Zero One Daily Intelligence Brief — April 27, 2026

TSMC Ramps 3nm Capacity 20% Above Plan as AI Demand Strains Every Link in the Chain

Top Market Signals

Nanya Tech enters Vera Rubin LPDDR supply

Nanya Technology (2408 TT) has become the first Taiwanese DRAM maker to enter NVIDIA's AI server main-memory supply chain, qualifying LPDDR product for the upcoming Vera Rubin platform with direct technical assistance from TSMC.

According to Economic Daily News reporting, NVIDIA is restructuring Vera Rubin's memory architecture around LPDDR5X/LPDDR to cut total system power and shrink board area, replacing portions of traditional DDR. Research cited in the article indicates a single AI server requires LPDDR capacity dozens of times that of a smartphone, which points to structural tightening as Vera Rubin ramps.

The platform itself is not a chip refresh but a full architecture move: Vera CPU runs 88 cores and 176 threads, while Rubin GPU delivers a claimed 5x inference and 3.5x training uplift on roughly 1.6x transistor count. NVIDIA has coordinated US, Korean and now Taiwanese memory vendors to route HBM4 stacking and LPDDR integration through TSMC's CoWoS line for yield reasons. Industry sources told the paper TSMC stepped in directly to coach Nanya on wafer flatness and warpage control, the gating items that previously kept Taiwanese DRAM out of AI specs.

The financial read-through is pricing rather than volume in the first instance. Nanya's specifications and unit prices on Vera Rubin LPDDR materially exceed consumer DDR, helping group gross margin push toward 70%. The strategic read-through is supply diversification: NVIDIA now has three regions of memory supply for its highest-volume AI platform, which lowers concentration risk in any single Korean or US vendor.

Linked stocks: 2408 TT, 2330 TT, NVDA US

Sources: 台灣首家！南亞科打入 Vera Rubin 鏈 毛利率叩關七成, 南亞科多路並進 下半年有成果

MediaTek joins Google TPU 8t training silicon

MediaTek (2454 TT) is supplying I/O Die and back-end design services for Google's eighth-generation TPU 8t training chip, and supply-chain checks suggest the ASIC business is on track to become its largest single revenue source by 2027.

Per CTEE and DigiTimes, Google has split its gen-8 TPU into separate training (TPU 8t) and inference variants, with TPU mass production stepping up from H2 2026 onward. MediaTek has begun adding Final Test (FT) and System-Level Test (SLT) capacity and is pulling more CoWoS allocation, with knock-on benefits flagged for KYEC (2449 TT), Chunghwa Precision Test Tech (6510 TT) and Hong Tee. CTEE adds that Morgan Stanley has moved its Taiwan semiconductor top pick from TSMC to MediaTek, helping accelerate the rotation already evident in the share price.

The 30 April law conference is the next catalyst. The market wants quantification of how fast ASIC will overtake the smartphone SoC line, with the most aggressive sell-side voices already modeling a 2027 crossover.

Linked stocks: 2454 TT, 2330 TT, 2449 TT, 6510 TT, GOOGL US

Sources: 聯發科參與Google TPU 8t專案ASIC傳捷報 供應鏈嗨, TPU效應2027年見真章！　聯發科ASIC業務有望成最大營收來源, 一個晶片打天下退潮　Google TPU雙軌並行的含金量何在？

Foxconn locks Groq 3 LPX rack exclusivity

Foxconn (2317 TT) has secured exclusive supply of compute trays for NVIDIA's Groq 3 LPX inference rack and the bulk of full-rack assembly, with first shipments pulled into Q3 2026 ahead of prior expectations.

Economic Daily News reports that 2026/2027 LP30 plus LP35 chip shipments are now tracking 1.5 million and 2.5 million units respectively, supporting a first-batch run of around 6,000 racks in Q3 2026 and a further 10,000 racks in 2027 (excluding the next-generation LP40). On the Vera Rubin side, Foxconn shipped roughly 12,000 GB-series racks in 2025 for ~43% share, and analysts cited expect that to rise to 55-60% on the VR200 NVL72, with mass production targeted at end of Q3.

Capacity is being built to match. Chairman Liu Young-Way previously said Foxconn can produce more than 1,000 AI data center racks per week and is scaling toward 2,000 per week by year-end. The company also remains a key vendor on Google TPU servers, with analysts citing roughly 15% share into 2026 as TPU agentic-workload demand expands.

