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: 搶食博通大單！聯發科化身「IC 設計界台積電」，憑兩大優勢決戰雲端晶片
MediaTek (2454) raised its 2026 ASIC revenue guidance from $1B to $2B at its April 30 call, with the Street reading this as confirmation it wrested Google's TPU training-chip order from Broadcom, lifting its market cap by NT$2 trillion in a month. Analysts now expect ASIC to surpass handsets as MediaTek's top revenue line by 2027, leveraging its TSMC (2330) advanced-node leverage from 1.5B annual mobile/IoT chip shipments to compete in cloud AI silicon.
Why it matters: MediaTek doubled ASIC revenue guidance and is reported to have won Google's TPU training-chip business from Broadcom — a major stock-moving contract win with NT$2T market-cap impact already realized.
Original: NVIDIA營運報告改版，認為偏利空解讀
NVIDIA posted a Q1'27 double-beat (revenue $81.6B vs $79B est, EPS $1.87 vs $1.77) with Q2 guide of $91B, but the analyst flags that the new segment structure—splitting Data Center into Hyperscale ($38B, +12% QoQ) and ACIE ($37B, +31% QoQ) instead of Compute/Networking—appears designed to mask a Compute miss offset by stronger Networking (a CPO tailwind). The shift also signals NVIDIA is leaning more on sovereign AI and Neocloud buyers as hyperscalers push ASIC alternatives, implying less stable order flow but stronger pricing power ahead.
Why it matters: Sector-level commentary on NVIDIA's segment reporting change and AI supply-chain dynamics—relevant to TW AI supply chain (TSMC, Hon Hai, CPO names) but no direct named ticker catalyst.
Original: 今天老虎五隻！千如⊕、友達⊕、國精化⊕、TPK⊕、富采⊕
Taiwan's TAIEX rose 899 points to 42,267 on NT$1.18T (~US$36B) turnover, with foreign investors net buying NT$76.1B (~US$2.3B) and a two-day total of NT$136.4B (~US$4.2B). The rally is driven by North American CSP capex forecast above US$800B in 2026 and approaching US$1T by 2027, plus AMD's pledge of over US$10B additional investment in Taiwan, lifting TSMC (2330), MediaTek (2454), ABF substrates, CoWoS, PCB/CCL, and memory names.
Why it matters: Market commentary and sector roundup citing macro AI capex trends and broad Taiwan supply-chain beneficiaries rather than a specific stock-moving event for tracked names.
Open source articleOriginal: 洪麗寗：AI時代下電力與產能密不可分 緯穎已將電力布局至2028年
Wiwynn (6669-TW) Chair Emily Hong said power supply has overtaken geopolitics as the top constraint on AI server capacity, noting AI server testing is as power-hungry as data centers themselves. Wiwynn has locked in power planning through 2028 across its Mexico, US, and Malaysia sites, fully exited China in 2022, and flagged July's USMCA renegotiation as a new uncertainty for its tariff-free US exports from Mexico.
Why it matters: Strategic capacity and power-planning color from the Wiwynn chair, plus a USMCA renegotiation flag, but no specific capex, contract, or guidance figures to move the stock immediately.
Original: 兩天吞噬缺口！這些股票為何這麼強？
Taiwan's TAIEX surged 899.76 points on 5/22 to a record 42,267.97, fully recovering a 2,000-point mid-May drop in just two sessions, driven by NVIDIA's blowout earnings/buyback, AMD's $10B Taiwan investment pledge, and a post-lockup rally in TSMC (2330) and MediaTek (2454). The piece also flags Inventec (2356) as an AI-server beneficiary with US CSP pull-in for NVIDIA's next-gen platform lifting AI server mix above 30%, plus Amtran (2489) and Taiflex (8039) as smaller idea names.
Why it matters: Broad index-recovery commentary with stock-tout flavor; macro drivers (NVIDIA earnings, AMD capex) are already-known catalysts and individual picks are advisory-style rather than fresh stock-moving events.
