850,000 — share count only (729.0M → 705.2M), net of ~₩0.95tn forgone after-tax interest2027E EPS 自 ₩300,000 修正為 ₩309,000、目標價自 ₩1,800,000 修正為 ₩1,850,000 — 只動股數(729.0M→705.2M)，並扣除約 ₩0.95兆放棄的稅後利息	◆ this analysis; arithmetic shown in TP derivation本研究；算式見目標價推導	—
129. Realized 3Q26 contract prices still unsettled as of 08-21; latest published figures remain the +13–18% conventional DRAM / +10–15% NAND forecast band. The 15–20% PC DRAM figure circulating in August dates from a 06-30 revision and predates this page's 07-03 basis — it is not new information3Q26 實際合約價截至 08-21 仍未定案；目前已公布的最新數字仍是 +13–18% conventional DRAM／+10–15% NAND 預測區間。8 月間流傳的 PC DRAM 15–20% 數字來自 06-30 的上修，早於本頁 07-03 的基準 — 不是新資訊	✓ TrendForce 07-09 + pcmasterinsider 08-03, both explicitly forecasts; kucoin flash dated 06-30TrendForce 07-09＋pcmasterinsider 08-03，兩者皆明示為預測；kucoin 快訊日期 06-30	trendforce/pcmasterinsider · 07-09, 08-03
130. CPO roadmap published in Nature Electronics 08-20 with Univ. of Virginia, UIUC, NTU, MIT and Yonsei; targets >100 Tb/s per node, <1 pJ/bit, <10 ns chip-to-chip; compute rises ~3×/2yr vs interconnect ~1.4×; path 2D → 2.5D interposer → 3D heterogeneous integration. No product, customer or dateCPO 藍圖 08-20 發表於 Nature Electronics，共同作者含維吉尼亞大學、UIUC、南洋理工、MIT、延世；目標為每節點 >100 Tb/s、<1 pJ/bit、晶片間延遲 <10 ns；算力約每 2 年 ×3、互連約 ×1.4；路徑為 2D→2.5D 中介層→3D 異質整合。無產品、無客戶、無時程	✓ company newsroom + The Elec + Seoul Economic Daily agree on every figure公司 newsroom＋The Elec＋首爾經濟，各項數字一致	news.skhynix.com/en/cpo-in-nature-electronics/ · 08-20

Counts (updated 08-04): 118 hard-number claims reviewed → 102 ✓ verified · 4 corrected · 12 ⚠ uncertain (single-source or flagged-partial: HBM4 unit price #66, NVIDIA/2nd-customer identity #67, Microsoft LTA #68, Section 232 scope #73, NAND-plant timeline #74, oversubscription claim #49, 1Q26 NI composition #31, LTA half-of-sales ratio #90, post-print TP average #106, HBF/HBM4 compatibility #109, competitor HBF plans #116, Kioxia prototype #117). The no-insider-trades finding moved out of ⚠ and into corrected. Plus separate ◆ judgment calls outside the enumerated set (TP formula, scenario probabilities, forward estimates, decision layer).統計（08-04 更新）：118 項硬數字查證 → 102✓已查證 · 4 項已修正 · 12⚠未核實(單源或部分標註：HBM4單價#66、NVIDIA/第二家客戶身分#67、微軟長約#68、Section232範圍#73、NAND廠時程#74、超額認購說法#49、1Q26淨利組成#31、LTA占營收比#90、財報後平均目標價#106、HBF與HBM4相容性#109、競爭者HBF計畫#116、Kioxia原型#117)。原「無內部人交易」一項已自 ⚠ 移入已修正。另有列舉外的◆判斷（目標價公式、情境機率、財測估計、操作框架決策層）。

Update log更新紀錄

✓ 2026-08-21 ₩40tn buyback and cancellation — the largest in Korean market history · TP ₩1,800,000 → ₩1,850,000 on share count alone · view stays range-bound · CPO roadmap published ₩40兆庫藏股取得並註銷 — 韓國市場史上最大・目標價 ₩1,800,000 → ₩1,850,000，只因股數・觀點維持區間・CPO 藍圖發表

In one paragraph. On 08-19 the board approved a ₩40tn (≈$28.7B) buyback with full cancellation, running 08-20 to 11-19 — ~3.3% of shares outstanding, the largest share retirement ever executed by a Korean listed company — and raised the 2025-27 shareholder-return floor from "within 50%" of cumulative free cash flow to "50% or more". The stock rose +12.73% to ₩1,691,000 on 08-20, carrying the KOSPI up 5.89%. This page takes the cancellation into the model — 2027E EPS ₩300,000 → ₩309,000, target ₩1,800,000 → ₩1,850,000 — and no further. Not one operating estimate moved: revenue, operating profit and margin for 2026E and 2027E are all carried forward, and the 6× multiple is unchanged. A +2.8% target against a +7.9% price leaves +9.4% of implied upside, so the view stays range-bound and adding stays paused for the third update running. Separately, a co-packaged optics roadmap was published in Nature Electronics on 08-20 — optionality, with no product and no date, treated exactly as HBF was. ◆ framing = judgment定位＝判斷

一段話講完。08-19 董事會通過 ₩40兆（≈$28.7B）庫藏股取得並全數註銷，執行期 08-20 至 11-19 — 約占流通股數 3.3%，是韓國上市公司史上最大規模的股份註銷 — 同時把 2025-27 年股東回饋下限，自累計自由現金流的「50% 以內」上調為「50% 以上」。股價 08-20 大漲 +12.73% 至 ₩1,691,000，並帶動 KOSPI 上漲 5.89%。本頁把註銷放進模型 — 2027E EPS ₩300,000 → ₩309,000、目標價 ₩1,800,000 → ₩1,850,000 — 就到此為止。沒有任何一項營運估計被動過：2026E 與 2027E 的營收、營業利益、營益率全數延用，6 倍 PER 不變。目標價 +2.8% 對上股價 +7.9%，剩下的隱含空間是 +9.4%，因此觀點維持「區間」，加碼連續第三次更新維持暫停。另外，08-20 有一份 CPO 藍圖發表於 Nature Electronics — 選擇權性質，無產品、無時程，比照 HBF 處理。◆ 定位＝判斷

