The owner of a PNY GeForce RTX 5090 in a DigitalStorm prebuilt PC found a hole in the 16-pin power connector after smelling burning during a game of Crypt of the NecroDancer. PNY denied the warranty claim, citing use of a non-original cable; a local repair was estimated at about $180.
Growing demand for optical interconnects in AI servers and the transition of co-packaged optics (CPO) to mass production are shifting silicon photonics testing from component characterization to wafer yield protection and early defect detection.
Rising AI infrastructure demand is pushing Taiwan's copper-clad laminate (CCL) makers to adopt Nvidia's high-frequency M9-grade materials. AI servers are moving to 52-layer PCBs, making CCL upgrades critical for system performance.
GMKtec announced the compact EVO-X5 Pro workstation with AMD Ryzen AI Max+ Pro 495, up to 192GB unified memory (up to 160GB for graphics). It claims offline inference of models up to 300B parameters; release on September 28, pricing not announced.
At the Tech in Asia conference in Singapore, Gigabyte demonstrated the compact AI computer AI TOP ATOM, slightly larger than a Mac mini. It is based on the Nvidia GB10 Grace Blackwell Superchip with a 20-core Arm processor and Blackwell GPU, 128 GB unified memory; the device can run models up to 200 billion parameters. Price is S$8799.
Samsung's Exynos processors are gradually regaining share in the global smartphone market, while flagship Galaxy phones are being tested on a 2nm process. Details and figures were not disclosed.
General Atomics unveiled the Wildfire project, a cheap and mass-producible reconnaissance-strike drone for the U.S. Air Force's Massed Modular Aircraft program. The vehicle has a ferry range of 14,816 km, carries two anti-ship missiles or four Joint Strike Missiles, and is controlled via pLEO satellite constellations.
Delos Data, a startup founded by former Intel engineers, raised over $100 million from Matrix, Playground, Socratic Partners, Capricorn Technology Impact Fund, Matter Venture Partners, and IAG. The funds will expand teams and accelerate development of networking chips for AI infrastructure with latency around 100 nanoseconds.
Princeton University engineers created a semiconductor a few molecules thick whose properties can be repeatedly changed by light. They built a 1-square-inch sample and arrays of programmable switches; the next step is connecting them into circuits.
Huawei announced a 3D data center concept in China, along with a monograph and a library of design images on the topic.
Samsung Electronics is considering using Micron's advanced 1γ LPDDR5X memory in Galaxy S27 smartphones. Rising prices for next-generation mobile DRAM are complicating supplier choice.
According to Digitimes, growing sizes of AI chip packages are pushing advanced packaging to a new level: in 2027, the market will be reshaped by glass substrates and through-glass via (TGV) technology.
South Korea's Hanmi Semiconductor has reportedly received an order to supply advanced chip packaging equipment for Elon Musk's US megaproject Terafab. This is the first known case of a Korean back-end equipment supplier entering this project.
Samsung has started prototype production of Tesla's AI5 chip at its Taylor plant using a 2nm process, beginning yield verification ahead of schedule. Full verification will be completed by year-end, with deliveries from 2026; the Tesla contract is valued at $16.5 billion.
US startup Kepler Computing emerged from stealth after seven years, disclosing $468 million in funding. The company offers a new way to produce dense HBM memory without EUV lithography, attracting interest from major chipmakers.
A Digitimes report notes record revenue for the top three memory makers in Q3 2026, driven by AI server demand. Main growth comes from enterprise SSDs transitioning to PCIe Gen6 and LPDDR6 memory.
TSMC Chairman C.C. Wei and co-COO Y.J. Mii spoke at the "Dialogue Across Generations: Q&A for the AI Era" forum on September 15. Mii said that when moving to advanced process nodes there are no ready-made solutions and each node requires new engineering approaches.
Taiwan's National Science and Technology Council (NSTC) is stepping up efforts to move beyond the 1nm process node. The focus is on new materials, device architectures, process technologies and advanced packaging.
According to Digitimes, Samsung has turned to TSMC to manufacture base logic dies for HBM stacks. This changes the balance of power in the contract memory manufacturing market.
A Taiwanese equipment manufacturer plans to use AI to challenge Japan's dominance in high-frequency multilayer polymer capacitors (MLPC), critical for powering AI servers and powerful GPUs/CPUs. Panasonic currently controls over 90% of this market.
KT and KISTI opened Korea's first dedicated 400 Gbps international link between Daejeon and Chicago. Capacity on this route of the national science and technology network quadrupled from 100 to 400 Gbps, needed for large data transfers in AI and astronomy.
At the SEMICON Taiwan 2026 sustainability and energy forum, participants discussed whether Taiwan has enough electricity for semiconductor expansion amid AI demand. Digitimes concluded the island can support the chip boom, but not through green energy.
Humanoid robot makers are cutting prices and ramping up output, but the industry remains unprofitable. Profits are concentrating among component suppliers, especially makers of dexterous robotic hands.
At China's CIOE exhibition, GlobalFoundries showed a compact metal transceiver marking data centers' shift from copper to optics. The company is entering the silicon photonics market as AI data centers hit power transmission limits.
Digitimes reports that Nvidia's announced NVHBM memory is not a new technology but an extension of existing custom HBM (cHBM). The company aims to platformize cHBM and establish itself in custom AI architectures.
BAE Systems announced the Shadow EW family of electronic warfare systems for small crewed and uncrewed platforms. The architecture uses commercial chips and open standards, supports post-deployment software updates, and is designed for mass production.
Nvidia's flagship RTX 5090 has risen to $9,000-$12,000 from third-party sellers in the US, while major retailers have no stock. A rumor suggests the GeForce RTX 6000 (Rubin) will launch in the first half of 2027, risking even higher prices.
A Geekbench result for a Mac Studio with a 36-core M5 Ultra chip and 256 GB of memory showed 3774 single-core and 52,516 multi-core scores, about 6 times higher than M1. The configuration costs $10,799, with limited supplies and shipping dates into January.
At AI Infra Summit, Cornelis Networks and Delos Data presented alternatives to Nvidia's proprietary NVLink interconnect. Cornelis announced the open Active Compute Fabric architecture, while Delos showed I/O chiplets with over 30 Tbps throughput, twice that of Nvidia and AMD.
Pratt & Whitney completed the Risk Reduction Design Review for the F135 Engine Core Upgrade on September 15. The modernization will give the F-35 more power and improved cooling for Block 4 systems without replacing the engine architecture.
