NVIDIA AI Updates: October 14, 2026

1. NVIDIA Open Sources the Full Run:ai GPU Orchestration Platform

NVIDIA. NVIDIA released the full Run:ai platform as open source under the NVIDIA DSX AI Factory GitHub organization (github.com/dsx-ai-factory/run-ai), with components including the Apache 2.0 KAI Scheduler (a CNCF Sandbox project), a GPU Fractioning Operator for CUDA-native GPU fractions with hard memory isolation, KAI Resource Management for projects, quotas, and node pools, and a Workload Map Go library. Scheduler features include gang scheduling, fair-share and quota allocation, topology-aware bin-packing, Dynamic Resource Allocation, and GPU sharing via MIG, MPS, time sharing, HaMi, and CUDA 13.4 MPS fractioning. The platform is available on NGC as a Community Product with a Helm chart and no subscription, but without enterprise support, while commercial Run:ai offerings continue. Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud, and vCluster are integrating it. Source

2. NVIDIA Readies Nemotron 3.5 Bolt 74B and Local AI Tools Ahead of RTX Spark Laptops

NVIDIA. Ahead of RTX Spark laptops launching October 16 with up to 128GB of unified memory, NVIDIA said Nemotron 3.5 Bolt 74B-A9B, a mixture-of-experts model with 9B active parameters packed into 24GB using 2-bit quantization-aware training, arrives October 15 with support in llama.cpp, vLLM, Ollama, Unsloth, and LM Studio, and with training data, weights, and recipes released openly. NVIDIA PAIR exits beta with llama.cpp support and direct Hugging Face downloads, NVIDIA Switchyard is a new open-source library for routing queries between local and cloud LLMs, and NVIDIA Broadcast 2.3 adds an MCP server and agent skill. Developer builds for Windows on Arm include the first PyTorch-CUDA build for the platform and a native Conda installer with over 2,000 packages, and a Nemotron 3.5 Bolt E4B model for ACE game characters follows on October 23. Source

3. NVIDIA Publishes Agent Skills for Rewriting Spark UDFs to Run on GPUs

NVIDIA. NVIDIA released cuDF agent skills for Apache Spark at github.com/NVIDIA/cudf-spark, installable with npx skills add NVIDIA/cudf-spark --skill '*', that guide coding agents through generating tests, converting a UDF, benchmarking CPU against GPU, and profiling, with failures looping back for correction. Rewrites target Spark SQL expressions the cuDF plugin maps to GPU, the RapidsUDF interface’s evaluateColumnar method using the Java cuDF API, or custom CUDA kernels via JNI. NVIDIA reports that a consumer technology company converted more than 30 production UDFs, with one UDF used by over 3,000 jobs dropping from 274.55 to 12.78 operator hours on 600 billion rows, a 22x speedup. Source

4. NVIDIA Announces Investments in Four French AI Startups

NVIDIA. NVIDIA announced planned investments in Doctolib (healthcare agents), Entalpic (atomic-scale materials engineering), Genesis AI (general-purpose robotics), and Pigment (business planning), adding to earlier stakes in Alice&Bob, Gradium, and Scintil Photonics, and said it has committed nearly $10 billion across Europe. Bull plans to extend European AI server production to Vera Rubin Ultra NVL72 systems in 2028, with rack integration at its Angers factory. The post also noted that LINAGORA and Docaposte’s Nemotron-based Twake.ai won the CANUT framework agreement covering more than 2,500 French public organizations, ahead of Jensen Huang’s GTC Berlin keynote on October 21. Source