TLDR Hardware 2026-09-04

How can we improve TLDR Hardware? We'd really appreciate it if you could fill out this three question survey to give us a bit of feedback!

Figure signed a deal with Nscale to deploy up to 100,000 Nvidia Vera Rubin GPUs, starting in Barstow, Texas, in the second half of 2027, an initial $3.5 billion compute commitment with intent to scale past $6 billion, with Nscale also making a strategic investment in Figure. It follows last week's launch of Index, Figure's crowdsourced humanoid training dataset now generating 35 minutes of data every second, addressing the other half of the bottleneck: Figure says it's now largely constrained by data and compute rather than robot hardware itself in training its Helix model.

Uber began deploying Wayve-powered autonomous vehicles across London on Thursday, marking Wayve's first commercial public service anywhere in the world and making London only the second European city with robotaxis, after Zagreb. The initial fleet, fewer than 20 Ford Mustang Mach-Es with plans to add Nissan Leafs later, keeps a licensed safety operator behind the wheel for now, and riders booking UberX, Comfort, or Electric may get matched with one at no extra cost.

Micron is reportedly investigating near-GPU NAND, high-endurance flash modules placed right on the GPU package where HBM sits today, aiming to fill the massive latency and bandwidth gap between HBM and a system's SSD. The pitch is memory-bound inference: large models buy extra accelerators purely to hold weights and KV cache, an expensive way to add capacity since HBM is the priciest, tightest-supplied memory around, so a cheaper flash tier fast enough to keep compute fed could let the same model run on fewer GPUs.

Nvidia's RTX Spark superchip fuses a 20-core Grace Arm CPU with a Blackwell RTX GPU on one 3nm package, delivering up to 1 petaflop of AI performance and 128GB of unified LPDDR5X memory, enough to run 120-billion-parameter models locally with up to a 1-million-token context window. It's Nvidia's first true client SoC for Windows, pairing native CUDA support with full RTX graphics tech, so the same chip targets AI development, creative work, and gaming. Jensen Huang promises at least 100fps at 1440p in modern AAA titles.

Arduino trained a generation of young engineers, and now, post-Qualcomm acquisition, they are professionalizing the framework. The recent Arduino Core on Zephyr 1.0.0 release bridges the familiar, beginner-friendly Arduino API with the industrial-grade Zephyr Real-Time Operating System. This integration allows developers to leverage enterprise capabilities like preemptive multi-threading, advanced power management, and standardized networking stacks while writing standard Arduino code.

Bimo is an open-source, 3D-printable bipedal robot kit built with custom quasi-direct drive actuators, ROS 2 integration, and simulation-ready environments for reinforcement learning locomotion policies. The repository provides complete mechanical models, electronics schematics, and training pipelines to enable reliable sim-to-real transfer on a desktop-sized robot. It offers an accessible, affordable testbed for experimenting with legged locomotion dynamics, custom motor control schemes, and modern reinforcement learning algorithms without requiring high-cost industrial hardware.

As multi-agent AI workflows spawn dozens of independent model calls from a single task, all competing for one local GPU's execution queue, Nvidia's new open-source PAIR router spreads those calls across every eligible RTX GPU, RTX PRO workstation, DGX Spark, or Apple M4+ device already sitting on a home network. It doesn't split one inference request across multiple GPUs or pool VRAM into a bigger virtual accelerator - each request still runs whole on one node, but it proxies existing Ollama and LM Studio interfaces so agents keep using the connection they already expect while PAIR handles discovery (via mDNS), MTLS-secured pairing, and live scheduling based on which paired node is actually ready.

Dyson's CameraJet packs a 100,000-pixel macro lens camera into the brush head, running its "Gap Optical Targeting" machine learning system to detect and predict gaps between teeth in real time, triggering a jet of water within 100 milliseconds of spotting one.

HTC's Vive Eagle glasses go on sale in the US, Europe, and Australia at $499 (or $599 with photochromic lenses).

AGS-PlantSeg is an open-access 3D point cloud dataset and benchmark designed for semantic segmentation in complex agricultural environments across varying plant growth scales.

Broadcom guided to $34.8 billion in fourth-quarter revenue, below the $35.1 billion analysts expected and well under some estimates that topped $36 billion, sending shares down about 4% after hours.

TSMC Senior VP Jeffrey Hou revealed at SEMICON Taiwan 2026 that AI-driven demand has pushed the company's equipment purchase forecasts to nearly double what they were just six months earlier, as it simultaneously builds roughly 13 fabs in Taiwan and 5-6 overseas, 4-5 times its historical pace.