Topic
Explore English articles about AI Semiconductor.
NVIDIA's NVL72 market is shifting toward a structure in which rising per-rack prices from the Vera Rubin generation transition, combined with long-term guarantees and third-party capital formation, underpin the rollout of costly AI computing infrastructure.
TSMC's advanced packaging technology CoWoS expanded monthly production capacity roughly ninefold in three years, yet a supply-demand gap of about 10% still remains. Drawing on primary sources and multiple estimates, this piece examines two constraints: the physical limits of yield and the supply-chain structure in which NVIDIA secures the majority of capacity.
South Korea's DRAM export unit price hit $92,183 per kilogram in August 2026, up 401% year-on-year, a value now matching about 620 grams of gold. Samsung and SK hynix's shift toward prioritizing HBM production has starved the general-purpose DRAM supply, sending price increases cascading from NVIDIA's AI servers to VAIO's PCs.
Finland's Posiva began drilling the world's first permanent disposal hole at the Onkalo facility on August 26, 2026. The hole's specifications—8.4m deep with a 25mm tolerance—reveal a design philosophy built on multiple barriers under the KBS-3 method, standing in contrast to Japan, which remains stalled at the literature survey stage.
Rapidus, a company with no mass-production track record, is attempting to reach the 2-nanometer chip generation by combining technology transfers from IBM and imec, a production method called RUMS, and roughly ¥2.9 trillion in government support — explained alongside the 36-year decline of Japan's semiconductor industry since 1988.
SK hynix has started construction on an HBM advanced packaging plant in Indiana, where Korean-made wafers will be stacked and tested for US customers. Mass production is now targeted for late 2029, about a year later than originally planned.
TechInsights' physical analysis confirms that China's CXMT has adopted High-k Metal Gate (HKMG) technology in its G4 DRAM process and LPDDR5X products. Drawing on patents and comparisons with rivals, this piece examines what the finding actually proves—and the multi-patterning, yield, and equipment hurdles still standing between CXMT and mass production of its next-generation G5 node.
Qualcomm's High Bandwidth Compute technology is reportedly gaining support from Samsung and SK hynix. The design, which moves computation closer to memory, will be tested with AI250 samples and third-party verification in mid-2027.
A reported $45 billion compute deal by Anthropic has come to light. Measured against Monarch's plan for up to 1.35GW, the 460MW figure represents roughly 34%, but payment obligations and whether capacity will be online in late 2027 remain undisclosed.
Facing a worldwide RAM shortage, Google is tightening rules across Android 17 and Google Play, and starting in February 2027 will limit the store visibility and publication of memory-wasting apps. Here's a look at the new optimization push aimed at breaking Android's dependence on ever-improving hardware.