Open source articleOriginal: 記憶體堆疊達極限！傳 HBM 不再緊貼 GPU，改走「光學互連、分離封裝」新架構
Industry sources say memory and packaging players are evaluating a new architecture that separates HBM from the GPU and bridges them with optical interconnects, freeing capacity from GPU shoreline limits as vertical stacking past 16-20 layers nears physical limits. OSAT executives view chip-to-chip optical links as the inevitable next step after rack-to-rack and server-to-server adoption, with discussions already underway with GPU vendors on form-factor changes — implicating HBM leaders SK Hynix and Samsung, advanced-packaging foundries, and CPO component suppliers.
Why it matters: Roadmap/sector story on a next-generation HBM-GPU packaging architecture with no formal roadmap or named contracts yet, but it directly affects HBM leaders, advanced packaging, and CPO suppliers.
Open source articleOriginal: 南亞科出貨輝達快達陣了，低功耗產品傳在台積亞利桑那廠驗證
Taiwan DRAM maker Nanya Tech said its LPDDR5X low-power memory is being validated at TSMC's Arizona fab for Nvidia's next-gen Vera Rubin AI platform, with advanced packaging and shipments to follow. CEO Pei-Ing Lee confirmed multiple international AI customers, marking Taiwan's first push into the AI core supply chain via low-power DRAM and a potential challenge to Korean incumbents Samsung and SK Hynix.
Why it matters: Named-customer AI supply-chain win for Nanya at TSMC Arizona is a clear stock-moving event with direct read-through to Korean DRAM incumbents losing share in low-power AI memory.
Open source articleOriginal: 大摩：輝達記憶體成本飆 435% Vera Rubin 恐天價
Morgan Stanley estimates each next-gen Nvidia VR200 NVL72 rack will cost hyperscalers ~$7.8M versus $4M for GB300 NVL72, with memory alone accounting for ~25% of BOM at roughly $2M per rack — a 435% jump from the prior generation. The rack packs 54TB of LPDDR5X (3x GB200) plus ~$1M of 3D NAND, a clear demand tailwind for DRAM/NAND suppliers including SK Hynix and Samsung, especially on premium SOCAMM2 modules.
Why it matters: Sell-side BOM analysis showing a 435% memory content surge per AI rack is a direct, quantifiable demand signal for Korean DRAM/NAND suppliers and LPDDR5X/SOCAMM2 leaders.
Open source articleOriginal: 三星李在鎔祕密來台挖台積電牆腳，拜會聯發科蔡力行爭取代工訂單
Samsung Chairman Lee Jae-yong led a top-level team to Taiwan on May 21 to meet MediaTek CEO Rick Tsai, offering preferential HBM/memory supply for next-gen Dimensity SoCs in exchange for foundry orders — the same bundling tactic that won Qualcomm. The visit, which also reportedly included poaching attempts at TSMC, comes as MediaTek shifts Google's inference TPU advanced packaging to Intel while keeping training TPU at TSMC, hinting at a cooling MediaTek-TSMC relationship.
Why it matters: Named foundry customer poaching attempt by Samsung's chairman targeting MediaTek, with concrete bundling strategy and direct competitive implications for TSMC's largest non-Apple customer relationship.
Open source articleOriginal: 〈台股盤後〉再刷史上第9大漲點899點收復4萬2 周線翻長紅漲1095點
The TAIEX jumped 899.76 points (+2.66% on the week, +1,095.61pts) to close at 42,267.97 on NT$1.18T turnover, the 9th-largest single-day gain in history. TSMC (2330) rose 1.12% to NT$2,255 and MediaTek (2454) jumped 8.73% to NT$3,860, while CCL/high-speed material names including Taiwan Union Tech (2383) and IC substrate maker Nan Ya PCB (8046) advanced on AI-driven demand.
Why it matters: Broad market wrap with sector rotation context (passives, CCL, IC substrates on AI demand) rather than a specific stock-moving catalyst.
Open source articleKioxia
285A
¥67,100
-12.86%