AI infrastructure and semiconductor supply chains led reader interest this week, from Taiwan's manufacturing recovery and Nexperia's legal battle to Sandisk's HBF, Micron's HBM expansion, and Marvell's silicon photonics roadmap. Samsung also refreshed its smartphone lineup with the Galaxy S26 FE. Below are the most-read DIGITIMES stories from the week of August 31-September 06, 2026.
Taiwan manufacturing returns to prosperity territory
Taiwan's manufacturing business-climate indicator moved into the red prosperity category in July, according to the Taiwan Institute of Economic Research, helped by stronger optimism among local manufacturers. Yet year-over-year export growth to the US slowed sharply to 28.0% from 267.6% in January, the lowest since February 2025. TIER said AI, semiconductors, and broader global demand still supported activity, though weak domestic demand in China remained a drag and the US-bound slowdown bears watching.
Nexperia-Wingtech dispute deepens in China
A Chinese court ordered more than CNY2.14 billion (≈ US$318 million) in domestic assets belonging to Nexperia and its equipment arm frozen, escalating the legal fight with parent Wingtech Technology. The order covers stakes in several Nexperia China entities and may remain in force until August 2029. The dispute follows Dutch and US regulatory pressure, Chinese export controls, and a governance crisis that disrupted supply chains and contributed to a 57.54% drop in Wingtech's 2025 revenue.
Sandisk pitches HBF for AI memory
Sandisk said its NAND-based high bandwidth flash can match HBM bandwidth while offering eight to 16 times the capacity at a similar cost, using BiCS NAND and advanced die stacking. The company presented HBF as a read-optimized memory tier for AI inference, not a full HBM replacement, and outlined architectures that mix HBM and HBF side by side. Sandisk also argued that longer context windows and KV caches are pushing flash closer to compute.
Micron accelerates HBM expansion for AI
Micron is reportedly preparing to roughly double HBM capacity by the end of 2026 to about 100,000 wafers a month, narrowing but not closing the scale gap with Samsung Electronics and SK Hynix. ETNews cited industry sources saying the company is installing new tools in Taiwan and Singapore and preparing investment in Hiroshima. The bigger change is product mix: 12-high HBM4 could rise to as much as 50% of output by year-end, tied closely to Nvidia's Vera Rubin platform.
Samsung refreshes Galaxy S26 FE pricing
Samsung launched the Galaxy S26 Fan Edition in the second half of 2026 with a starting price of KRW1.045 million (≈ US$714) and global sales of the 256GB model were set to begin on September 4. The phone is meant to strengthen Samsung's mid- to upper-tier lineup as component costs rise. It includes most flagship AI and camera features, a 6.7-inch AMOLED display, and a 4,900mAh battery, while reports on a MediaTek chip were later contradicted.
Marvell maps silicon photonics growth from Taiwan
Marvell CTO Radha Nagarajan said near-packaged optics is likely to reach volume production between late 2027 and early 2028, with full co-packaged optics following later. He said Taiwan remains central because heterogeneous integration and advanced packaging are strongest there, while TSMC's Arizona and Dresden capacity is focused on silicon processing. Nagarajan also stressed that Marvell's work with Google TPU customers does not replace other suppliers, and that Nvidia investment is helpful but not required.
Onto Innovation targets AI chip process control
Onto Innovation used a SEMICON Taiwan 2026 promotional piece to argue that AI-era manufacturing needs better process control at lower cost of ownership. The company said its tools address advanced node logic, memory scaling, silicon photonics, advanced packaging, and AI packages through inspection, metrology, and analytics. It highlighted Dragonfly G5 as a single platform for wafers and 310×310mm panels, with sub-micron defect sensitivity, multiple inspection modes, and machine-learning classification for front-end and back-end uses.
Article edited by Levi Li
