The AI infrastructure race has spent the past few years fixated on high-bandwidth memory as one of the most indispensable building blocks of accelerated compute. Yet as model sizes, context windows and inference workloads continue to expand, the bottleneck is increasingly shifting from how much memory can be stacked toward how efficiently data can move between memory and compute.
Buried in 238 pages of targets and inventories are the things nobody puts in a press release: a formal corporate policy on shoe soles, a fish that lives in a fab, a lizard evacuation in Germany, and a workforce that is quietly out-reproducing the rest of Taiwan by a factor of four. Nine small stories from the world's most important chipmaker.
Prosecutors and investigators in Taiwan searched IC substrate maker Unimicron Technology (TWSE: 3037) on Aug. 28 in a case involving the labeling of product origin, the company said in exchange filings today.
The four major US cloud service providers (CSPs) have already rewritten their 2026 capital expenditure records, and ASIC vendors are increasingly confident that demand will remain strong through 2028 as customer spending and supply-chain visibility both extend farther out.
Apple on August 25 unveiled the M6 and M5 Ultra chips without holding a launch event, with the M6 becoming the company's first processor built on TSMC's 2nm process and debuting in the new Mac mini. But industry watchers say the M6's short-term significance lies mainly in showcasing 2nm design technology, not in delivering immediate commercial gains.
In an effort to accelerate its custom silicon initiative and curb reliance on external hardware suppliers, artificial intelligence developer Anthropic held discussions to acquire AI chip startup MatX for roughly US$7 billion, according to Reuters. While the acquisition talks have since evolved into potential partnership discussions, the move underscores Anthropic's ambition to expand its proprietary hardware capabilities.
China's memory-chip industry is narrowing parts of the technology gap with global leaders on two fronts. YMTC is approaching SK Hynix in NAND revenue share while running a more advanced layer mix, and CXMT is moving deeper into DRAM process technology with mass-produced high-k materials and a broader HBM roadmap.
According to Bloomberg, Nvidia has formed a federal political action committee (PAC), expanding its Washington presence as governments debate artificial intelligence rules that could affect businesses, workers, and consumers worldwide. The move underscores how major technology firms are preparing for a longer policy battle over AI, competition, and market access.
Nvidia's latest earnings and forecast suggest AI demand remains far ahead of available supply, with implications for memory, power, cooling, and networking worldwide. The chipmaker's guidance points to continued price pressure and a broader buildout of infrastructure that will shape AI adoption across markets and industries.
Samsung Electronics expects products based on its first-generation zHBM specifications to appear in 2029 or later, giving the 3D memory concept unveiled earlier this month an initial product window. Heat remains a key engineering hurdle as Samsung works to stack HBM directly above AI accelerators.
Samsung Electronics' packaging technology lead, Lee Hee-seok, outlined the company's advanced semiconductor packaging roadmap at Advanced Semiconductor Packaging & Chiplet Show (ASPS) 2026. Beyond sharing target production timelines and specifications for fan-out panel-level packaging (FOPLP) and glass substrates, Lee emphasized AI's pivotal role in driving packaging R&D forward.
More coverage
