NVIDIA AI Updates: October 23, 2026

1. NVIDIA Unveils Minerva, an AIOps Platform for AI Factory Observability

NVIDIA. NVIDIA launched Minerva, an AIOps platform that combines infrastructure and workload telemetry and builds a dynamic knowledge graph mapping GPUs, SuperNICs, switches, cables and the jobs running on them, so operators can trace a slow training job to the component causing it. It analyzes billions of data points per hour with supervised and unsupervised models to detect anomalies, predict failures and group related events into incidents with a likely root cause, and it includes a native AI assistant plus connections to external agentic tools and automation APIs. IREN is an early adopter, Aolani, Bitdeer AI, IBM, Jio, Lambda, Quake AI and Tesla plan to deploy it, and Dell, Lenovo and Supermicro will offer it in their AI factory portfolios. Source

2. Firebird Secures Export License for 45,000 Grace Blackwell GPUs in Kazakhstan

NVIDIA. NVIDIA reported that Firebird, which previously built an AI factory in Armenia, has secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs for a Kazakhstan buildout. Phase one is under construction with Kazakhtelecom at Data Center Valley, targeting commercial launch in Q3 2027, and a second phase would bring Firebird’s potential capacity in the country to about 500 megawatts. Kazakhtelecom already runs 400 H200 and 480 B300 GPUs in Astana and Almaty and plans more than 1,000 Vera Rubin GPUs in the Ekibastuz region next year. Source

3. Cosmos Adds Data Factory Blueprints, Cosmos 3 Nano Simulation Variants and Faster NIM Microservices

NVIDIA. NVIDIA expanded its Cosmos world model platform with Cosmos Data Factory, including five cloud-native blueprints (autonomous vehicle dataset generation, image attribute augmentation, event video generation, defect image generation and video data augmentation) available on GitHub, Microsoft Azure, Nebius and CoreWeave. The Cosmos Framework’s training, distillation and quantization recipes are now packaged as agent skills covering Cosmos 3, GR00T, Alpamayo, Qwen and Pi0.5, and new Cosmos 3 Nano variants target multiview, multisensor simulation of AV and bimanual robot policies. Updated Cosmos 3 NIM microservices deliver up to 2x faster inference in validated configurations, and Metropolis VSS 3.3 monitors 40% more video streams per GPU; Porsche says a Cosmos-based data flywheel across Volkswagen Group brands could cut repeated physical test drives by up to 90%. Source

4. Energy Companies Adopt NVIDIA DSX to Co-Design Grid Power and AI Factories

NVIDIA. NVIDIA detailed partner adoption of its DSX AI Factory Platform, which includes DSX Sim, DSX Flex for coordinating AI factory demand with grid conditions, and DSX MaxLPS and DSX Exchange for compute-per-watt optimization. Hitachi Energy’s battery storage system qualified as DSX Ready, Schneider Electric validated Vera Rubin power and cooling reference architectures, and Siemens Energy is developing a grid-aware power reference design. Phaidra reports about 10% more compute within the same power envelope using DSX components, and Emerald AI’s Conductor integration with DSX Flex is modeled to unlock up to 100 GW on the existing U.S. grid. Source

5. Earth-2 Releases StormScope EU for European Short-Term Weather Forecasting

NVIDIA. NVIDIA released StormScope EU, a European version of its StormScope nowcasting model that uses EUMETSAT satellite observations to predict cloud conditions across 16 Meteosat channels at 10-minute intervals. Adopters of Earth-2 open models include AXA (cyclone risk across more than 1,000 simulated seasons with FourCastNet 3), Vaisala Xweather, TotalEnergies (solar radiation nowcasts at photovoltaic sites) and the University of Manchester, which adapted CorrDiff for U.K. air-quality modeling. Space42 also built SARDiff4Height, which estimates surface heights from a single radar image, using CorrDiff and PhysicsNeMo. Source

6. European Governments Build Sovereign AI Services on NVIDIA Open Models

NVIDIA. NVIDIA highlighted European public sector projects running open models on government-controlled infrastructure, including deepset’s Haystack Enterprise Platform behind the European Commission’s AI@EC and a Polish agentic factory for more than 400 labor offices. secunet, NVIDIA, Red Hat and deepset are developing a sovereign cloud on secunet’s SINA Cloud, with security components authorized for classified information, combining Nemotron models and Red Hat OpenShift AI. New synthetic datasets include Nemotron-Personas-UK from Locai Labs (1 million personas) and Nemotron-Personas-Germany from the Fraunhofer-led Soofi initiative, based on the 2022 census. Source

7. NVIDIA Benchmarks 753B-Parameter GLM 5.3 on an Eight-Node DGX Spark Cluster

NVIDIA. A technical post details official eight-node DGX Spark clusters (1 TB of unified memory) connected via ConnectX-7 with RoCE, running the full 753B-parameter GLM 5.3 in NVFP4 (about 470 GB) at 1,600 tokens per second prefill and 31 tokens per second decode for a single user. NVIDIA introduced RoCEnante, a Spark-optimized collective backend merged into FlashInfer, which raised Qwen3.8-27B decode speedup at eight nodes to 4.43x versus 3.05x without it. Multi-node nanochat pre-training scaled from 8,741 to 68,079 tokens per second across eight nodes, or 97.4% scaling efficiency, and NVIDIA Sync’s Cluster Assistant is expected to support eight-node setup by year end. Source