Zero One Daily Intelligence Brief — May 21, 2026

NVIDIA tops $81bn as Samsung defuses strike and AMD's Venice arrives on TSMC 2nm

Top Market Signals

NVIDIA breaks out Physical AI, signals GPU demand beyond hyperscalers

NVIDIA (NVDA US) used the Q1 FY27 call to reframe its forward demand story as a broadening of the GPU buyer base, not deeper pockets from the same five hyperscalers. Physical AI was broken out as its own reporting line for the first time, an ACIE (AI Compute and Inference Engine) sub-segment was introduced, Anthropic was named as a partner alongside the hyperscaler set, and a standalone Vera CPU addressable market was flagged separate from the Vera Rubin GPU platform.

Q1 FY27 revenue was US$81.6bn (+85% YoY), data center revenue US$75.2bn, Q2 FY27 guide US$91bn, with a new US$80bn buyback and a dividend hike. Q2 is the first full quarter after April US export restrictions effectively closed the China market, with CEO Jensen Huang telling CNBC the company has "largely conceded" China to Huawei. NVDA shares slid in extended trade despite the print.

Linked stocks: NVDA US, 2330 TT, 2454 TT, 9984 JT

Sources: Nvidia, 美出口管制影響，黃仁勳：中國 AI 市場已讓給華為, 輝達財報一人救半導體全村, 〈台股盤後〉AI飆風狂吹 大漲1347點創史上第五大漲點收41368點

AMD's EPYC Venice goes into volume production on TSMC's 2nm node

AMD (AMD US) announced its sixth-generation EPYC processor "Venice" is now in volume production on TSMC (2330 TT) N2 in Taiwan, making it the first HPC product to ramp on TSMC's 2nm process. AMD plans to extend N2 production to TSMC's Arizona fab and confirmed that the follow-on Verano CPU will also use N2 and integrate LPDDR for memory-bandwidth-constrained workloads. The disclosure landed on the eve of Computex, with AMD CEO Lisa Su personally in Taipei to brief supply chain partners.

Alongside the Venice ramp, AMD pledged over US$10bn of investment into Taiwan's AI ecosystem, covering advanced packaging capacity and joint development with TSMC on SoIC-X and CoWoS-L. The commitment positions Taiwan as the primary build-out site for AMD's data center roadmap.

Lisa Su brought AMD's first-ever AI Dev Day outside the United States to Shanghai on 19 May (with a Kai-Fu Lee dialogue on AI agents), then flew straight to Taiwan ahead of Jensen Huang's expected 27 May return. AMD is keeping developer mindshare in China and production volume in Taiwan, sitting opposite NVIDIA's posture of having conceded the China market.

Venice on N2 also fills TSMC 2nm capacity that AI accelerator customers would otherwise contest. The 11 May Blade flagged AMD picking up TSMC capacity from MediaTek and Qualcomm; today's announcement extends that pattern up to the next node.

Linked stocks: AMD US, 2330 TT, INTC US

Sources: AMD EPYC Venice量產 搶下台積2奈米HPC首發, AMD蘇姿丰率先來台高調報喜 宣布採用台積2奈米的處理器進入量產, 黃仁勳不期待大陸 超微蘇姿丰雙邊布局，訪陸後先來台爭台積2奈米, 蘇姿丰：超微在台灣產業體系投資逾100億美元

Samsung defuses strike with stock-based bonus pool tied to AI memory profits

Samsung Electronics (005930 KS) reached a preliminary labor agreement 90 minutes before a planned 18-day strike by roughly 48,000 workers was due to begin, after South Korea's Employment and Labor Minister stepped in to mediate. The deal averts what UDN estimated at KRW 1tn per day in lost production, and removes a near-term tail risk on global AI memory supply.

Samsung's Device Solutions division (semiconductors) gets an uncapped special performance bonus pool sized at 10.5% of its operating profit, layered on top of the existing OPI program, with payout conditioned on the DS unit exceeding KRW 200tn of cumulative operating profit over the next three years. Forty percent of the pool goes to roughly 78,000 DS workers in equal shares; sixty percent flows by sub-division performance. Per-worker bonuses could reach KRW 600m (about US$440k), paid entirely in vested Samsung treasury stock with one-third available immediately, one-third after a year, and one-third after two years.

Loss-making units (Foundry, System LSI) still collect bonus this year because the penalty mechanism is delayed to 2027, which has drawn industry pushback for breaking the performance-pay link. Former Samsung semi head Gyung Gye-hyun separately warned that aggressive China memory capacity could re-introduce oversupply by 2H 2027 and that hyperscaler capex ROI could compress demand by 2028.

