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.
Startup Euclyd closed a Series A round of over €200 million, with Samsung and Scaleup Europe Fund as co-investors. The company is developing the CRAFTWERK inference chip, claiming 100x energy efficiency compared to Nvidia Vera Rubin; the figure has not been independently confirmed. Former ASML CEO Peter Wennink became chairman of the board.
San Francisco company Foundation signed a $24 million contract with the Pentagon to develop humanoid combat robots. Its flagship Phantom MK1 robot is not yet armed, but the company has not ruled out installing weapons in the future.
Pratt & Whitney has licensed startup Hermeus to manufacture F100-PW-229 turbofan engines to its standards, creating a second U.S. production source. Demand for the F100 could exceed 180 engines per year by fiscal year 2034.
Tokyo startup MW has installed ceiling-mounted robotic arms in residential homes and sold its first units at 200-400 million yen. It received over 1,000 applications, raised 3 billion yen in seed funding and plans a full launch in 2028.
NVIDIA published first production results for its DSX AI factory platform: Lambda on a 19-node HGX B200 cluster gained 24% more token throughput at the same power budget. Silicon Valley Power sent 200+ power demand signals at the Eos factory.
Seagate and WD published research on storage's role in AI infrastructure. Seagate found 99% of companies expect storage requirements to grow over three years, with 43% calling storage the main AI deployment problem versus 27% for compute. WD/IDC found 94.7% already store more data due to AI, and 74% expect volumes to grow at least 25%.
At SEMICON Taiwan 2026, Google discussed AI infrastructure, saying TPU systems are scaling to 1 million chips and power supply, not compute, is becoming the main constraint. The discussion went beyond next-generation TPUs.
The Korea Semiconductor Industry Association assessed China's memory gap with the West: 1 year in NAND, 2 years in DRAM, and 3 years in HBM. YMTC produces 270-layer NAND versus 300 layers from Samsung and SK hynix, and will move to 400 layers by 2027. CXMT began DDR5 production in 2025 and is already developing LPDDR6, but HBM3E yield is only about 25%.
The first Apple M6 result appeared in Geekbench 7: 4071 single-core and 22,783 multi-core, up to 12% and 27% faster than M5 respectively. The M6 will debut in a new Mac mini launching September 22.
An Onnec survey found that 77% of UK data center operators face critical cable issues for AI workloads, and 72% believe their solution will take years. 83% said that by the time equipment launches, AI requirements already demand immediate upgrades.
SK Hynix is considering Intel as an additional maker of HBM base dies starting with HBM4E to reduce reliance on TSMC. Digitimes estimates Intel may not offer a price advantage over TSMC.
Japan's NTT Docomo has begun commercial operation of Nokia's newest base stations for multi-vendor Open RAN networks, becoming the first operator worldwide. The gear supports 4G and 5G, uses Dual Boost Massive MIMO antennas and Levante and Ponente software.
AMD released the six-core Ryzen 5 5500F on the Zen 3 (Cezanne) architecture without integrated graphics. In RandomGaminginHD tests with a Radeon RX 9060 XT 16GB, it was up to 16.1% faster than the Ryzen 5 5500 in 1080p gaming, thanks to PCIe 4.0 support and a 200 MHz higher boost. Price is under £90.
AirTrunk signed a memorandum of understanding with Emergence Quantum to deploy cryogenic cooling in data centers. The technology is expected to boost silicon chip speed and energy efficiency and prepare infrastructure for quantum computing.
Akash Systems uses GaN-on-Diamond diamond substrates to cool servers, with diamond about five times more thermally conductive than copper. Servers with AMD Instinct MI350X GPUs have been delivered to Indian operator NxtGen, and a $300 million initial order was received in the US.
NextNav conducted field tests of 5G positioning based on standard Positioning Reference Signals at 17 urban and suburban locations. Outdoor accuracy was within one meter of GPS, bringing commercialization of a terrestrial 3D PNT system as a GPS backup closer.
Developer Paulneja ran Linux kernel 6.11 on an ESP32-S3 microcontroller (240 MHz, 16 MB flash, 8 MB PSRAM) in NOMMU configuration with a custom fork() implementation. Version 0.8 boots stably for the first time, leaving 3.7 MB free memory; Linux gets only one core, the second is used by FreeRTOS for Wi-Fi and Bluetooth.
Ceva and LG Electronics have partnered to simplify the integration of ultra-wideband (UWB) into systems-on-chip. The joint platform combines Ceva-Waves UWB base IP and software with LG's RF front-end.
Meta said it will begin deploying its third-generation custom AI chip, MTIA 450, in data centers in the first half of 2027, followed by a broad rollout of MTIA 500 by the end of the year.
Best Buy lists Nvidia RTX 5090 cards at $7494.99, over triple the $1999 launch price. The increase is attributed to GPU shortages driven by AI server demand; the card launched in January 2025.
Meta detailed the benefits of its new in-house AI chips: a processor will run in data centers in 2027, with another chip model due by the end of 2027. The company aims to cut costs by reducing reliance on external accelerators.
By 2035, US data centers are projected to burn up to 15 billion cubic feet of gas per day, a 117% increase from the previous estimate of 6.9 billion. That is about 20% of US electricity consumption, or 194 GW.
GF Securities analyst Jeff Pu estimated Intel's 18A yield at 80%, with 14A output at 6,000 wafers per month by the end of 2027 and 24,000 in 2028. He also suggested possible Intel-NVIDIA collaboration on an AI-CPU based on the Rosa Feynman architecture.
Micron has demonstrated the first 512GB DDR5 RDIMM with speeds up to 9200 MT/s, validated by AMD and Intel. The module uses 3D-stacked DRAM with TSV, enabling up to 12TB in a 24-slot dual-socket server. Production is planned for the second half of 2027.
Apple M5 Ultra with 36 cores outperformed 96-core AMD Ryzen Threadripper PRO 9995WX in Geekbench 7: 10.8% higher multi-core and 46.4% higher single-core (3774 vs 2578). In Ray Tracer and Clang tests, AMD led by up to 58.8%.
Astera Labs expanded its Leo memory controller lineup with three new products to accelerate KV cache processing for agentic AI. The solution addresses HBM memory shortages on accelerators as inference shifts from single queries to continuous AI agent cycles.
Rambus announced an IP solution for PCIe 7 switches with time division multiplexing (TDM), aimed at connecting SuperNICs and accelerators in AI infrastructure. The technology aims to provide PCIe 7th generation bandwidth for next-generation AI systems.
Dutch firm Axelera AI has begun shipping its second-generation Europa AI processor: 629 TOPS (INT4/INT8/INT16) at 45W TDP across eight AI cores and 16 RISC-V cores, with 128MB L2 SRAM and 200GB/s DRAM on Samsung 5nm. It is available as a chip, the Edge 232p card and the Server 250p board with four AIPUs; Dell XE5 and Supermicro 111AD servers are validated.
U.S. companies VulcanForms and Specter Aerospace have partnered to create large-scale production of hypersonic air-breathing systems using metal 3D printing. The partnership will cover munitions, interceptors, and cruise missiles, including full assembly and integration. The location and cost of the future plant have not been disclosed.
