CONNECT WITH US
Sep 15
TSMC 3nm, 2nm, and CoWoS stay tight as AWS, MediaTek, TPU shift capacity
TSMC's 3nm, 2nm, and CoWoS advanced packaging capacity remained in short supply in late 2026's third quarter, as changes in major customer orders and front-end and back-end supply models prompted the foundry to adjust its advanced capacity allocation. Supply chain sources said Nvidia and Apple continue to dominate TSMC's advanced process and packaging resources, while AWS has recently expanded AI chip tape-outs, giving Annapurna more 3nm capacity.
MediaTek has officially launched the Dimensity 9600 Pro, introducing a major architectural overhaul for its 2026 flagship mobile chip. Beyond the shift to a 2nm process, the new platform adopts a 2+3+3 all-big-core CPU configuration, upgrades its cache architecture, and doubles its NPU resources with a dual-NPU design to support increasingly demanding AI workloads.

The logic die at the bottom of a high-bandwidth memory (HBM) stack has become the semiconductor industry's newest contested territory. Within weeks, SK Hynix has been reported to be weighing Intel alongside TSMC, Samsung Electronics has been reported to be pairing its own foundry with TSMC, and Micron has confirmed that TSMC will manufacture its next-generation base die. What began as a single-vendor arrangement is turning into a multi-front sourcing contest — and the assumption that a second source automatically means a cheaper one is already being challenged.

Fu Chun Shin Machinery Manufacture (FSC), a major plastic injection molding machine maker, is expanding further into semiconductors by teaming up with Korean-Chinese joint venture Zhongke Xianfeng on through-glass via (TGV) equipment for glass substrates, used in next-generation advanced packaging. The two sides signed a cooperation agreement on September 14, initially targeting major printed circuit board (PCB), IC substrate, and OSAT makers in Taiwan and Southeast Asia.
High-end multilayer polymer capacitors (MLPCs) — increasingly critical to power delivery in AI servers and high-wattage GPUs and CPUs — have historically been dominated by Japan's Panasonic, which has held over 90% of the market. Now, a Taiwan-based machinery maker is aiming to challenge that grip.

South Korea's neural processing unit (NPU) startups are moving from chip development toward commercial deployment, but limited experience operating at scale remains a major hurdle to winning global customers.

Taiwan's AI server supply chain continued to show strong year-over-year growth in August, although performance varied significantly across baseboard management controller (BMC), cable/connector, and server-case suppliers.

DeepSeek is reportedly set to hire Wentao Yan, a partner at venture capital firm GL Ventures, as its first CFO. If the appointment takes place, it would happen at an important time for the Chinese AI startup as it seeks to go public, begins a large expansion of its workforce, and invests in chipmaking and computing capacity.

At the China International Optoelectronics Expo (CIOE), a palm-sized metal transceiver represents the shift from copper to light in data center infrastructure, one of the biggest transformations in the industry amid the AI boom.

DIGITIMES observed that Nvidia's newly announced NVHBM is not a brand-new memory technology, but rather an extension of existing custom HBM (cHBM). It is a managed cHBM design that moves the memory controller, which originally sat on the XPU, into the HBM logic base die, and pairs it with custom physical-layer (PHY) and die-to-die (D2D) interfaces. In strategic terms, Nvidia is extending its moat from GPUs to the memory architecture of custom ASICs, and is trying to define how future custom AI chips connect to memory, hook into scale-up networks, and come together as complete AI racks.

Taiwan's smaller IC design firms posted a strong revenue recovery in the first half of 2026, but the gains may prove difficult to extend. DIGITIMES said higher prices and early customer buying lifted sales, even as consumer demand remained soft. The second half now looks exposed to weaker momentum.

The semiconductor industry is on track to nearly quadruple in revenue by 2031 while shipping barely more silicon than it does today, a divergence that makes pricing power, not capacity, the defining variable of the AI buildout.