Linked stocks: 2317 TT, NVDA US, GOOGL US

Sources: 經濟日報, 鴻海 COMPUTEX 2026 秀人工智慧肌肉, 鴻海 AI PC 業務夯 出貨喊衝

SK Hynix posts ~72% op margin as Samsung strike looms

SK Hynix (000660 KS) reported a Q1 2026 operating margin of nearly 72%, a record, just as around 40,000 Samsung Electronics (005930 KS) workers prepare for an 18-day strike from May 21 that Korean media warn could shift AI memory orders toward TSMC.

DigiTimes characterizes Hynix's margin as the highest in company history and among the strongest in global manufacturing, framing the result as evidence that memory has entered a structurally high-margin phase. Liberty Times, citing Yonhap, reports the Pyeongtaek gathering on April 23 and union demands centered on bonus-cap removal, a 7% base-salary increase and a larger share of operating profit. The grievance fueling the action is the bonus gap with Hynix, with workers reportedly defecting between the two firms.

For investors, the binary is clear: an 18-day stoppage at Pyeongtaek would tighten an already strained AI memory market and pull demand toward Hynix and toward TSMC's foundry-adjacent stack. Samsung says it will keep negotiating to reach a wage agreement.

Linked stocks: 000660 KS, 005930 KS, 2330 TT

Sources: 記憶體產業暴賺時代　南韓雙雄SK海力士、三星比拚70%營益率, 三星罷工訂單恐跑掉！韓媒：台積電受益, 三星罷工恐演變為36天停產 動搖客戶信任和競爭力、供應混亂

TSMC sprints 3nm and 2nm capacity, declines High-NA EUV

TSMC (2330 TT) is raising 3nm Taiwan monthly capacity from a planned 150,000 wafers to 180,000 by year-end and pushing 2nm toward 100,000 wafers per month, while telling ASML it will not adopt High-NA EUV before 2029.

Per Economic Daily News supply-chain checks, the 3nm step-up amounts to roughly 20% above the original year-end plan, and 2nm, which began mass production in late 2025, is on a sharp ramp toward ~100,000 wafers per month by end-2026. The acceleration is feeding through to consumables vendors including 新應材, 中砂, 昇陽半, 家登 and 意德士, all of which are expanding capacity.

On lithography, TSMC global business SVP Y.P. Chiang told the company's North America Technology Symposium that there is no plan to introduce ASML's 0.55-NA High-NA EUV before 2029, citing cost. The takeaway for ASML is that High-NA push-out into the late 2020s lengthens the runway for Low-NA throughput tools and pure-play EUV optics suppliers, while DUV and supporting metrology continue carrying the 3nm/2nm ramp.

Separately, Taiwan's economy remains supply-led: AI server-driven electronic and optical product output rose 146.32% YoY in March per the Ministry of Economic Affairs, and Taipei Economic Research has lifted its 2026 GDP forecast to 7.56% from 3% previously.

Linked stocks: 2330 TT, ASML NA, 3037 TT, 6510 TT

Sources: 台積電先進製程神速擴產 可望憑藉技術優勢「贏者全拿」, 台積電向ASML天價High-NA EUV設備「說不」　三大底氣其來有自, 台積電先進製程持續放量 耗材協力廠營運起飛, 打造AI無所不在的世界　台系供應鏈暫時無可替代

Memory squeeze cracks PC demand; SSD second hike of April

Research houses now project 2026 global PC shipments could fall 10-15% as DRAM and NAND prices surge and CPU supply remains tight, while Samsung Electronics (005930 KS) and Kingston have pushed through their second SSD hike this month, raising prices 10%+.

DigiTimes reports that the previously assumed 2026 PC recovery has reversed, with multiple research firms now modeling a contraction. The exception is Apple (AAPL US), where MacBook shipments rose 21.7% YoY per Sigmaintell, supported by the unified memory architecture insulating Apple from the worst of the DRAM price move; Asustek (2357 TT) is also growing against the trend.

In storage, CTEE reports that Samsung and Kingston notified Chinese distributors of fresh price increases of at least 10% on SSD products, the second move this month, against a cumulative 2-3x band rise. The driver is structural: AI data center pull is squeezing consumer storage supply, not a short-term inventory swing. NAND Flash imbalance has shifted from cyclical to structural as cloud customers absorb capacity. The investor implication is that consumer PC OEMs face margin compression while integrated handset and notebook makers with proprietary memory strategies extend their lead.

