Global semi news — Korea, China, Taiwan, the US, and Japan. Government policy, export controls, capex moves, supply-chain shifts, and macro events. AI-classified and tagged with affected tickers. All headlines link back to the originating publisher.
Original: 即使大摩認為輝達 GPU 物有所值，但 AI ASIC 發展仍是不可忽視勢力
Morgan Stanley validates NVIDIA's premium pricing, finding that while Blackwell-based data center builds cost 2× more than those using Google TPUs or Amazon Trainium, NVIDIA's performance-per-watt advantage spans 2–8×. Benchmark scores put NVIDIA's upcoming Vera Rubin (FP4) at 19.5 versus Google TPUv7 at 4.3 and Amazon Trn3 at 2.5. However, inference-cost metrics are gaining traction: Groq prices at $0.05–0.10 per million tokens at 800 tok/s, versus Blackwell's $0.25 per million tokens at 450 tok/s, signaling a credible cost-efficiency challenge in the inference segment.
Why it matters: An analyst benchmarking note on GPU vs. ASIC efficiency is useful sector context for NVIDIA's supply chain (TSMC, SK Hynix), but discloses no new capex commitments, contracts, or earnings-moving data.
Original: 英特爾連五黑 分析師仍樂觀、看好 CPU 市場成長
Intel has pulled back ~16% from last week's all-time high of $129.44 over five sessions, but Citi (target lifted to $130) and Benchmark Equity Research (target $140) remain bullish, projecting AI-driven CPU TAM growth of 35% annually to ~$132B by 2030 as inference workloads elevate CPU demand. Reports indicate Intel has reached a preliminary wafer foundry agreement with Apple, facilitated by the US government, a development that — if confirmed — would introduce direct competition for TSMC's dominant Apple foundry relationship.
Why it matters: Analyst upgrades and CPU TAM forecasts are sector commentary, but the unconfirmed Apple–Intel foundry deal rumor introduces a supply-chain threat to TSMC (2330) that warrants monitoring without yet rising to a confirmed stock-moving event.
Original: 陳立武稱 Intel 18A 是國家寶藏，Intel 14A 定 2028 年量產與台積電 A14 競爭
Intel CEO Lip-Bu Tan disclosed multi-year foundry contracts with Apple and Elon Musk's TeraFab using Intel 18A, with yield improvement now hitting industry-standard monthly rates and defect-density targets met ahead of schedule, enabling Panther Lake CPU volume shipments. Intel 14A is scheduled for risk production in 2028 and volume ramp in 2029, deliberately synchronized with TSMC's A14 node, with PDK 0.5 already available to design partners. EMIB advanced packaging yield reached 90%, but an ABF substrate shortage is tightening supply chains and prompting select customers to prepay for materials.
Why it matters: Named customer contracts (Apple, TeraFab) and an explicit competitive roadmap synchronized with TSMC's A14 are material events with direct stock-price implications for TSMC and rival foundries.
Open source articleOriginal: 蘋果找英特爾代工也難撼動台積電地位
Bernstein analyst Mark Li argues that Apple's rumored shift of M7 (2027, Intel 18A-P process) and A21 (2028, Intel 14A) production to Intel is too small in scale to threaten TSMC's leadership, noting Intel has not meaningfully closed the technology gap. TSMC remains the only foundry capable of stable 2nm mass production, while Samsung's GAA 2nm is assessed as comparable to TSMC's 3nm in real-world performance. Bernstein reiterates Outperform on TSM with a raised PT of $430, and separately notes AMD may have shifted some 2nm Venice/Veranos CPU orders to Samsung, keeping advanced-node competition in focus.
Why it matters: Analyst PT raise and competitive roadmap colour are noteworthy, but the story explicitly reaffirms the status quo with no new contract, capex announcement, or earnings data to move the stock.
Open source articleOriginal: 英特爾 Wildcat Lake 首批筆電現身中國，榮耀 X14 價位直逼 MacBook Neo
Intel's Wildcat Lake processors have debuted commercially in China via Honor's X14 2026 laptop at 4,399 RMB (~$645), matching Apple's MacBook Neo base price (~$600) while offering superior specs of 16GB RAM and 512GB SSD versus MacBook Neo's 8GB/256GB. Benchmark leaks suggest the Core 7-350 runs ~9% faster in multi-core than Apple's A19 Pro, though Apple retains a ~19% single-core advantage, reflecting different design priorities. ASUS models are also appearing on JD.com, signaling rapid Wildcat Lake penetration into the entry-level thin-and-light segment and a broader Intel push to challenge Apple at the ~$600 price point.
