Daily Edition · Page 8

THE DAILY INDEX

Technology, AI, software, security, crypto and business

GENERAL

TLDR Hardware

FIGURE SECURES UP TO 100,000 GPUS FROM NSCALE TO TRAIN ITS HUMANOID ROBOTS

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.

BIMO, AN OPEN-SOURCE BIPEDAL ROBOTICS KIT WITH SIM-TO-REAL TRANSFER

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.

ARDUINO WITH ZEPHYR

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.

NVIDIA'S PAIR TURNS SPARE RTX GPUS ON YOUR HOME NETWORK INTO ONE INFERENCE CLUSTER

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.

TLDR IndexPage 8