Silicon Nexus

ASRock Rack (Yongqing Electronics) announced a new generation of AI infrastructure platforms featuring NVIDIA's upcoming Vera CPU at COMPUTEX 2026 in Taipei. The launch positions ASRock Rack as an early ODM partner for NVIDIA's next-gen Vera Rubin platform, reinforcing Taiwan's central role in AI server supply chains.

Why it matters: COMPUTEX product launch tied to NVIDIA's next-gen Vera CPU signals continued Taiwan ODM leverage in AI server buildouts, though ASRock Rack itself is not in the tracked universe.

Marvell confirmed it is engaging TSMC to use the A14 process (mass production targeted for 2028) for next-generation AI data center connectivity chips, deepening its single-foundry bet on TSMC. Marvell already runs 1.6Tb DSPs on TSMC's 3nm and is the first to use 2nm for DSP/DCI modules; it has committed up to $1B in prepayments to secure capacity, while Nvidia has invested $2B to deepen collaboration.

Why it matters: Roadmap/supply-chain story confirming TSMC's leading-edge customer pipeline through A14 (2028); supportive for TSMC but not an immediate stock-moving event.

Qiushi (CCP central organ) and CCID jointly published a rare 6,700-character report admitting China's AI industry, now at ~1.2 trillion RMB (~$165B) core scale, suffers from four structural gaps: chip/CUDA ecosystem lock-in, weak vertical accuracy, scarce Chinese-language training data (just 1.3% of global 5B-scale corpora), and broken monetization across compute-model-application layers. The candid self-critique signals Beijing will steer policy and capital toward industrial AI QC, AI medical imaging, AI financial risk, data labeling, and synthetic data — while acknowledging NVIDIA CUDA migration costs remain prohibitive despite Huawei Ascend reaching 2nm.

Why it matters: Sector-level policy signal on China AI structural gaps and CUDA dependence — relevant to NVIDIA/TSMC supply chain and HBM demand context, but no specific contract, capex, or earnings event for tracked Taiwan/Korea names.

Taiwan analytical-services firm MA-tek (3587-TW) will complete its first silicon photonics wafer and chip optoelectronic test platform by August 2026, with second and third tools to follow by end-2026 and early 2027, and a CPO module test platform by Q2 2027. The company says its top-10 silicon photonics customers contributed roughly NT$191M (~US$6M) in 2025, and global SiPh customer coverage already exceeds 50%, positioning MA-tek as a key analytics supplier to the AI-driven silicon photonics ramp.

Why it matters: Capex/roadmap disclosure for a non-tracked Taiwan analytics vendor (3587) with no direct revenue or contract impact on tracked tickers, though it signals continued silicon photonics/CPO ramp relevant to the AI infra supply chain.

FactSet's latest survey of 16 analysts raised the 2026 EPS consensus for Silergy (6415-TW) from NT$10.2 to NT$10.27, with a high of NT$12.31 and a low of NT$8.5. The consensus price target stands at NT$600, a mild positive revision for the analog power IC designer.

Why it matters: Mild consensus EPS revision and reiterated target price — directional but not a stock-moving catalyst on its own.

FactSet's latest poll of 15 analysts lifted Novatek's 2026 EPS consensus median to NT$30.07 from NT$29.64 (range NT$24.21–34.61), with an average target price of NT$470. The upward revision signals improving display-driver IC earnings visibility heading into 2026.

Why it matters: Sell-side consensus revision on a single tracked name (Novatek) — meaningful for the stock but not a hard catalyst like a contract, capex, or earnings release.

Intel CEO Tan Lip-Bu told a podcast that Intel will fabricate chips for Elon Musk's 'Terafab' project on its 14A node to feed Tesla EV, robotics and space data-center demand, and cited TSMC's use of NVIDIA AI in fab operations. He flagged four AI supply-chain bottlenecks — power, helium, severe memory shortages and CPU/GPU tightness — and warned chip prices will rise and be passed to customers, with new fab capacity taking years to come online.

Why it matters: CEO commentary flags memory shortage and pricing power as the dominant near-term theme — directly supportive of HBM/DRAM names — and confirms TSMC 14A engagement on Terafab, but it's interview color rather than a formal contract or capex announcement.

A Taiwan brokerage note frames TSMC's shift from CoWoS to CoPoS (panel-level) and glass core substrates as a 5-year capex super-cycle peaking 2028-2032, with material utilization rising from 50-65% to 75-95% and AI chip density per panel jumping from ~4 to 9-16 units. Named Taiwan equipment beneficiaries include Sunyne (3583), Gallant Precision (3131), and All Ring Tech (6187) as POR candidates, with ABF substrate makers Unimicron (3037), Nan Ya PCB (8046), Kinsus (3189) and panel makers Innolux (3481), AUO (2409) framed as coexistence (not replacement) winners. Note is commentary/promotional from a Taiwan investment advisory, not a confirmed order or capex announcement.

Why it matters: Sector/roadmap commentary on TSMC's advanced packaging transition with named beneficiaries, but no confirmed orders, capex figures, or earnings catalysts — and the source is a promotional advisory note rather than a primary disclosure.

UMC (2303) is partnering with Intel to jointly develop 12nm and 3nm processes at Intel's Arizona Fab 52, marking UMC's return to advanced nodes without heavy capex. The 12nm PDK is due in 2026 with mass production targeted for end-2027, while the 3nm node aims to directly challenge TSMC (2330) in leading-edge foundry.

Why it matters: Named strategic alliance with concrete roadmap (12nm PDK 2026, 3nm targeting TSMC parity) materially reshapes UMC's business model and the foundry competitive landscape.