Why it matters: Competitive product launch story with indirect supply-chain read-through to TSMC (likely Wildcat Lake compute-tile fab) and SK Hynix (LPDDR memory), but no named contract, capex event, or direct earnings catalyst for tracked names.
Open source articleOriginal: AI 搶光記憶體，比起價格蘋果現在更怕買不到 DRAM
The AI data center buildout has pushed DDR5 prices up over 400% in some markets, stripping Apple of its historically dominant supply-chain leverage and forcing it to prioritize securing DRAM allocation over price negotiation — a shift CEO Cook has publicly acknowledged. Samsung Electronics and SK Hynix are direct beneficiaries as memory suppliers tighten contract structures to five-year terms with stricter breach penalties. Hana Securities analyst Kim Rok-ho warns LPDDR competition will intensify further when NVIDIA's Rubin AI platform begins shipping in Q4 2026, with prices seen as having only one direction to go.
Why it matters: Supply-chain and market-data story confirming structural DRAM tightness with a specific analyst price outlook, but lacks a primary capex announcement, contract disclosure, or earnings event to qualify as high.
Open source articleOriginal: 成熟製程進上行循環！外資上調聯電、世界先進與環球晶目標價
An American brokerage upgraded UMC's rating and raised target prices for UMC, Vanguard International Semiconductor (5347), and GlobalWafers (6488, downgraded but TP raised), citing mature-process nodes entering a new upcycle as AI-server-driven PMIC demand offsets weak consumer electronics. UMC wafer prices are forecast to rise 5–10% in H2 2026 with further increases in 2027, and capacity is expected to remain tight through 2028. TSMC outsourcing interposer production to Vanguard and exercising more disciplined legacy-node capacity management are flagged as structural supply tighteners.
Why it matters: Named analyst rating changes (UMC upgraded, GlobalWafers downgraded) with quantified target price revisions and a specific 5–10% wafer price forecast for H2 2026 constitute clear stock-moving events across multiple tracked tickers.
Open source articleOriginal: AI 推論晶片成矚目焦點！傳新創 Tenstorrent 吸引英特爾、高通收購興趣
AI inference chip startup Tenstorrent is in early-stage talks with Intel, Qualcomm, and other industry players about a potential acquisition, with sources indicating a deal could value the company above $5 billion; the startup is simultaneously pursuing a new funding round and working with investment banks on strategic options. The story unfolds amid a broad wave of AI chip M&A: NVIDIA recently struck a ~$20B licensing deal covering Groq assets, while Cerebras Systems—which rejected acquisition overtures from Arm and SoftBank—completed a $5.55B IPO last week and now trades at roughly a $65B market cap. No TW or KR names are directly named, but the consolidation trend underscores intensifying competition in the AI inference accelerator space.
Why it matters: Signals accelerating M&A consolidation in the AI inference chip sector, relevant to the competitive landscape for chipmakers, but no named TW or KR company is a direct party to the transaction.
Open source articleOriginal: 輝達財報在即 選擇權示警：恐成美股關鍵轉折
Morgan Stanley raised its Nvidia target from $260 to $285 and KeyBanc lifted its target from $275 to $300, both maintaining Overweight ratings ahead of Wednesday's (May 20) after-hours print, citing robust Blackwell Ultra GPU demand and early Rubin GPU shipments. Options strategist Brent Kochuba (SpotGamma) warns that heavy call positioning could trigger broad profit-taking, potentially making the earnings a near-term inflection point for the broader equity rally. The outcome is closely watched across the AI GPU supply chain, with TSMC, SK Hynix, and AI server ODMs among the most exposed names.
Why it matters: Concrete analyst target-price upgrades ahead of a named earnings date, combined with an options-market inflection warning, constitute a clear stock-moving event for Nvidia's direct supply-chain partners—TSMC, SK Hynix, and AI server ODMs—within the tracked universe.
Original: 法院裁定關鍵產線不得停擺，三星罷工施壓空間大幅縮小
South Korea's Suwon District Court granted Samsung Electronics a partial injunction on May 18, ordering unions to maintain normal staffing and operations in critical semiconductor areas—safety systems, equipment maintenance, and wafer protection—throughout the planned May 21 strike. Violations carry fines of KRW 100M/day (~$73K) per union entity and KRW 10M/day (~$7.3K) per union leader. The court stopped short of banning the 18-day walkout by roughly 50,000 workers, but the ruling materially curtails union leverage and limits the production-disruption threat to Samsung's fabs.
Why it matters: Operationally significant for Samsung's semiconductor output risk, but the partial injunction is a risk-limiting rather than a clear earnings-moving catalyst, and no production guidance or financial figures were revised.
Kioxia
285A
¥67,100
-12.86%