The full entry — the buyback weighed, the TP arithmetic, the ADR premium breach, catalyst scoring, and everything refreshed (i)–(viii)完整條目 — 庫藏股的秤重、目標價算式、ADR 溢價穿頂、催化劑計分，與全部更新項目 (i)–(viii)

(i) The news. Board resolution 08-19: acquire ₩40tn of treasury stock between 08-20 and 11-19, then cancel all of it within one to two weeks of completion. At the ₩1,662,000 reference price (the 08-18 close) that is ~24.07M shares, 3.30% of the 728,865,500 outstanding. Samsung Securities called it the largest domestic share cancellation on record and the second after the ₩12.2tn buyback earlier in 2026. The board simultaneously lifted the 2025-27 cumulative-FCF return floor from "within 50%" to "50% or more", and stated that the company's intrinsic value was "not sufficiently reflected in the current share price" ✓ disclosure + five outlets公告＋五家media.

(i) 消息本身。08-19 董事會決議：08-20 至 11-19 之間取得 ₩40兆庫藏股，執行完畢後一至兩週內全數註銷。以參考價 ₩1,662,000（08-18 收盤）計算約 24.07M 股，占 728,865,500 股的 3.30%。三星證券稱其為韓國史上最大規模的股份註銷，也是繼 2026 年初 ₩12.2兆之後的第二次。董事會同時把 2025-27 累計 FCF 回饋下限自「50% 以內」上調為「50% 以上」，並表示公司內在價值「未充分反映於目前股價」✓ 公告＋五家media。

(ii) The tape. 08-20 close ₩1,691,000, +12.73% from ₩1,500,000 — cross-checked two ways before use (1,500,000 × 1.1273 = 1,690,950, and a separate outlet's 08-20 intraday ₩1,640,000 = +9.33% off the same base). KOSPI closed 6,852.58 (+5.89%), Samsung Electronics +9.49%, foreign net buying ₩1.7–2.2tn. At publication the 08-21 session was still open at an intraday ₩1,758,000 (+3.96%, 10:25 KST); per the 08-04 lesson the masthead anchors on the last completed session and the intraday print is stated separately ✓ five outlets五家media.

(ii) 盤面。08-20 收 ₩1,691,000，自 ₩1,500,000 大漲 +12.73% — 使用前已兩路交叉驗算（1,500,000×1.1273=1,690,950；另一家media 08-20 盤中 1,640,000＝同基期 +9.33%）。KOSPI 收 6,852.58（+5.89%）、三星電子 +9.49%、外資淨買超 ₩1.7–2.2兆。發布時 08-21 盤中仍在交易，₩1,758,000（+3.96%，10:25 KST）；依 08-04 的教訓，masthead 錨定最後一個已完成交易日，盤中價另行標示 ✓ 五家media。

(iii) What changed in the model, and what did not. Only the denominator. ₩40tn retires ~23.7M shares at the prevailing ₩1,691,000 (fewer than the ~24.07M the ₩1,662,000 reference implied), taking the 2027E divisor from 729.0M to 705.2M. Net of ~₩0.95tn of after-tax interest income forgone by spending the cash, 2027E EPS goes ₩300,000 → ₩309,000 (+3.0%) and the target ₩1,800,000 → ₩1,850,000 (+2.8%) at the same 6× multiple. Revenue, operating profit and operating margin for both 2026E and 2027E are untouched, which is why the T3 revision table now shows a column of 0.0% beside a share-count line. All three scenario EPS figures were restated onto the post-cancellation share count (bear ₩199k→₩204k, bull ₩420k→₩433k), lifting every scenario TP by the same ~3%; the probability weights 35/45/20 were not touched, because a buyback moves the share count, not the memory cycle ◆.

(iii) 模型動了什麼、沒動什麼。只動分母。₩40兆以現價 ₩1,691,000 買回約 23.7M 股（少於參考價 ₩1,662,000 所隱含的約 24.07M 股），2027E 除數自 729.0M 降至 705.2M。扣除花掉現金所放棄的約 ₩0.95兆稅後利息收入後，2027E EPS 自 ₩300,000 升至 ₩309,000（+3.0%），目標價在同樣 6 倍 PER 之下自 ₩1,800,000 升至 ₩1,850,000（+2.8%）。2026E 與 2027E 的營收、營業利益、營益率都沒有動 — 這正是 T3 修正表現在會出現「一整排 0.0% 搭配一列股數」的原因。三個情境的 EPS 都改用註銷後股數重述（bear ₩199k→₩204k、bull ₩420k→₩433k），使每個情境 TP 同步抬高約 3%；35/45/20 的機率權重未動，因為庫藏股改變的是股數，不是記憶體循環 ◆。

(iv) Why the view does not improve. The target rose 2.8%; the price rose 7.9% since the 08-03 anchor. The Buy line, defined as TP ÷ 1.15, re-derived itself from ₩1,565,200 to ₩1,608,700 — the first time the mechanism installed on 08-04 has actually fired — but the close sits 5.1% above it, against 0.11% above at the last update. So the gap to Buy widened rather than closed, and the view stays range-bound. The asymmetry check is now the worst this page has recorded: doubling requires 98% of the entire bull scenario (93% on 08-03, 82% at the 07-29 low), and halving corresponds to 2027 OP about 16% below the bear case (21% on 08-03, 30% at the low) ◆.

