Zero One Daily Intelligence Brief — May 15, 2026

TSMC sees US$1.5tn 2030 chip market as Jensen lands in Beijing empty-handed

Top Market Signals

TSMC unveils US$1.5tn 2030 chip market and COUPE roadmap

TSMC (2330 TT, TSM US) used the Hsinchu leg of its 2026 Technology Symposium on 14 May to forecast a US$1.5tn global semiconductor market by 2030, with HPC and AI applications taking 55% of that pool, and to put silicon photonics on the same strategic tier as advanced packaging.

According to UDN reporting on the symposium and TechNews coverage of senior vice president Y.C. Chang, TSMC framed CoWoS as the platform that solves compute density (GPU and HBM stacked on a 2.5D interposer) and COUPE (Compact Universal Photonic Engine) as the platform that solves the data-movement bottleneck inside an AI data centre. Chang told the audience that COUPE will "become as well known as CoWoS".

The COUPE roadmap is what to anchor on. The first stage, on-PCB pluggable optics, is the 2025 baseline. From 2H26, TSMC plans to move the optical engine onto the substrate, which on its own delivers roughly 4x the power efficiency of traditional copper and cuts transmission latency by about 90%, per data TSMC presented. Stage three sits the COUPE engine directly on the CoWoS interposer, with TSMC modelling 10x power efficiency and 95% latency reduction. The first 200Gbps micro-ring modulator built on COUPE is targeted for production this year, with TSMC pointing at 4Tbps/mm bandwidth density by 2030.

Customer pickup is already wide. NVIDIA (NVDA US) has been integrating COUPE into its Spectrum-X switch line, Broadcom (AVGO US) into its Tomahawk 6-Davisson generation, AMD (AMD US) is in deep evaluation, and silicon-photonics startup Lightmatter has multiple dies through TSMC tape-out. The same Hsinchu session also disclosed that the largest CoWoS reticle size will reach 14x the standard photomask by 2028, framing the capacity envelope behind the 2030 forecast.

The demand-side validation arrived the same week from Applied Materials (AMAT US), which raised its FY26 semi-equipment revenue growth target to above 30% (from above 20%) and its advanced-packaging revenue growth to 50% YoY, with CEO Gary Dickerson telling the call that customers are now sharing eight-quarter rolling demand forecasts and that AMAT is tracking more than 100 fab construction projects (10+ added last quarter).

Linked stocks: 2330 TT, AMAT US, AVGO US, NVDA US

Sources: 台積電估2030年半導體產值1.5兆美元　COUPE成關鍵字, TechNews, 台積電CoWoS良率飆逾98%、產能急擴　14倍光罩尺寸2028量產, 應材業績、財測優於預期，看旺今年半導體設備市場

Jensen lands in Beijing on Trump's plane; H200 still has zero China deals

NVIDIA (NVDA US) CEO Jensen Huang joined US president Donald Trump's China delegation at the last minute, and the H200 export-licence path that markets read as a positive surprise has produced no purchases yet from the Chinese buyers it covers.

According to TechNews, white house officials originally circulated a CEO list on 11 May that did not include Huang. He boarded Air Force One in Anchorage on 12 May after Trump personally called and asked him to come. The travelling group also included Tesla's Elon Musk, Apple's Tim Cook, BlackRock's Larry Fink and Boeing's Kelly Ortberg, with 15 to 17 US executives present in the Trump-Xi sessions per US trade representative Jamieson Greer's Bloomberg TV interview on 15 May.

The substance the market wanted is not what showed up. Greer told Bloomberg that "we did not discuss chip export controls in the meeting" and that the topic was not a main part of the China agenda. Reuters reported that the US commerce department has approved roughly ten Chinese companies (Alibaba, Tencent, ByteDance, JD.com among them) to buy NVIDIA H200, with Lenovo and Foxconn (2317 TT) cleared as distributors, but per DigiTimes none of those approved buyers have actually placed orders. Asked directly on 14 May whether NVIDIA would sell to Huawei, Huang paused for three seconds and answered "what a strange question" without committing either way, per Liberty Times.

Markets read the trip as bullish anyway. NVIDIA closed at US$235.74 on 14 May, an all-time high giving roughly US$5.7tn in market cap, on the H200-approval headline. TSMC ADR rose 4.48% to US$417.72 on the same news, and Korea's KOSPI broke 8,000 for the first time on AI-semi strength.