Stocks rallied on the deal. KOSPI surged 5.3% in morning trade, Samsung Electronics added 6%, SK Hynix (000660 KS) jumped 11%, and Micron (MU US) added 4.76% to US$731.99 overnight. The labor-cost floor moved up at the same time: Micron is now offering Korean HBM design engineers up to KRW 300m per year (Seoul-based contracts) to defect, lifting the per-engineer cost base across the entire HBM4 cycle.

Linked stocks: 005930 KS, 000660 KS, MU US, 2408 TT

Sources: 三星電子驚險達勞資協議　員工最高年領1400萬獎金, 三星勞資協議內容出爐！半導體員工人均可領1260萬特別績效獎金, 三星勞資協議分紅包括虧損部門, 三星罷工急煞引爆記憶體！美韓指標瘋漲 南亞科成交值一度超車台積電

Nanya plans 30,000 wpm new fab, sees DRAM tight through end-2027

Nanya Technology (2408 TT) chairman Tsou Ming-jen used the company's AGM to declare that the DRAM industry has shifted from its historical cyclical pattern to "structural growth" driven by AI demand. He flagged 2026 capex above NT$52bn, on top of a NT$78.7bn private placement completed in April (where three of four subscribers are cloud SSD operators), and forecast that the supply-demand gap will persist through end-2027 and growth into 2028.

President Lee Pei-ying said Nanya's new 5A fab will begin equipment install in Q1 2027 and ramp into production in H2 2027, adding roughly 30,000 wafers per month. If schedule holds, Nanya's total capacity could expand 80-100% over the next three years, with the current 1B node (10nm second-gen) shipping 16Gb DDR5, 8Gb DDR4 and 4Gb DDR4; 1C entering verification in H2 2026; and 1D 16Gb DDR5 entering pilot in Q2 2026.

Nanya is also pursuing a differentiated AI memory line. The 128GB DDR5 RDIMM 5600/6400 Mb/s server module using TSV has passed functional test, and Nanya is co-developing what Tsou called "ultra-high-bandwidth memory" using Wafer-to-Wafer Bonding, positioned above current JEDEC HBM standards rather than as a standard HBM3/HBM4 follower. Lee said the product is already contributing small revenue and should scale into H1 2027.

Nanya opened up at NT$291.5 on the AGM and at one point traded more value than TSMC, while Powertech (6239 TT) hit limit-up and Phison (8299 TT) and ADATA (3260 TT) also rose. The 27 Apr Blade flagged Nanya as the first Taiwanese DRAM vendor in NVIDIA's Vera Rubin LPDDR supply chain; today's commentary frames that entry as the start of a multi-year build-out.

Linked stocks: 2408 TT, 2344 TT, 6239 TT, 8299 TT

Sources: TechNews, 新廠 2027 年下半年量產！南亞科預期 DRAM 缺口延至明年底, 南亞科預估DRAM缺貨到明年底 後年也看成長

Panel-level packaging coalition hardens around TSMC's CoPoS

Innolux (3481 TT) hit limit-up at NT$40.6, a 15-year high, after Korean media confirmed that SK Hynix-affiliated materials maker SKC will commit roughly NT$25bn to glass substrates for panel-level packaging, with around KRW 589.6bn flowing through US subsidiary Absolics for three-year mass production. The TSMC-Innolux-SKC triangle has been building since Innolux's FOPLP partnership with TSMC at Longtan was first surfaced in the 11 May Blade.

Chairman Hong Chin-yang told shareholders that beyond the existing FOPLP work, Innolux is now actively developing RDL interposer (redistribution-layer interposer) for large-die AI accelerators, with technical validation already underway for major packaging customers. Innolux is also co-developing TGV (through-glass via) with a global tier-1 customer. Both RDL interposer and TGV are core enablers of TSMC's CoPoS (chip-on-panel-on-substrate), the panel-based successor to CoWoS, which TSMC has guided for mass production as early as 2028.

Lam Research (LRCX US) announced the opening of a Panel-Level Packaging Center of Excellence in Salzburg, Austria on the same day, built on its 2022 acquisition of Semsysco. The CoE focuses on wet chemical processes for square substrates of various sizes and thicknesses, using Lam's Kallisto and Phoenix platforms for electrodeposition, etch and cleaning. Capacity for FOPLP is now being built up at the toolmaker level alongside the substrate, packaging house, and foundry layers, mirroring the HBM build-out sequence of 2022.