(iv) 為什麼觀點沒有變好。目標價漲 2.8%；股價自 08-03 錨點漲 7.9%。定義為 TP ÷ 1.15 的買進線自 ₩1,565,200 自動重推至 ₩1,608,700 — 這是 08-04 裝上去的那套機制第一次真的啟動 — 但收盤高出它 5.1%，上次更新則只高出 0.11%。距離「買進」是拉開、不是收攏，因此觀點維持「區間」。不對稱性檢驗是本頁記錄以來最差：翻倍需要吃掉整個 bull 情境的 98%（08-03 為 93%、07-29 低點為 82%），跌一半則對應 2027 OP 僅比 bear 情境低約 16%（08-03 為 21%、低點為 30%）◆。

(v) The buyback, weighed rather than cheered. Three things are true at once. It is mechanically accretive, and this page has taken that into the target. The announcement raised the company's own execution cost — roughly a tenth of the intended retirement was consumed by the market's reaction to announcing it. And it retires stock at ≈4.7× book with peak-cycle cash, into the 2027-29 supply wave this page's bear case describes. Set beside July's ≈₩37tn equity raise, the share count over Jul–Dec is close to a round trip: the company raised equity for capacity six weeks before committing ₩40tn to destroying equity. The two pools are not the same money — ADS proceeds are earmarked for fabs, the buyback comes from operating cash — but the sequencing is a statement worth reading ◆.

(v) 庫藏股，用秤的、不用喊的。三件事同時成立。它機械上是增益的，本頁也已算進目標價。這則公告抬高了公司自己的執行成本 — 原本要註銷的量，大約十分之一被「宣布這件事」的市場反應吃掉。而它是在約 4.7 倍股價淨值比、用循環高點的現金註銷股票，迎面而來的正是本頁空方情境描述的 2027-29 供給波。與 7 月約 ₩37兆的增資並排看，7 到 12 月的股數幾乎是一趟來回：公司在承諾用 ₩40兆消滅股權的六週前，才剛為了產能增資。兩筆錢不是同一筆 — ADS 募資款專用於廠房，庫藏股來自營運現金 — 但這個順序本身是有態度的，值得讀 ◆。

(vi) Catalyst scoring. A new row was added for the buyback with its own pre-committed bar: full ₩40tn executed and cancelled by ~December confirms the accretion; executed short or cancellation deferred reverses the 08-21 TP revision back toward ₩1,800,000, since the whole upgrade is the share count. The realized TrendForce 3Q26 contract print — due "late August" — has not settled as of 08-21; the latest published figures are still the +13–18% conventional DRAM / +10–15% NAND forecast band, so nothing arrived to score. Note for the record: the 15–20% PC DRAM figure circulating in August dates from a 06-30 revision, predates this page's 07-03 basis, and is therefore not new information. That row stays open with its window now live. 3Q26 earnings on ~10-28 remains the key operating data point.

(vi) 催化劑計分。新增一列庫藏股，並附上自己的預先承諾門檻：₩40兆全額執行並於約 12 月完成註銷，則增益確立；若執行未達額或註銷遞延，則把 08-21 的目標價修正倒回 ₩1,800,000 方向，因為這次上調的全部就是股數。原訂「8 月下旬」的 TrendForce 3Q26 實際合約價，截至 08-21 仍未定案；目前已公布的最新數字仍是 +13–18% conventional DRAM／+10–15% NAND 的預測區間，沒有東西可以計分。附帶備查：8 月間流傳的 PC DRAM 15–20% 來自 06-30 的上修，早於本頁 07-03 的基準，因此不是新資訊。該列維持開啟，視窗現已生效。約 10-28 的 3Q26 財報仍是最關鍵的營運數據點。

(vii) The ADR premium broke its ceiling. SKHY closed 08-20 at $163.08 against ≈$121.3 of intrinsic value per ADS, a +34.5% premium — through the 30% hard ceiling rather than at it, and the widest reading since listing (4.9% → 21.6% → ≈31% → 30.0% → 34.5%). It widened even though the ADR rose only +4.43% against Seoul's +12.73%, because it had already run ahead. The consequence is stated plainly in the playbook: no ADR buying at any size while the premium is above 30%, regardless of the KRX price — a passing KRX leg does not rescue a failed premium leg. Buying at $163.08 and reaching the ₩1,850,000 target returns +1.7% / -2.4% / -18.6% as the premium ends at 25% / 20% / zero, against +10.4% / +6.0% / -11.6% one update ago. Nothing here instructs an existing holder to sell.

(vii) ADR 溢價穿破上限。SKHY 08-20 收 $163.08，對照每 ADS 內含價值 ≈$121.3，溢價 +34.5% — 是穿過 30% 硬上限，而不是踩在上面，也是掛牌以來最寬的一次（4.9%→21.6%→≈31%→30.0%→34.5%）。它是在 ADR 只漲 +4.43%、首爾漲 +12.73% 的情況下擴大的，因為它本來就跑在前面。後果在操作框架裡寫得很直白：只要溢價高於 30%，ADR 一股都不買，不管 KRX 價格怎麼走 — KRX 那條腿過關，救不了溢價這條腿。以 $163.08 買進、到達 ₩1,850,000 目標價，溢價收在 25%／20%／歸零 → 報酬 +1.7%／-2.4%／-18.6%，而上次更新同樣三個數字是 +10.4%／+6.0%／-11.6%。以上沒有一條指示既有持有人賣出。

(viii) Everything refreshed. Close ₩1,567,000 → ₩1,691,000 · market cap ₩1,142兆 → ₩1,232兆 (≈$799B → ≈$884B) · FX ₩1,427.53 → ₩1,394.47/$ (page-rounded ₩1,430 → ₩1,395, every ≈US$ annotation re-rendered) · fwd PER 2027E 5.2× → 5.5× · PBR 4.3× → 4.7× · off-peak -47.5% → -43.4% · YTD +141% → +160% · SKHY $142.72 → $163.08 · premium 30.0% → 34.5% · entry zone <₩1,565,000 → <₩1,608,700 and the whole ADS level table with it · LTA customer count ~10 → 11 (2Q26) · scenario diagram redrawn with the Buy line now shown as its own marker · a new CPO section added with a bandwidth-wall diagram · sources ledger 118 → 130 claims.

