NVIDIA AI Updates: May 22, 2026

1. NVIDIA sweeps COMPUTEX Best Choice Awards on Vera Rubin NVL72, Jetson Thor, and Alpamayo ahead of June 1 GTC Taipei keynote

NVIDIA. NVIDIA opened its COMPUTEX 2026 run by collecting three top honors at the Best Choice Awards: Vera Rubin NVL72 took both the Golden Award and a Sustainable Tech Special Award, Jetson Thor won a Golden Award for edge and physical AI, and the Alpamayo autonomous-vehicle platform won the Vehicle Technology and Smart Cockpit category. The Vera Rubin NVL72 pitch centers on a rack-scale system combining 36 Vera CPUs and 72 Rubin GPUs with a cable-free, liquid-cooled design that NVIDIA says cuts assembly to roughly five minutes and delivers up to 10x higher inference performance per watt and 10x lower cost per token versus the prior generation. Jensen Huang’s GTC Taipei keynote is set for 11 a.m. local time on June 1 at Taipei Music Center, with the broader conference running June 1 through June 4 and Jetson Thor positioned as a 2,070 FP4-teraflop platform already shipping into hundreds of robotics applications. Source

2. NVIDIA open-sources GPU Usage Monitor, a one-command Helm stack for Kubernetes GPU observability

NVIDIA. NVIDIA released GPU Usage Monitor, an Apache 2.0 open-source tool that bundles DCGM Exporter, kube-state-metrics, Prometheus, and Grafana into a single Helm chart for real-time visibility into GPU allocation, compute utilization, memory consumption, and pod status across Kubernetes clusters. The chart deploys in three commands, ships pre-built Grafana dashboards, and supports filtering by GPU type including Hopper and Blackwell, with optional integration into an existing external Prometheus and configurable resource allocation and credential handling via Helm values. The stated goal is closing the “observability gap” that has dogged GPU operators on Kubernetes 1.19+ and Helm 3.0+ environments, where fragmented metric sources made it hard to see compute and memory trends against thresholds without bespoke dashboard work. Source

3. NVIDIA shows a NeMo Agent Toolkit multi-agent system that autonomously discovers quant trading signals

NVIDIA. NVIDIA published a reference design for a self-improving quant research pipeline built on the NeMo Agent Toolkit, NVIDIA NIM, and Nemotron, using three specialized agents that hypothesize alpha signals, generate Python implementations, and backtest them in a continuous loop. The Signal Agent composes hypotheses from a structured library of 66 mathematical operators spanning arithmetic, rank, and time-series functions, the Code Agent converts blueprints into executable Python, and the Evaluation Agent runs backtests, computes Information Coefficient metrics, and feeds refinement suggestions back upstream, with Arize Phoenix providing LLM observability and tracing. In the published run, a Rank-Adjusted Return Momentum signal multiplying price rank by return rank achieved a mean Information Coefficient of -0.0134 with p less than 10 to the minus 7 over 3,504 trading days, and the entire system is config-driven through YAML so researchers can swap models and thresholds without touching code. Source