Connectivity aggregation from FORT Robotics merges cellular, satellite, Wi-Fi and radio into one link for remote machines. What it claims and what’s unproven.
Category: AI & Emerging Tech
Modular Wafer Fabrication: GLS Debuts ChipForge Project
Modular wafer fabrication gets a first test: Great Lakes Semiconductor’s ChipForge project targets AI vehicle sensors with APES, BlueDot and FMC. Caveats.
Quantum Error Correction: GPUs Cut Calibration to 2.8 s
Quantum error correction calibration dropped from 3 minutes to 2.8 seconds on an NVIDIA H100 GPU, Riverlane says. What the result shows and what it leaves out.
Co-Packaged Optics: Lightmatter, Flex Show Open Rack Trays
Co-packaged optics reach open racks: Lightmatter and Flex show a 204.8 Tbps switch tray, plus a light-linked 4-XPU tray, at the 2026 OCP Global Summit.
Maskless Lithography: Nikon DSP-110 Speeds Chip Packaging
Maskless lithography gets faster: Nikon’s DSP-110 for chip packaging promises 1.5 µm resolution and about 40% more throughput. It ships in April 2027.
Light-Powered Brain Reads Hidden Patterns in Light
A light-powered brain from Wits University recognised 81 light patterns after training on 15 examples. How it works and what it could mean for communications.
AI Data Center Power Monitoring: TI Nanosecond ADC
AI data center power monitoring gets faster: TI says its new ADC34DSO75 spots nanosecond power glitches remotely. Here is what it does and why it matters.
NVIDIA and Cisco USFP: Pluggable Optics for AI Factories
NVIDIA and Cisco unveiled USFP, pluggable optics with 400Gb/s per lane and up to 12.8Tb/s, offering 4x bandwidth density for AI data centres. Specs and limits.
InfiLoop: A Tiny AI Model That Thinks in 20,000+ Loops
InfiLoop, a 7M-parameter looped Transformer, keeps improving past 20,000 loops and scores 97.9% on Sudoku-Extreme in a new preprint. How it works and limits.
New MIT Algorithm Lets AI Follow Strict Safety Rules
MIT’s HardFlow algorithm enhances pretrained generative AI models, enabling them to meet strict safety requirements without retraining. Unlike traditional approaches that constrain every step, HardFlow reformulates the task as a trajectory-optimization problem, allowing models to produce high-quality outputs while satisfying hard constraints effectively across various applications.