(viii) 全部更新項目。收盤 ₩1,567,000 → ₩1,691,000・市值 ₩1,142兆 → ₩1,232兆（≈$799B → ≈$884B）・匯率 ₩1,427.53 → ₩1,394.47/$（頁面取整 ₩1,430 → ₩1,395，所有 ≈US$ 標註重算）・fwd PER 2027E 5.2× → 5.5×・PBR 4.3× → 4.7×・距峰 -47.5% → -43.4%・YTD +141% → +160%・SKHY $142.72 → $163.08・溢價 30.0% → 34.5%・進場區 <₩1,565,000 → <₩1,608,700，ADS 價位表整張跟著重算・長約客戶數 ~10 → 11（2Q26）・情境圖重繪並把買進線獨立標出・新增 CPO 章節與頻寬牆示意圖・資料來源表 118 → 130 項。

✓ 2026-08-04 HBF standard published with Sandisk · a +11.8% three-session rebound · view Buy → range-bound with the target unchanged · one correction to the 07-29 numbers 與 Sandisk 發布 HBF 標準・三個交易日反彈 +11.8%・觀點買進→區間而目標價不變・07-29 數字一則更正

In one paragraph. SK hynix and Sandisk published the first HBF standard — real, but with no customer and no volume before ~2030, and with Samsung and Kioxia contesting the same format, so it changes nothing in the 2027 model. What changed is the price: +11.8% in three sessions, including the first limit-up since 2009, which left the stock ₩1,783 above ₩1,565,000 — the line where this page's Buy band starts. So the view steps back to range-bound with the ₩1,800,000 target and every estimate unchanged: a downgrade of the entry price, not of the company. Separately, the 07-29 close is corrected to ₩1,401,000, which does not change that day's conclusion.

一段話講完。SK 海力士與 Sandisk 發布了第一份 HBF 標準 — 是真的，但沒有客戶、2030 年之前沒有量，而且三星與 Kioxia 正在爭同一個格式，所以它對 2027 模型沒有任何影響。真正改變的是價格：三個交易日 +11.8%，其中一根是 2009 年以來首次漲停，把股價推到 ₩1,565,000 之上 ₩1,783 — 而 ₩1,565,000 正是本頁買進區的起點。因此觀點退回「區間」，₩1,800,000 目標價與所有估計值都不變：下修的是進場價格，不是這家公司。另外，07-29 收盤價更正為 ₩1,401,000，不改變那天的結論。

The full entry — the news, the tape, catalyst scoring, the correction, and everything refreshed (i)–(vii)完整條目 — 消息、盤面、催化劑計分、更正，以及全頁更新項目 (i)–(vii)

(i) The news. On 08-03 SK hynix and Sandisk(SNDK.O) published the first High Bandwidth Flash technical specification through the Open Compute Project: up to 512GB per stack in 8-high and 16-high configurations, three grades spanning ≈0.4–3.0TB/s, over a UCIe chiplet interconnect, in a package Sandisk describes as matching HBM4's pin layout, dimensions and power profile ✓ both partners' releases雙方 release · 08-03. The consortium (with Google and Tenstorrent) formed in 2026-02. Separately, the company showed its 375-layer 4D NAND for the first time at FMS 2026 and guided eSSDs built on it to mass production in early 2027 ✓. Full treatment, with a diagram of where HBF sits between HBM and the SSD, is in the new HBF section.

(i) 消息本身。08-03，SK 海力士與 Sandisk(SNDK.O) 透過 Open Compute Project 發布第一份 High Bandwidth Flash（HBF）技術規格：8 層與 16 層堆疊、單疊最高 512GB，三個等級橫跨約 0.4–3.0TB/s，走 UCIe chiplet 互連；Sandisk 並稱其封裝的腳位、尺寸、功耗規格與 HBM4 相同 ✓ 雙方 release · 08-03。這個聯盟（另有 Google 與 Tenstorrent）於 2026-02 成立。另外，公司在 FMS 2026 首度展出 375 層 4D NAND，並指引以此打造的 eSSD 於 2027 年初量產 ✓。完整處理與「HBF 落在 HBM 與 SSD 之間哪一層」的圖解，見新增的 HBF 章節。

(ii) How it bears on the thesis: it does not move the model, and this page will not pretend otherwise. A specification is not a product. A technical roadmap exists and predates this spec (Sandisk: first samples 2H CY2026, AI-inference device samples early 2027), but the 08-03 material carries no named customer, no price and no volume commitment, and commercialisation is guided to ~2030 ✓. Nor does SK own the format: Sandisk has MOUs with Samsung as well, Samsung began HBF development in 2025-10, and Kioxia demonstrated a 5TB prototype in 2025-08 ✓ reported轉述. It is a genuine strengthening of the NAND leg — NAND is 22.0% of revenue and the structurally weaker half of this business, and HBF is an attempt to sell NAND bits into a bandwidth-priced socket rather than a capacity-priced one — but that belongs in the optionality column and contributes exactly zero to 2027E ◆. Optionality is not a catalyst.