Two countercurrents to keep on the radar. Activist short Culper Research published a 40-page report titled "The China Problem" on 15 May claiming that more than 20% of NVIDIA's FY26 compute revenue is still tied to China through Singapore (Megaspeed, Apex), Malaysia (Speedmatrix) and Taiwan (Giga Computing) intermediaries linked to Alibaba, and predicting Beijing will move to "exclude NVIDIA" in CY26. Wells Fargo, by contrast, has modelled US$25bn a year of incremental China revenue. Anthropic separately used the Trump-China optics to call publicly for tougher US chip controls, per Nikkei Asia.

Linked stocks: NVDA US, 2330 TT, 2317 TT, AVGO US

Sources: H200 銷中長路漫漫？葛里爾：晶片出口非談判重點, 傳美國核准H200銷售卻零成交　黃仁勳赴中國尋突破, 最後一刻加入 CEO 隨行團，黃仁勳：川普總統要求我來, 點名台灣公司涉轉運！放空機構40頁報告狙擊輝達 直指有「中國問題」

Foxconn 1Q26 record profit; 2026 capex +30%, NT$30bn Kaohsiung supercomputing

Foxconn (2317 TT) posted 1Q26 net profit of NT$49.9bn, a record for the same quarter, with gross margin lifting to 6.18% and operating margin to 3.57%; management guided 2Q26 revenue +4% to +15% QoQ and above +15% YoY, and announced a NT$30bn capital injection into a Kaohsiung Asia Bay supercomputing build that begins phase one in 3Q26.

Per Liberty Times coverage of the Hon Hai analyst meeting on 14 May, rotating CEO Cheng Chi-heng described AI servers, switches and AI rack demand as "continuing to heat up" and reiterated a full-year stance of "strong growth, cautiously optimistic". AI servers already accounted for over 50% of cloud-network product revenue in 1Q26. Management expects full-year AI rack volumes to roughly double from 2025, and 800G-and-above switch shipments and revenue could double too. New-generation GPU AI rack shipments begin meaningful ramp in 3Q26.

The capex number reframes the multi-year investment picture. After lifting 2025 capex by about 27% YoY, Foxconn now expects another roughly 30% YoY increase in 2026, focused on regional manufacturing, automation and AI core capacity. The Americas (US plus Mexico) is now Foxconn's largest AI-server production base, with Taiwan staying as the high-end R&D, qualification and ramp centre.

The Asia Bay supercomputing announcement is a separate signal worth flagging. Per cnYES, Foxconn has injected NT$30bn into the Kaohsiung-based facility, which it intends to operate as Taiwan's largest AI supercomputing centre with phase-one focus on AI inference and generative AI. The cluster is positioned as enterprise-grade Taiwan-sovereign compute, and pairs with Foxconn's earlier announcement of strategic partnerships with Amini and Bull to address sovereign-AI infrastructure demand in Africa and the Global South.

Our 6 May 2026 Blade flagged Hon Hai's April monthly revenue print and management's "Q2 not the slow season" framing; today's 1Q26 net-profit record and Q2 +>15% YoY guidance turn that posture into a number. Morgan Stanley raised its Foxconn target to NT$310 (from NT$290) on the print, citing operating leverage and 2026-2028 EPS upgrades; the stock briefly touched NT$260, a roughly seven-month high, before settling at NT$254 (+3.89%) by 9:45 a.m.

Linked stocks: 2317 TT, 2330 TT, NVDA US, 2382 TT

Sources: 鴻海增資亞灣超算300億元 首階段Q3啟用衝刺本土主權AI, 焦點股》鴻海：法說會釋樂觀展望 帶量大漲, 看好 AI 伺服器與 800G 交換器爆發成長，大摩升鴻海目標價至 310 元

TSMC's smart-glasses call electrifies Taiwan optics

TSMC vice president Y.C. Chang named smart glasses as "the most promising" of the next-generation AI device categories during the 14 May Hsinchu symposium, and Taiwan optics names took the cue immediately on 15 May.

Per Economic Daily News, Chang argued that human-to-environment interaction is most efficient through vision, and that smart glasses can carry data-centre intelligence directly to the user's eye in real time. Chang positioned smart glasses, autonomous vehicles and humanoid robots as the three end-user surfaces that "physical AI" will land on, and explicitly singled out smart glasses as the highest-potential of the three.

