TLDR Hardware 2026-09-16

Agility Robotics has introduced Digit 5, a next-generation humanoid robot featuring a 50-pound payload capacity, swappable end-effectors, and a rapid nine-minute recharge capability. The system incorporates a holistic safety architecture powered by onboard sensors and advanced safety controllers to autonomously slow down, avoid, or stop when people are detected nearby.

SpaceX has set Starship Flight 14 as its first attempt at a full orbital mission, currently targeted no earlier than September 22 (an earlier FCC filing had pointed to September 15), with Booster 21 and Ship 41. The flight is expected to deploy next-gen Starlink V3 satellites, which SpaceX says carry over ten times the capacity of current satellites for "fiber-like" speeds, marking the first time Starship has released operational payloads in orbit rather than flying a suborbital trajectory. Musk confirmed the upper stage won't attempt a tower catch this time, that capability is being saved for Flight 15, where he's given it a 50-60% success chance, while Booster 21 will most likely attempt a soft splashdown in the Gulf of Mexico following an anomaly during Booster 20's boostback and landing.

Move fast and break things is great, until your robot hurts a person. Creating robots that can safely operate around people is incredibly important, but it is also incredibly time-consuming and expensive. The Open Source Safety Consortium aims to fix this using open source. The consortium is an industry collaborative initiative dedicated to defining standardized safety frameworks, reproducible verification procedures, and validation benchmarks for open-source physical AI software stacks. The organization is attempting to bridge the gap between open-source robotic software innovation and rigorous industrial safety certifications. The hope is that future robotics projects can use turn-key products and libraries to build safe robots out of the gate.

NVIDIA OSMO is an open-source, Kubernetes-native Physical AI orchestrator that connects data generation, RL training, simulation, evaluation, and hardware-in-the-loop testing across heterogeneous compute. Developers describe the pipeline in one YAML file, while OSMO schedules work across data-center GPUs, RTX simulation systems, cloud infrastructure, and edge platforms including Jetson and Arm. The system also exposes pipeline state, datasets, and available GPU capacity to coding agents.

Older stretchable screens only flexed the wiring, so fixed pixels thinned out and picture quality collapsed as you pulled. A UNIST/DGIST/IBS team printed intrinsically stretchable quantum-dot pixels at about 16,000 PPI and 53,300 nits that keep working at 1.65× stretch, so wearable and e-skin displays can stay outdoor-bright when bent.

Waymo started providing fully autonomous public rides in Las Vegas, expanding commercial deployment of its physical-AI driving stack. Las Vegas is among the first cities operating primarily with Waymo's Ojai vehicle platform powered by its sixth-generation Waymo Driver, across a nearly 24-mile service area. More than 100,000 people had registered interest.

Meta shares rose after Bloomberg reported the company plans to deploy its third-generation MTIA chip, internally called MTIA 450 or Arke, in data centers during the first half of 2027, with the next version, MTIA 500 (Astrid), set to finish design in about a month and reach data centers by the end of 2027. Twelve new sample chips arrived from TSMC on September 1 performing within 2-3% of simulations, and Yee Jiun Song, the VP overseeing Meta's custom silicon program, said each generation delivers better performance per watt and per dollar than current Nvidia offerings running the same AI models. Meta is developing the chips with Broadcom on design and TSMC on manufacturing, part of a broader push to diversify its silicon supply and reduce reliance on Nvidia.

Astera Labs' second-gen Leo 2 CXL memory controllers move from PCIe Gen5/CXL 2.x to Gen6/CXL 3.2 and double memory channels from 2 to 4, doubling both bandwidth and total capacity to up to 768GB of DDR4 or 4TB of DDR5 per expander, with new RAS features like hardware DIMM testing and soft-ECC on top of hardware ECC, aimed at economically reusing cheap, aging DDR4 pulled from retired servers rather than buying expensive new DDR5 amid the ongoing memory shortage.

The deep tech event listing for the back-to-back San Francisco and LA Tech Weeks has been announced.

NVIDIA just slipped RTX PRO 5500 Blackwell into the workstation lineup with 84GB ECC GDDR7 and a 600W budget.

Cloud software has long had programs that randomly degrade components of infrastructure to test its durability.

Micron demonstrated a 512GB DDR5 RDIMM, doubling per-slot capacity by vertically stacking DRAM dies inside the package and connecting them with through-silicon vias, the same die-stacking technique HBM uses, letting a 24-slot dual-socket server hold 12TB without changing the module count or footprint.

A preview of what's still coming from Apple through early 2027, including a redesigned MacBook Pro (rumored as "MacBook Ultra") with a tandem OLED touchscreen, M6 Pro/Max chips, and a Dynamic Island.