(ii) 它對論點的意義：它沒有動到模型，本頁也不會假裝有。規格不等於產品。技術時程是存在的、而且早於這份規格（Sandisk：2026 下半年首批樣品、2027 年初 AI 推論裝置樣品），但 08-03 的材料沒有具名客戶、沒有價格、沒有量的承諾，商業化指引在約 2030 年 ✓。SK 也並不擁有這個格式：Sandisk 跟三星同樣簽了 MOU、三星 2025-10 已展開 HBF 開發、Kioxia 2025-08 就展示過 5TB 原型 ✓ 轉述。它確實補強了 NAND 這條腿 — NAND 占營收 22.0%，是這門生意結構上較弱的一半，而 HBF 是把 NAND 的 bit 賣進「按頻寬計價」而非「按容量計價」位置的一次嘗試 — 但這該放進「選擇權」欄位，對 2027E 的貢獻正好是零 ◆。選擇權不是催化劑。

(iii) The tape did something far larger. From the 07-29 close of ₩1,401,000: 07-30 ₩1,322,000 (-5.64%), then 07-31 limit-up +29.95% to ₩1,718,000 — SK hynix's first daily limit since 2009 and its first under the 30% band adopted in 2015, on Microsoft/Amazon results clearing the AI-capex fear (SOX +8.19%) and a ₩4.90bn open-market share purchase by chairman Chey Tae-won the previous session — then 08-03 -8.80% to ₩1,567,000 on profit-taking ✓ Korean press ×3韓媒×3. Net: +11.8% in three sessions, a ±30% round trip, and not one new piece of fundamental data inside it.

(iii) 盤面做的事大得多。自 07-29 收盤 ₩1,401,000 起：07-30 ₩1,322,000（-5.64%），接著 07-31 上限漲停 +29.95% 至 ₩1,718,000 — SK 海力士 2009 年以來首次漲停，也是 2015 年放寬至 30% 後的第一次，動因是微軟／亞馬遜財報消除了 AI capex 疑慮（費半 +8.19%），以及前一交易日崔泰源會長 ₩49.0億的場內買股 — 然後 08-03 因獲利了結 -8.80% 至 ₩1,567,000 ✓ 韓媒×3。淨結果：三個交易日 +11.8%、一趟 ±30% 的來回，而這中間沒有出現任何一筆新的基本面數據。

(iv) Scoring the pre-committed catalyst. This page committed on 07-29 to reading the TrendForce 3Q26 contract price against a bar of "conventional DRAM QoQ ≥ +high-teens%". The forecast leg turns out to have been on the record since 07-03: +13–18% QoQ for conventional DRAM, +10–15% for NAND ✓ TrendForce ×2. Only the top of that band reaches the bar, so it scores at-or-below, not as a confirmation. Reaction executed as pre-committed: the second tranche is withheld and adding is paused. The estimates were left alone — that same band implies 3Q group revenue of roughly ₩91–95tn, which brackets this model's ₩92tn, so it confirms the path rather than moving it. The realized print in late August still arbitrates.

(iv) 給預先承諾的催化劑打分。本頁 07-29 承諾以「conventional DRAM QoQ ≥ +高十位數%」這條門檻來讀 TrendForce 3Q26 合約價。結果它的預測腿早在 07-03 就已在案：conventional DRAM +13–18% QoQ、NAND +10–15% ✓ TrendForce ×2。只有區間上緣搆得到門檻，因此計分為「打平或不足」，不算確認。依承諾執行反應：第二批保留、加碼暫停。估計值不動 — 同一個區間隱含 3Q 集團營收約 ₩91–95兆，正好把本模型的 ₩92兆夾在中間，它確認了路徑而不是移動它。8 月下旬的實際數據仍是最終仲裁者。

(v) The conclusion changed — and it changed for one reason only: price. View moves Buy → range-bound ◆, target price unchanged at ₩1,800,000. The mechanism is arithmetic, not opinion: this page's view bands are defined on 6-month absolute return, so Buy begins at +15%, which is a price of ₩1,800,000 ÷ 1.15 = ₩1,565,217. The 08-03 close of ₩1,567,000 sits ₩1,783 — 0.11% — on the range-bound side of that line. Nothing about the company was downgraded; the entry price was. The entry zone is re-derived accordingly from the stale ad-hoc ₩1,600,000 (a level calibrated against the superseded ₩1,960,000 target) to <₩1,565,000, which now recomputes itself automatically whenever the target moves. Stated plainly, because it matters: this is a boundary call. The 08-04 session was trading at ₩1,559,000 as this update published — ₩6,000 below the ₩1,565,000 line, i.e. back inside Buy. A label that flips on a 0.2% price move is not a signal — the pre-committed levels are, and they did not move.

(v) 結論改變了 — 而且只因為一個理由：價格。觀點自「買進」轉為「區間」 ◆，目標價維持 ₩1,800,000 不變。機制是算術、不是意見：本頁的觀點區間定義在 6 個月絕對報酬上，買進自 +15% 起算，對應價格 ₩1,800,000 ÷ 1.15 ＝ ₩1,565,217。08-03 收盤 ₩1,567,000 落在這條線的「區間」那一側，只高出 ₩1,783、即 0.11%。被下修的不是這家公司，是進場價格。進場區也據此自過期的 ₩1,600,000（那是對著已被取代的 ₩1,960,000 目標價校準的）重推為 <₩1,565,000，而它今後會隨目標價自動重算。直說，因為這很重要：這是一個邊界判定。本次更新發布時，08-04 盤中價為 ₩1,559,000 — 比 ₩1,565,000 低了 ₩6,000，也就是已經回到買進區內。一個會因為 0.2% 價格變動而翻面的標籤不是訊號 — 預先承諾的價位才是，而那些價位沒有動。

(vi) A correction to the 07-29 entry. That update published the 07-29 close as ₩1,377,000. The official close was ₩1,401,000 (-9.61%); ₩1,377,000 was an intraday print ✗→✓ two Korean outlets, and the arithmetic checks two ways兩家韓媒，且雙向驗算相符. Root cause, recorded so it does not recur: the price source labels a live quote "closing price" while the Seoul session is still open — reproducible on 08-04, where the same source called the intraday ₩1,514,000 a close. Downstream figures for 07-29 are all restated: market cap ₩1,004tn→₩1,021tn, fwd PER 4.6×→4.7×, PBR 3.8×→3.9×, off-peak -53.9%→-53.1%, implied upside +30.7%→+28.5%, ADS premium ≈34%→≈31%. The 07-29 conclusion is unaffected — +28.5% still clears the +15% Buy bar by a wide margin, so the upgrade made that day stands on the corrected number too.