Linked stocks: 3481 TT, 2330 TT, LRCX US, AMAT US

Sources: SK 集團重押玻璃基板 群創衝破40元關卡、刷15年新高, 科林研發在奧地利成立面板級封裝研發中心，強化先進封裝布局

ASML's Fouquet says chip shortage is multi-year, equipment stocks rip

ASML (ASML NA) CEO Christophe Fouquet, in a rare Reuters interview, said AI, satellite and robotics demand has run far ahead of capacity and that supply tightness will persist long-term, with the chip market reaching US$1.5tn by 2030. He flagged Elon Musk's "TeraFab" plans and Starlink as driving a new demand wave on top of the AI build-out, and confirmed ASML will publish High-NA logic and memory product data this year while developing a second advanced packaging tool for larger AI dies.

Equipment stocks moved on the message. ASML ADR rose 6.21% to US$1,550, with Applied Materials (AMAT US) up 4.9%, Lam Research up 6.84% and KLA (KLAC US) up 5.11%. ASM International and BE Semiconductor also closed higher in Europe. UBS raised ASML's price target to EUR1,900 (about US$2,200), implying over 40% upside, and named the stock its top semi pick on the view that AI capex hand-off will lift equipment book-to-bill through 2026-2027.

NVIDIA, Nanya and ASML all framed the cycle as supply-constrained in the same 24-hour window: the customer, the memory IDM, and the lithography monopolist on the same page. That alignment is rare outside the peaks of prior super-cycles.

Linked stocks: ASML NA, AMAT US, LRCX US, KLAC US

Sources: ASML 執行長稱晶片供應將長期吃緊 設備股嗨

先聲 First Word — Exclusives from Chinese-Language Sources

Global Unichip raises ASIC guide above GPU, Tseng Fan-cheng steps down

Global Unichip (3443 TT) CEO Tsao Shang-yi told the company's AGM that customized compute is entering a fast-growth phase and that ASIC shipment growth will exceed GPU's, citing Google's TPU as the example, with GUC providing CoWoS and Physical Design support on customer 2nm and 3nm projects. Founding chairman Tseng Fan-cheng retired after 23 years; ex-TSMC Senior Director Tsai Li-su was unanimously elected to succeed him, with TSMC holding four board seats including CFO Wendell Huang and R&D VP Lu Li-chung. (21 May 2026) Source: 創意總座：「客製化算力」快速成長期來了！ ASIC成長率將超越GPU

MediaTek doubles 2026 AI ASIC guide to US$2bn on Google TPU 8th-gen

MediaTek (2454 TT) CEO Rick Tsai, in a recent investor call, doubled the 2026 AI ASIC revenue guide from US$1bn to US$2bn after winning Google's 8th-generation TPU, with a second AI ASIC project targeting end-2027 production and a global ASIC TAM projection of US$70-80bn by 2027. MediaTek's 2nm SoC is on track for Q3 2026 launch; Q1 26 results were revenue NT$149.2bn and EPS NT$15.17, with handset revenue down 17% QoQ on memory-driven cost pressure. (21 May 2026) Source: 聯發科關禁閉出關，Google ASIC 大單拉抬預期業績帶動股價攻上漲停

Tripod tracks to 38% Q2 server mix as analysts hike full-year forecasts

Tripod Technology (3044 TT) printed Q1 26 revenue of NT$20.96bn (+22% YoY) and EPS NT$5.6, with gross margin up 1.1pp QoQ to 26.5%; April revenue of NT$8.2bn (+28% YoY) sets up roughly 17% QoQ growth into Q2. Analysts now model server revenue mix climbing to 38% in Q2, full-year 2026 revenue +29% YoY, and capex above NT$10bn split between China and Vietnam. Server and networking order visibility already extends into H2 2026. (21 May 2026) Source: 伺服器需求強勁 法人調高健鼎獲利預估及目標價

AMD takes first AI Dev Day outside the US to Shanghai, then crosses to Taipei

AMD ran its first-ever AI Dev Day outside the United States in Shanghai on 19 May, with CEO Lisa Su in dialogue with 01.AI founder Kai-Fu Lee on AI agents, multi-agent collaboration, and inference architecture; AMD says its Shanghai R&D centre is now one of its global priority sites. Su then flew straight to Taipei the next day to brief TSMC ahead of the Venice production announcement. AMD operates R&D centres in Beijing, Shanghai, Shenzhen and Taipei with over 4,000 engineers combined. (21 May 2026) Source: 黃仁勳不期待大陸 超微蘇姿丰雙邊布局，訪陸後先來台爭台積2奈米