Largan Precision (3008 TT) opened at NT$3,290 and went limit-up at NT$3,445 within the morning session. Ability Enterprise (英濟, 6166 TT) opened at NT$36.4 and reached limit-up at NT$36.5. Young Optics (揚明光, 3504 TT) briefly touched limit-up before settling at +7%, Altek (華晶科, 3059 TT) was up over 7%, and Genius Electronic Optical (玉晶光, 3406 TT) traded up over 5%.

A second wave hit later in the morning around the robotics names adjacent to the same comments. Per Economic Daily News, Solomon Tech (所羅門, 2359 TT), Mirle Automation (盟立, 2464 TT), Roselin (羅昇), and TBI Motion (全球傳動) hit limit-up, with Techman Robot (達明), Altek and Hiwin Mikrosystem (大銀微系統) up over 8%. Solomon is named as the AI-vision partner for Unitree's mainland-China humanoid platform; Altek and Canon Taiwan (佳能精密) are also positioned as machine-vision suppliers.

The read-through is that Taiwan's optics chain is being repriced as the camera-and-lens layer of a coming smart-glasses and humanoid-robotics build, on top of its existing smartphone-camera base. Largan, the longstanding flagship, is again the most-volume name in the lift.

Linked stocks: 3008 TT, 6166 TT, 3406 TT, 3059 TT

Sources: 台積電喊智慧眼鏡最具潛力 大立光、英濟飆漲停 這幾檔也強漲, 機器人最重要的是視覺系統？概念股大噴發 所羅門等攻漲停

Samsung's offer rejected; Seoul floats emergency adjustment right

Samsung Electronics (005930 KS) management sent the union an "unconditional" proposal to restart talks at 10 a.m. on 15 May, the union rejected it, and Korea's industry minister has now publicly floated the country's "emergency adjustment right" as a possible last-resort intervention before the planned 21 May to 7 June general strike.

According to TechNews, Samsung had offered the union a choice between calculating the operating performance incentive (OPI) bonus pool as 10% of operating profit or 20% of economic value added (EVA), and to layer an "uncapped special compensation system" on top of the existing OPI to address the union's "remove the cap" demand. Samsung Electronics super-enterprise union chair Choi Seung-ho replied that the proposal "does not feel like it was sent to us" and that talks could resume "after June" only, leaving the planned 18-day stoppage on schedule.

The macro framing has changed materially in the past 24 hours. Korean industry minister Kim Jung-kwan said on 14 May that the strike "must be stopped at all costs" and that if it goes ahead the government "will inevitably activate emergency adjustment". Per TechNews, that authority normally sits with Korea's labour minister, so an industry minister raising it directly is a first and signals how seriously Seoul is treating the AI-memory supply hit. DigiTimes coverage estimates the daily output cost at roughly KRW1tn (~US$680m) and the 18-day exposure at the equivalent of roughly 4% of global DRAM supply.

The supply chain is already pre-positioning. Per DigiTimes, Samsung memory-line management has committed to keeping wafer fabs at minimum operating mode through any stoppage, but spot prices have started moving in advance of the strike, and Apple and HP are reported to have called Samsung directly to ask about supply-continuity contingencies. Customers asking that question this far ahead is itself a signal.

The Samsung Pyeongtaek strike has been on the AI-memory watch list for two weeks; today's exchange of an "unconditional" management offer and a "talk after June" union reply makes the next 24 hours look more decisive than another sequence of pre-strike negotiation. The most-asked question among AI-memory buyers is now where SK Hynix (000660 KS) and Micron (MU US) capacity goes if Samsung's HBM and DRAM lines run any meaningful idle hours.

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

Sources: 麻煩大了！三星提無條件再次協商，工會代表：罷工後再談！, 三星半導體大罷工倒數　最終手段「緊急調整權」浮上檯面, 三星罷工對峙升溫　記憶體產線力守運作、現貨價格提前躁動, 三星總罷工背後危機　良率下滑、客戶出走「隱形技術債」

NVIDIA Vera Rubin design fixed; supply chain points to 3Q26 ramp

NVIDIA's (NVDA US) next-generation Vera Rubin platform has cleared the thermal-architecture redesign overhang that had been weighing on cooling and assembly suppliers, and the supply chain now expects shipments to ramp quarterly from 3Q26.

According to DigiTimes, recent market reporting that Vera Rubin shipment timing was at risk because of cooling-architecture changes had spilled into share-price moves at thermal and ODM names. Supply-chain sources now say the design adjustments are "largely resolved" and that NVIDIA has worked through the changes with its ODMs and key suppliers. The implied schedule lines up volumes building quarter-by-quarter through 2H26.