(vi) 對 07-29 條目的一則更正。該次更新把 07-29 收盤登載為 ₩1,377,000。官方收盤是 ₩1,401,000（-9.61%）；₩1,377,000 是盤中價 ✗→✓ 兩家韓媒，且雙向驗算相符。成因記錄如下以免重演：價格來源在首爾盤中會把即時報價標成「closing price」— 08-04 可複現，同一來源把盤中的 ₩1,514,000 稱為收盤。07-29 的所有衍生數字一併更正：市值 ₩1,004兆→₩1,021兆、fwd PER 4.6×→4.7×、PBR 3.8×→3.9×、距峰 -53.9%→-53.1%、隱含空間 +30.7%→+28.5%、ADS 溢價 ≈34%→≈31%。07-29 的結論不受影響 — +28.5% 仍大幅越過 +15% 的買進門檻，那天的上調用更正後的數字一樣成立。

(vii) Also refreshed throughout. Masthead (close, market cap ₩1,142tn, fwd PER 5.2×, PBR 4.3×, off-peak -47.5%, YTD +141%); FX to ₩1,427.53/$ from ₩1,450, so every ≈US$ annotation on the page was re-rendered; the ADR box (SKHY $142.72, premium ≈+30.0% — at the 30% hard ceiling, with the level table and return math recomputed); the sell-side picture, where the post-print reset has begun and split rather than converged (7 cuts, 2 raises, 9 held; ₩1.48M–4.70M, a 3.2× high-to-low ratio; roster average ≈₩3,296,000 vs +148%→+110% implied); the management section, where the standing "no insider trades found ⚠" is retired by chairman Chey's purchase; the scenario diagram and asymmetry check; a new HBF section with a mechanism diagram; and the sources ledger, now 118 claims. Hero image reviewed and kept — the thesis it illustrates is unchanged. The official SK hynix newsroom write-up of the FMS keynote was checked against every figure here: it corroborates the spec in full and, notably, does not claim HBM4 pin/dimension/power compatibility — that claim appears only on the Sandisk side, which is why it keeps its ⚠.

(vii) 全頁同步更新。頁首（收盤、市值 ₩1,142兆、fwd PER 5.2×、PBR 4.3×、距峰 -47.5%、YTD +141%）；匯率自 ₩1,450 改為 ₩1,427.53/$，因此全頁 ≈US$ 標註一律重算；ADR 換算框（SKHY $142.72、溢價 ≈+30.0% — 正好踩在 30% 硬上限，價位表與報酬數學一併重算）；賣方版圖，財報後重置已經開始，而且是分裂不是收斂（7 降、2 升、9 維持；₩1.48M–4.70M，最高是最低的 3.2 倍；名單平均約 ₩3,296,000，隱含空間自 +148% 降為 +110%）；經營層章節，長期掛著的「未發現內部人交易 ⚠」因崔泰源會長買股而解除；情境圖與不對稱檢驗；新增 HBF 章節與機制圖；資料來源表擴充至 118 項。Hero 圖已檢視、維持不變 — 它所描繪的論點沒有改變。SK 海力士官方 newsroom 對 FMS keynote 的報導已逐項比對：規格全部吻合，但值得注意的是它並未宣稱與 HBM4 腳位／尺寸／功耗相容 — 該說法只出現在 Sandisk 一方，這正是它保留 ⚠ 的原因。

What would change the view back. Price below ₩1,565,000 (mechanically), or — the version that would actually mean something — the realized late-August TrendForce 3Q contract print landing at ≥+high-teens% QoQ, which is what the second tranche has been waiting on since 07-29. ◆ = judgment/estimate throughout this entry unless marked otherwise.

什麼會讓觀點翻回去。價格跌破 ₩1,565,000（機械性地），或者 — 真正有意義的那個版本 — 8 月下旬 TrendForce 3Q 實際合約價落在 ≥+高十位數% QoQ，那正是第二批自 07-29 起一直在等的東西。◆ ＝本紀錄除另有標註外，皆屬本研究判斷/估計。

Earlier entries (08-04 HBF standard · 07-29 2Q26 earnings · 07-23 restructure · 07-22 initial coverage) — expand較早的紀錄（08-04 HBF 標準・07-29 2Q26 財報・07-23 改版・07-22 初次評估）— 點開展開

✓ 2026-07-292Q26 earnings: record quarter, first miss in 10 — TP ₩1,960,000→₩1,800,000, view upgraded 區間→買進2Q26 財報：創紀錄的一季、也是連 10 季 beat 後的首次 miss — TP ₩1,960,000→₩1,800,000，觀點區間→買進

✗→ 2026-08-04 Entry preserved as written; two figures in it were later found wrong. The 07-29 close is ₩1,401,000 (-9.61%), not the ₩1,377,000 used below, and the ADS premium was ≈31%, not ≈34% — the price source labelled an intraday quote as the close. Every derived figure in this entry (market cap ₩1,004tn, fwd PER 4.6×, PBR 3.8×, off-peak -53.9%, implied upside +30.7%, two-day -21.7%) is off by ~1.7% accordingly. The upgrade to Buy stands on the corrected number. See the 08-04 entry, item (vi), and claim #99.條目照原樣保留；其中兩個數字事後查出有誤。07-29 收盤是 ₩1,401,000（-9.61%），不是下文使用的 ₩1,377,000；ADS 溢價是 ≈31%、不是 ≈34% — 價格來源把盤中報價標成了收盤。本條目中所有衍生數字（市值 ₩1,004兆、fwd PER 4.6×、PBR 3.8×、距峰 -53.9%、隱含空間 +30.7%、兩日 -21.7%）因此都偏離約 1.7%。以更正後的數字計算，當日上調為「買進」的結論仍然成立。詳見 08-04 條目第 (vi) 點與查證表 #99。

