Zero One Daily Intelligence Brief — April 15, 2026

Bloomberg pegs AI capex at $1T as Taiwan prices rise and Nvidia's Rubin slips

Top Market Signals

Meta locks Broadcom MTIA to 2029 with first 2nm custom AI silicon

Meta extended its custom AI chip partnership with Broadcom through 2029, initially committing to deploy up to 1GW of MTIA silicon built on Broadcom's training and inference technology and scaling to multiple GW in follow-on generations. The announcement cements Broadcom's position as the leading custom-silicon partner for hyperscalers, coming two weeks after it confirmed a long-term Google TPU agreement that includes up to 3.5GW of capacity accessible to Anthropic. Broadcom CEO Hock Tan, who joined Meta's board in 2024, will not stand for re-election, and former Estée Lauder CFO Tracey Travis is also exiting.

The MTIA chips will be the first custom AI silicon built on 2nm process technology, placing Meta on the same leading-edge node as Nvidia's next-generation parts and anchoring additional demand for TSMC's most advanced capacity. Meta CEO Mark Zuckerberg framed the collaboration as spanning chip design, packaging, and networking, aligned with the company's pledge to spend up to $135 billion on AI in 2026. Meta's broader buildout now includes roughly 6GW of AMD GPU deployment, millions of Nvidia chips, a new Arm-designed custom chip, and 31 data centers — 27 of which are in the United States.

The news is a marker for the broader hyperscaler ASIC cycle. Unlike Google and Amazon, whose custom chips are integrated into their cloud platforms and rented to external customers, Meta's MTIA silicon is exclusively for internal use, meaning the 1GW+ commitment sits on top of — not alongside — Meta's external-facing AI investments. Tan also signaled that OpenAI's first custom chip, also being built with Broadcom, is scheduled to come online in 2027.

Linked stocks: AVGO US, META US, 2330 TT Sources: 博通與 Meta 合作協議延長至 2029 年, Google News

SK Hynix evaluates 20–30% HBM4 cut to Nvidia as Rubin production slips

SK Hynix is evaluating a reduction of roughly 20% to 30% in 2026 HBM4 shipments to Nvidia versus its original plan, according to reporting from DigiTimes and Korea's ZDNet. The Korean memory maker is Nvidia's primary HBM supplier and was expected to be the dominant provider of the sixth-generation high-bandwidth memory. A cut of this magnitude at the anchor customer would reshape 2026 HBM allocation decisions across Micron and Samsung and could extend the commercial life of HBM3e by two to three quarters.

The move reflects the Rubin ramp bottleneck that has been building in public reporting over the past several weeks. Nvidia revised HBM4 per-pin speed requirements upward to above 11 Gbps in Q3 2025, forcing all three HBM suppliers to modify designs and pushing HBM4 mass production into late Q1 2026 (Tom's Hardware, I-Connect007). BigGo Finance reported Nvidia subsequently lowered its 2026 Rubin GPU production target from 2 million to 1.5 million units, with Rubin now expected to account for 22% of Nvidia's 2026 high-end GPU mix, down from a prior 29% forecast. The Register on April 8 flagged additional friction around the ConnectX-8 to ConnectX-9 NIC migration, rising power consumption, and more advanced liquid-cooling requirements at the system level.

Implications run both ways. On the bearish side, a further Rubin slip pushes out the biggest volume catalyst for 2026 HBM capacity. On the bullish side, if HBM4 capacity is cut but underlying AI demand remains unchanged, that capacity either serves existing HBM3e customers or is reallocated to other hyperscaler custom programs — Google's TPU and Meta's MTIA being the obvious candidates. Today's Nvidia-SiFive investment and NVLink Fusion expansion to RISC-V (Signal 5) reinforce the broader theme that architectural diversification in AI compute is accelerating regardless of Rubin's ramp timing.

Taiwan's memory chain remains tight regardless — see today's First Word item on Q1 slow-season revenue records across the group.

Linked stocks: 000660 KS, NVDA US, MU US, 2408 TT Sources: NVIDIA Rubin傳量產受阻　SK海力士擬下修HBM4出貨量約2成, Theregister, Biggo, Tom's Hardware

Taiwan component suppliers kick off broad Q2 price hikes in CCL, passives, and ABF

Taiwan's leading CCL suppliers (Elite Material, Taiyao Technology, Iteq), passive-component makers, and ABF substrate suppliers are all raising prices in Q2, with upstream glass-fiber and copper-foil cost pressure driving a multi-quarter cycle that DigiTimes projects will continue through the end of 2026. Taiyo Yuden notified distributors that MLCC, inductor, ferrite bead, RF inductor, SAW/FBAR, and aluminum capacitor prices will rise from May 1, following Murata's earlier April hike — a sequence that typically benefits Yageo (2327 TT) and Walsin Technology (2492 TT) in the MLCC peer group.

On the PCB-upstream side, glass-fiber prices rose roughly 30% in 2025 and are expected to more than double in 2026, with high-end Low Dk, Low Dk2, and T-glass grades seeing 20% to 30% annual increases at Nittobo, Asahi Kasei, Taiwan Glass, and Taishan. HVLP copper-foil supply-demand gaps are projected at 48% in 2026 and 43% in 2027, prompting CCL producers to raise process charges. M6, M7, and M8 CCL grades are all seeing double-digit percentage hikes per round. China's Kingboard already raised board and PP prices 10% on April 3.

Further downstream, BofA raised price targets on ABF substrate names Nan Ya PCB (8046 TT), Unimicron (3037 TT), and Kinsus (3189 TT) to near NT$1,000 following its projection of a 13% ABF supply gap. The piece of this chain to watch most closely is glass fiber — if E-glass doubles, CCL gross margins hold only if the trio passes through pricing at the same cadence, which is what customer letters in the past week suggest.

