CONNECT WITH US

Weekly news roundup: AI boom strains chips; Google's TPUs pay off, China eyes 3nm without EUV

Staff reporter, Taipei
0

Credit: Shihmin Fu

This week’s reports centered on intensifying competition and bottlenecks across the semiconductor and AI supply chains, from Sandisk’s unusual engineering push in Korea and Google’s TPU-driven ASIC demand to equipment shortages and packaging debates that could slow capacity expansion. Taiwan’s Physical AI discussions emphasized system integration over components, while China’s semiconductor researchers outlined a DUV-based sub-3nm GAA path. Beijing also moved to restrict civilian drones starting in November sharply. Below are the most-read DIGITIMES stories from September 14 to September 20, 2026

Sandisk recruits engineers in Korea as memory talent race heats up

Sandisk is unusually recruiting engineers in South Korea for NAND system architecture, NAND core design, VLSI design and mask design, with HBM, LPDDR and GDDR experience among the qualifications. The move marks a shift from the company’s historically sales-and-marketing-focused Korean operations. Sandisk’s push reflects growing AI-oriented storage ambitions, stronger fiscal 2026 results, and the rising importance of flash and HBF in inference workloads, alongside wider competition for Korean memory talent.

Google TPU pays back investment in 1 year, creating upside for MediaTek, Broadcom ASIC revenues

Google Cloud CEO Thomas Kurian said Google’s overall AI server payback period is under two years, and under one year for servers using in-house silicon such as TPUs. That rapid return is strengthening the case for larger custom-silicon investments and creating upside for ASIC partners including MediaTek and Broadcom. Industry sources say TPU order momentum is rising, new SKUs are being requested, and the ecosystem is expanding beyond AI labs into finance, banks, biopharma, and HPC.

Deloitte Taiwan seminar highlights system integration as Physical AI moves into industrial applications

At a Deloitte Taiwan seminar in Taichung, speakers said Taiwan’s Physical AI opportunity depends less on individual robot parts than on integrating hardware, software, sensing and control into deployable systems. Officials from NIAR, Taiwan AI Academy, Deloitte Taiwan, Nvidia and Techman Robot stressed verification, simulation, data collection and practical use cases. They said Taiwan should build system capabilities, test environments and scalable industrial applications to move from monitoring production to executing tasks.

Chip equipment makers face waits of up to 40 months for critical parts

Semiconductor equipment makers are facing longer waits for critical parts as fab investment accelerates, with some component lead times more than doubling. Reported examples include a four-month delivery stretching to 10 months, waits of up to 40 months for certain Japanese parts, and local components taking about 50% longer. The strain comes as TSMC, ASML, and Applied Materials expand capacity and SEMI forecasts record equipment sales, especially for wafer fab and DRAM tools.

Google reportedly weighs CoWoS backup for 2027 TPU as EMIB-T substrate yields lag

Google’s Humufish TPU project is scheduled to use Intel’s EMIB-T packaging with MediaTek, but market rumors suggest Google is weighing CoWoS as a backup because substrate yields are lagging. Industry sources say EMIB-T yields have improved, yet dedicated substrates still face challenges, with Unimicron previously targeting 50% yield. Sources still expect EMIB-T to remain the main plan, arguing there is time before 2027 production and that dual packaging would add cost and complexity.

Exclusive: China's CAS institute charts DUV route to 3nm GAA without EUV

China’s Institute of Microelectronics under the Chinese Academy of Sciences says it has established an early-stage path to sub-3nm gate-all-around devices without EUV lithography, using DUV and stacked nanosheet-channel GAA CMOS. Chief engineer Ye Tianchun said process optimization pushed Ion/Ioff above 5×10?, with reported values of 9.7×10? and 7.6×10?. The work is presented as technology validation, not mass-production readiness, and still faces major integration and yield hurdles.

Beijing municipality bans civilian drones starting mid-November

Beijing will impose its strictest drone restrictions yet from November 15, 2026, designating the entire municipality as controlled airspace and banning drone flights, possession, storage, and transport of drones and core components. The revised rules add a citywide storage ban and require existing equipment to be removed before the deadline. Authorities are offering limited subsidies for buybacks and scrapping, with lower rates after October 31, and free EMS shipment out of the city.

Article edited by Jerry Chen