(i) The print. Rev ₩79.32tn (≈$54.7B) / OP ₩60.54tn (≈$41.8B) (OPM 76.3%) — all-time highs, but -5.5%/-5.6% below consensus (₩83.93tn (≈$57.9B)/₩64.16tn (≈$44.3B)), ending a 10-quarter beat streak ✓ official + FnGuide · 07-29. Net profit ₩93.92tn (≈$64.8B) exceeds revenue (118% margin) because ₩63.27tn (≈$43.6B) of one-off investment gains — the Kioxia stake exit + SPC revaluation — landed this quarter ✓ sedaily. Net cash is now an official ₩69.4tn (≈$48B). Capex guidance raised to high-₩40tn; LTAs signed with ~10 customers; HBM4 mass shipments began. Blended DRAM ASP +~30% QoQ (vs 1Q's +60-70%): still rising, visibly decelerating ✓ call recaps ×2.

(i) 財報本身。營收 ₩79.32兆（≈$54.7B）／OP ₩60.54兆（≈$41.8B）（OPM 76.3%）— 雙雙創史上新高，但低於共識（₩83.93兆（≈$57.9B）／₩64.16兆（≈$44.3B））各 -5.5%／-5.6%，連 10 季 beat 就此中止 ✓ 官方+FnGuide · 07-29。淨利 ₩93.92兆（≈$64.8B）比營收還高（淨利率 118%），因為 ₩63.27兆（≈$43.6B）的一次性投資利益 — Kioxia 持股出場＋SPC 評價 — 在本季入帳 ✓ sedaily。淨現金成為官方實報的 ₩69.4兆（≈$48B）。Capex 指引上修至 high-₩40兆；與約 10 家客戶簽長約；HBM4 開始量產出貨。DRAM 混合 ASP 季增約 +30%（1Q 是 +60-70%）：還在漲、但肉眼可見地減速 ✓ 法說轉述×2。

(ii) Scoring our own tripwire. This page pre-committed on 07-22: OP ≥₩64tn = base confirmed / <₩60tn = migrate toward bear. The actual ₩60.54tn landed in the gray zone between the two lines — neither branch fired cleanly. Ruling ◆: roll the actuals through the model mechanically (2026E OP ₩258→240tn, 2027E OP ₩300→275tn, 2027E EPS ₩327,000→₩300,000 — the T3 table shows every 調整前/調整後), reweight the tails 30/45/25 → 35/45/20, and cut the TP to ₩1,800,000 (= ₩300,000 × the unchanged 6×; the cut is 100% EPS, 0% multiple).

(ii) 給自己的絆線打分。本頁 07-22 預先承諾：OP ≥₩64兆＝基準確認／<₩60兆＝轉向 bear。實際的 ₩60.54兆落在兩條線之間的灰色地帶 — 兩個分支都沒有乾淨觸發。裁決 ◆：實績機械性滾入模型（2026E OP ₩258→240兆、2027E OP ₩300→275兆、2027E EPS ₩327,000→₩300,000 — 每一筆調整前/調整後見 T3 表），尾部權重 30/45/25 → 35/45/20，目標價下修至 ₩1,800,000（＝₩300,000 × 不變的 6 倍；下修 100% 來自 EPS、0% 來自倍數）。

(iii) The tape. The market cut faster than the model: -11.9% on 07-28, -11.16% on 07-29 to a ₩1,377,000 close — two days -21.7%, -53.9% off the June peak ✓ stockanalysis + computed. That close is 13.9% below the ₩1,600,000 entry zone this page pre-committed to.

(iii) 盤面。市場砍得比模型快：07-28 -11.9%、07-29 -11.16% 收 ₩1,377,000 — 兩日 -21.7%，距 6 月峰 -53.9% ✓ stockanalysis+計算。這個收盤價比本頁預先承諾的 ₩1,600,000 進場區還低 13.9%。

(iv) The conclusion changed — and here is exactly why. View upgraded range-bound → Buy ◆. The estimate fell -8%, the price fell -27% since the last update; the cut TP now implies +30.7%, which sits above the +15% line this page's own view-band table defines as Buy. The evidence leg came in gray, so the entry is staged, not full: first tranche at market, second on the late-Aug TrendForce 3Q contract-price print (or <₩1,250,000). The remaining honest risk: the market may be sniffing the bear path (a 2027 price rollover starting early) — that is exactly what the late-Aug print and the 10-28 3Q report (trim line: OP <₩65tn) are pre-committed to arbitrate.

(iv) 結論改變了 — 理由完整攤開。觀點上調：區間 → 買進 ◆。自上次更新以來，估計值跌了 -8%、價格跌了 -27%；下修後的 TP 隱含 +30.7%，已越過本頁自己的觀點區間表所定義的買進門檻（>+15%）。證據腿只到灰色地帶，因此進場分批、不全倉：第一批現價，第二批看 8 月下旬 TrendForce 3Q 合約價（或 <₩1,250,000）。誠實面對剩下的風險：市場可能正在嗅 bear 路徑（2027 價格 rollover 提前開始）— 而這正是 8 月下旬讀數與 10-28 3Q 財報（減碼線：OP <₩65兆）被預先指定去仲裁的事。

(v) Also refreshed throughout: masthead (close/market cap ₩1,004tn/PBR ≈3.8× on the post-Kioxia ₩363,000 BVPS/fwd PER 4.6×), scenario diagram, asymmetry check, playbook, and the sources ledger (now 97 claims). Consensus figures cited on this page remain pre-print and will reset. Hero image reviewed, kept.