Linked stocks: 2383 TT, 6213 TT, 2327 TT, 2492 TT, 3037 TT, 8046 TT, 3189 TT Sources: 玻纖布、銅箔成本接連上揚　2026全年CCL漲勢「未完待續」, 太陽誘電 加入被動元件漲價浪潮, ABF缺口擴大…南電、欣興 外資喊準千金

Bloomberg pegs next-year AI infrastructure capex at $1T, Taiwan ODMs running full orderbooks

Bloomberg's latest aggregate estimate for hyperscaler AI infrastructure capex in the next twelve months is approximately US$1 trillion (NT$32 trillion), with Amazon, Microsoft, and Meta leading the charge. Taiwan's AI-server assembly chain is the most direct downstream beneficiary, with Hon Hai (2317 TT), Quanta Computer (2382 TT), Wistron (3231 TT), Inventec (2356 TT), and Wiwynn (6669 TT) all running at full orderbook visibility into H2 2026. Component suppliers feeding AI server BOMs — thermal (Auras 3324 TT, Asia Vital Components 3017 TT), power (Delta 2308 TT, Lite-On 2301 TT), and cable assemblies (Bizlink 3665 TT) — are also all seeing accelerated pull-through.

Wiwynn (6669 TT) specifically called out strong sustained pull from core customers Meta and Microsoft in recent commentary, with analysts projecting full-year 2026 profit at a new record. Chenbro Micom (8210 TT), which rose 6.7% on April 14 to NT$961, continues to benefit from Nvidia HGX/GB platform ramps and ASIC-customer generation changes.

This signal reinforces our March Sales Signals data — the ODM/EMS ex-Hon Hai group posted +57% aggregate revenue growth YoY in March, with Wistron alone up 118% and Quanta up 88%. Orderbook visibility is consistent with that trajectory extending through Q2.

Linked stocks: 2317 TT, 2382 TT, 3231 TT, 2356 TT, 6669 TT, 8210 TT, 2308 TT, 3017 TT, 3324 TT, 2301 TT, 3665 TT Sources: 亞馬遜、微軟等 AI 投資大潮總金額上看32兆元 帶旺緯創、英業達等台鏈, 雲端大咖啟動 AI 大投資 關鍵元件廠全速喊衝, 緯穎 挑價內外20%

Nvidia takes SiFive stake and opens NVLink Fusion to RISC-V

Nvidia has invested in SiFive, the RISC-V processor IP company founded in 2015 by the three inventors of the RISC-V architecture, and opened NVLink Fusion to SiFive designs — giving RISC-V a first-class path into Nvidia-centric AI compute infrastructure. The move follows Arm's announcement that it will enter the AGI CPU design business directly, and together the two developments signal a strategic rebalancing of CPU options in the AI era. Rather than fostering a competitor to Arm, the SiFive alignment positions Nvidia to benefit from any architecture that ends up in AI host CPU sockets, while giving RISC-V a credible path to the performance-per-watt scale that AI workloads now require.

For Taiwan, the implication is that design-service and IP names with RISC-V exposure — Andes Technology (6533 TT) in particular — gain optionality as NVLink Fusion adoption broadens. For Arm, the signal is a reminder that the Nvidia relationship is strategic but not exclusive.

Linked stocks: NVDA US, ARM US, 6533 TT Source: DigiTimes

先聲 First Word — Exclusives from Chinese-Language Sources

Taiwan memory chain sets Q1 slow-season revenue records

DigiTimes reports Taiwan memory supply chain companies have set Q1 2026 revenue records across the board, with year-on-year growth rates broadly doubling or more despite what has historically been a seasonal slow quarter. Nanya Technology (2408 TT) already disclosed a sharp Q1 profit jump; broader memory peers are expected to follow when monthly and quarterly results roll in. The inflection reflects AI-driven memory scarcity rather than a normal cyclical rebound, and the absence of a Q1 dip this year is itself the signal. (April 15) Source: 記憶體1Q26淡季創最強紀錄　營收飆漲推動獲利驚喜連連

Taiwan equipment makers pivot to silicon photonics and panel-level packaging on Nvidia Rubin ramp

DigiTimes reports Taiwan equipment suppliers historically well behind international peers are now securing design-ins across TSMC and Nvidia's Rubin Ultra platform, with silicon photonics (SiPh) co-packaged optics (CPO) and panel-level advanced packaging (PLP) as the two main transformation axes. Scientech (3583 TT) CEO Yi Jin-liang disclosed the company has shipped the world's first 310×310mm panel-level PVD tool, targeting ASE CoPos advanced packaging. ASE Technology Holding (3711 TT) is leading the transition from wafer-based to panel-based advanced packaging for AI chips. (April 15) Source: 台系設備廠轉型兩大方向明確　先進封裝與矽光子列車猛衝

Tong Hsing guides AI-driven fiber-optic modules to outgrow satellite business in 2026

At its April 14 earnings call, Tong Hsing Electronic (6271 TT) guided fiber-optic module revenue will be the company's fastest-growing segment in 2026, driven by AI communications demand, and is expected to surpass the company's satellite business by year-end. RF products, ceramic substrates, and automotive/mobile CIS sensors are also contributing. Tong Hsing expects Q2 2026 revenue to grow by a high-single-digit percentage quarter-on-quarter and full-year revenue to exceed prior expectations. Management also announced a new Philippines plant to support automotive-customer ramp. (April 14) Source: AI通訊挹注光纖模組成長顯著　同欣電估2026年比重超越衛星