Linked stocks: 005930 KS, AAPL US, 2357 TT, 2317 TT

Sources: PC寒冬恐現15%年出貨衰退　蘋果、華碩逆勢衝成長, AI排擠效應擴大 SSD漲勢續燒, 蘋果MacBook出貨年增21.7%　定價與記憶體策略成突圍關鍵, 記憶體封測稼動率屢高不墜　載板漲價2Q26不排除反映成本

Intel surges on AI data center beat, with caveats

Intel (INTC US) shares jumped 23.6% to $82.54 after Q1 results led by AI-driven data center revenue, drawing street-high price targets and an HSBC reset, while reports surfaced that Intel boosted yields by selling lower-quality "scrap" CPUs into a tight market.

Coverage from TipRanks, IndexBox and Tom's Hardware describes data-center strength as the centerpiece of the print, with multiple top analysts raising targets after the report. HSBC reset its Intel target for the rest of 2026, and Tom's Hardware separately reports that Intel's investor relations team confirmed customers are now accepting CPUs that would normally be classified as scrap or low-expectation, a tangible revenue tailwind given the current CPU shortage. DigiTimes notes that Lenovo (992 HK) is now flagging server lead times longer than six months as AI infrastructure pulls high-end Intel CPU supply tight.

The skeptical bench is also vocal. Seeking Alpha pieces this weekend frame the move as the AI narrative running ahead of fundamentals, and one Robert Castellano note explicitly warns to "pump the brakes" on the AI-foundry story. Jim Cramer credited CEO Lip-Bu Tan, but the more durable signal for portfolios is that AI capex is now creating spillover demand for second-tier CPU silicon, not just GPU and HBM. That widens the beneficiary set across the x86 ecosystem.

Linked stocks: INTC US, 992 HK, NVDA US

Sources: Google News, Tom's Hardware, Google News, 聯想部分伺服器產品交期逾半年　籲客戶提早下單鎖定供應, Intel: Pump The Brakes (NASDAQ:INTC)

先聲 First Word — Exclusives from Chinese-Language Sources

Foxconn Industrial Internet overtakes Huawei as Shenzhen's top manufacturer

FII, Foxconn's mainland industrial unit, posted RMB 902.8 billion in 2025 revenue, +48.2% YoY, surpassing Huawei and BYD to become the largest of Shenzhen's "Hua-Fu-Bi" trio for the first time. Confirms how AI server orders are concentrating in Foxconn's mainland operating entity. (27 Apr 2026) Source: 工業富聯營收首度超越華為 躍上深圳製造業一哥

Foxconn and Mitsubishi Electric sign MOU on automotive components

Foxconn (2317 TT) and Mitsubishi Electric have signed an MOU to evaluate joint operation of vehicle components, marking another step in Foxconn's auto-supply diversification beyond contract manufacturing. Adds a Japanese leg to Foxconn's automotive partner network. (27 Apr 2026) Source: 鴻海、三菱電機簽MOU 評估共同營運車用零組件

NTT-NEC alliance frays as Japan's "Denden Family" reshuffles

NTT is increasing overseas equipment purchases while NEC is scaling back its base-station business, signaling a crack in the post-Huawei domestic-supply alliance Tokyo had been pushing. Implies Japan's intent to rebuild a closed comms supply chain is not surviving commercial reality. (27 Apr 2026) Source: 日本通訊業重組生變：NTT增購海外設備、NEC縮減基地台業務

TI VP says 800V data-center power adoption "cannot wait", GaN volumes inevitable

Texas Instruments (TXN US) high-voltage power VP Kannan tells DigiTimes that 800V deployment in AI data centers is unavoidable and GaN volume ramp is now structural, ahead of TI's COMPUTEX 2026 showcase. Reinforces the day's signal that power architecture is the new AI bottleneck. (27 Apr 2026) Source: DigiTimes

Co-Tech (金居) flagged as HVLP4 copper-foil winner

Guangfa Securities argues copper foil will follow the glass-fabric playbook of high-end shortage driving price increases, naming Co-Tech as the HVLP4 ramp beneficiary at ~14x 2027E P/E. Adds copper foil to the AI BOM tightness list alongside ABF substrate and glass cloth. (27 Apr 2026) Source: 銅箔也喊漲…HVLP4放量 金居升勢壯