(v) 全頁同步更新：頁首（收盤／市值 ₩1,004兆／PBR 以 Kioxia 入帳後 BVPS ₩363,000 計約 3.8×／fwd PER 4.6×）、情境圖、不對稱檢驗、操作框架、資料來源表（現為 97 項）。頁上引用的共識均為財報前口徑、將重置。Hero 圖已檢視、維持不變。

(vi) Reader-facing additions, same day: every load-bearing ₩ figure now carries a ≈US$ approximation (₩1,450/$, 07-29, rounded), and the Playbook gains an ADR (SKHY) translation box — per-ADS values, the observed premium history (+4.9% IPO → +21.6% 07-28 → ≈+34% pre-market), TP/entry zones in ADS terms, and a premium-discipline rule (≤25%).

(vi) 同日新增（讀者需求）：全頁主要 ₩ 數字附 ≈US$ 換算（₩1,450/$，07-29，四捨五入），操作框架新增 ADR（SKHY）換算框 — 每 ADS 內含價值、溢價觀測史（IPO +4.9% → 07-28 +21.6% → 盤前 ≈+34%）、目標價／進場區的 ADS 版本、以及溢價紀律（≤25%）。

(vii) Layout compacted, same day: the default read is now the decision spine (Summary → Numbers → Catalysts → Playbook, ~8 min); the full argument (business & tech, moat, bull/bear, variant, management, competition) is collapsed under "Full case", older update-log entries fold away, and two argument-shaped duplicates were removed outright (the bear rebuttal diagram; the leading-indicators table, whose one unique row — TSMC CoWoS — moved into the catalyst table). Nothing substantive was deleted — collapsed, not cut. ◆ = judgment/estimate throughout this entry unless marked otherwise.

(vii) 同日版面精簡：預設閱讀動線縮為決策主幹（摘要 → 數字 → 催化劑 → 操作框架，約 8 分鐘）；完整論證（事業與技術・護城河・多空・變異・經營層・競爭）收合於「完整論證」，更新紀錄較早條目收合，並直接移除兩個重複結構（空方「反駁結構」示意圖；「領先指標」表 — 其唯一不重複的一列 TSMC CoWoS 併入催化劑表）。實質內容零刪除 — 是收合、不是砍掉。◆ ＝本紀錄除另有標註外，皆屬本研究判斷/估計。

✓ 2026-07-23Restructured to the pruned format + numbers refreshed — view and TP unchanged結構修剪＋數字更新 — 觀點與目標價不變

(i) Rebuilt per the 2026-07-23 layout: the forecast, valuation, and scenario sections were merged into one "Numbers, valuation & scenarios" chapter; the summary dropped its key-numbers grid and annual table; moat and leading indicators were promoted into a standalone, expanded section; the jargon decoder moved to the back of the page; and the sources ledger now collapses by default (every underlying model detail is still preserved in the research files — nothing was deleted from the record, only from the page).

(i) 依 2026-07-23 版式修剪：財測/估值/情境三章整合為「數字、估值與情境」，摘要移除關鍵數字卡與年度表，護城河與領先指標升為獨立章節並擴充，名詞解碼移至頁尾，來源表改為預設收合（所有模型細節仍保存於研究檔案，未刪除）。

(ii) Price refresh: ₩1,839,000 (07-22 close) → ₩1,882,000 (07-23 intraday, +2.84%), -37.0% off the peak, TP ₩1,960,000's implied headroom moved from +6.6% to +4.1%. Research view stays range-bound.

(ii) 價格更新：₩1,839,000（07-22 收）→ ₩1,882,000（07-23 盤中 +2.84%），距峰 -37.0%，TP ₩1,960,000 隱含空間 +6.6% → +4.1%，觀點區間不變。

(iii) The two mechanism illustrations (token-flood, epoxy-pour) were regenerated in a more photoreal style; filenames are unchanged.

(iii) 兩張機制概念圖（token-flood、epoxy-pour）改以寫實風格重新生成，檔名不變。

(iv) This page is published to /learn as of today. ◆ = judgment/estimate throughout this log entry unless marked otherwise.

(iv) 本頁自本日起發布於 /learn。◆ ＝本紀錄除另有標註外，皆屬本研究判斷/估計。

◆ initial初次評估2026-07-22 — Initial coverage2026-07-22 — 初次評估

First build of this page. Research view: range-bound. Target price ₩1,960,000 (6-month horizon, 2027E EPS ₩327,000 × 6× PER), implied upside +6.6% vs the ₩1,839,000 close. No prior model exists to revise against.

本頁首次建置。研究觀點：區間。目標價 ₩1,960,000（6個月，2027E EPS ₩327,000 × 6倍PER），對收盤 ₩1,839,000 隱含漲幅 +6.6%。無前一版模型可供對照。

Date日期	Action動作	View觀點	TP	Close收盤
2026-07-22	Initial coverage初次評估	Range-bound區間	1,960,000	1,839,000
2026-07-23	Restructure + price refresh結構修剪+價格更新	Range-bound區間	1,960,000	1,882,000
2026-07-29	2Q26 earnings update — estimates cut, tails reweighted, entry zone triggered2Q26 財報更新 — 估計下修、權重重估、進場區觸發	Buy (prior: range-bound)買進(前次:區間)	1,800,000 (prior前次:1,960,000)	1,401,000 ✗→
2026-08-04	HBF standard + three-session rebound — estimates unchanged, entry zone re-derived to the Buy boundary, 07-29 close correctedHBF 標準＋三日反彈 — 估計不變、進場區改繫於買進邊界、07-29 收盤價更正	Range-bound (prior: Buy)區間(前次:買進)	1,800,000 (unch.維持)	1,567,000

Append-only ledger. The first estimate-revision table (調整前/調整後) appears in §Numbers as of 07-29. The 07-29 close is shown corrected (✗→ from ₩1,377,000); entries are never deleted, only restated on the record.附加式(append-only)紀錄。首張修正對照表（調整前/調整後）自 07-29 起見〈數字〉章 T3。07-29 收盤價為更正後數字（✗→ 原載 ₩1,377,000）；條目永不刪除，只在紀錄上更正。