This matters most for the Taiwan thermal and rack chain. Asia Vital Components (3017 TT), Auras Technology (3324 TT) and Sunon (3552 TT) are the companies with the most thermal exposure to the redesign. The same Hsinchu TSMC symposium that put COUPE on the strategic track named Vera Rubin's interposer and packaging path, so the photonics layer and the power-and-thermal layer are now both lined up for the same ramp window.

The cross-event read goes back to our 29 April 2026 Blade, which framed Vera Rubin's outsized memory appetite as forcing handset-chip designers into LPDDR allocation panic. Today's signal is the supply-side counterpart: the platform's mechanical and power architecture is now finalised, and the squeeze on memory and on advanced packaging that was already visible in handset prices and wafer-foundry guidance now has a calendar attached.

A second reading is for Quanta Computer (2382 TT), whose 1Q26 gross-margin compression to 4.78% (down 1.54 ppt QoQ, down 3.14 ppt YoY) was largely attributed to AI-server product mix on the analyst call. Quanta also flagged that ASIC contribution accelerates from 2H26 and that some AI-server memory contracts will shift to consignment accounting from H2 to clean up the gross-margin print. Both align with NVIDIA's now-firmed Vera Rubin schedule.

Linked stocks: NVDA US, 2330 TT, 2382 TT, 3017 TT

Sources: NVIDIA Vera Rubin設計問題排除　供應鏈指3Q26逐季放量, 〈焦點股〉廣達首季毛利率遜於市場預期 跌逾半根停板力守季線之上

先聲 First Word — Exclusives from Chinese-Language Sources

Phison runs a 20B-parameter LLM on MediaTek's Dimensity 9500 with aiDAPTIV

Phison (8299 TT) and MediaTek (2454 TT) demonstrated at the MediaTek Developer Conference 2026 that Phison's aiDAPTIV solution running on the Dimensity 9500 mobile platform can execute a 20-billion-parameter large language model on a single device, a step-change for on-device generative AI inference. Phison frames it as the bridge from cloud-only generative AI to mobile-edge AI on flagship Android. (15 May 2026)
Source: 群聯aiDAPTIV搭上聯發科天璣9500平台　加速邊緣AI落地

NXP and Quanta build SDV regional network for next-gen vehicles

NXP Semiconductors (NXPI US) and Quanta Computer (2382 TT) announced a joint software-defined-vehicle (SDV) reference design built on NXP's S32 processor family, targeting end-to-end real-time communication inside the next-generation electrical-electronic vehicle architecture. The deal pulls Quanta beyond its AI-server franchise into Tier-1 auto electronics, on the same template that took it from notebook ODM into hyperscaler servers a decade ago. (14 May 2026)
Source: 恩智浦聯手廣達打造SDV區域網路方案　加速實踐E2E實時通訊

Shunsin's fifth straight limit-up tracks TSMC's COUPE-CPO disclosure

Shunsin Technology (6451 TT) hit limit-up at NT$577 in late-morning trade on 15 May, +9.9% on the day, with cumulative gain of 30.83% over five sessions; the move came on TSMC's COUPE-CPO disclosure during the Hsinchu symposium and Shunsin's own April revenue rebound to NT$641m (+11.8% MoM) on rising high-speed optical-transceiver-module shipments. The fundamental backdrop is still soft (1Q26 EPS NT$-1.45, four-month revenue down 16.76% YoY), so the move is theme-driven rather than earnings-led. (15 May 2026)
Source: 焦點股》訊芯-KY：CPO利多題材 亮燈漲停

Unimicron 1Q26 above expectations on ABF cycle and AI ASIC demand

Unimicron (3037 TT) reported 1Q26 EPS of NT$3.20 with gross margin 18% on ABF substrate revenue +14% QoQ; April revenue was NT$13.93bn (+6.5% MoM, +27.6% YoY). Analysts now model 2Q26 revenue +12.3% QoQ and gross margin lifting to 21.1% (+3.1 ppt QoQ), and lift Unimicron's price target by more than 60% as the ABF chain runs into pricing pass-through and the high-end-share mix tilts on AI ASIC demand. Capacity adds at Guangfu Phase 1 third tranche (2H26), Guangfu Phase 2 (2H27) and Yangmei Phase 2 (2028) point to high-end capacity doubling between 2025 and 2028, with key customers signing "guarantee volume not price" long-term contracts. (15 May 2026)
Source: 高階ABF載板需求強勁 法人調高欣興目標價逾六成