Particle.news
Download on the App Store

Nvidia’s GTC 2026 Extends AI From Data Centers to Orbit, Pairs Open Models With an Enterprise Agent Platform

Nvidia frames a unified stack to push AI inference from cloud to edge to orbit.

Overview

  • The new Space‑1 Vera Rubin module targets in‑orbit computing with up to 25× higher AI inference performance versus H100, while IGX Thor, Jetson Orin, and the RTX PRO 6000 Blackwell server GPU are listed as available now and Rubin modules will follow later.
  • Nvidia introduced NemoClaw, an enterprise version of the open‑source OpenClaw agent project, adding network guardrails and privacy controls as Jensen Huang touted OpenClaw’s momentum in interviews.
  • Nemotron 3 models, including Ultra claiming 5× throughput efficiency on Blackwell, are released on GitHub and Hugging Face, with CrowdStrike and ServiceNow cited as adopters; BioNeMo adds Proteina‑Complexa and contributes about 1.7 million complex predictions to the AlphaFold database.
  • Huang projected that Blackwell and Rubin chips could generate at least $1 trillion in sales by 2027, presenting the figure as a forecast tied to surging compute demand.
  • Industry moves track the shift to localized and embodied AI: China’s Softtone Huafang launched Ascend‑based appliances integrating OpenClaw for on‑premise agents, Unitree’s founder predicted sub‑10‑second humanoid 100‑meter sprints this year and a 2–3 year window for major embodied‑AI breakthroughs, and BenchLife reported DLSS 5 real‑time neural rendering is planned for fall 2026.