NVIDIA AI Updates: October 28, 2026

1. NVIDIA and Oracle Make Vera Rubin NVL72 and Vera CPUs Generally Available on OCI, Plan Multi-Gigawatt Expansion

NVIDIA. NVIDIA and Oracle announced that Vera Rubin NVL72 rack-scale systems and NVIDIA Vera CPUs are generally available on Oracle Cloud Infrastructure, with plans to expand OCI Superclusters built on Vera Rubin to multi-gigawatt scale. The Superclusters combine NVL72 racks with Oracle Acceleron networking, Spectrum-X Ethernet with Spectrum-6 switches and BlueField-4 DPUs, building on existing GB200 and GB300 deployments. OCI also plans to offer HGX Rubin NVL8, an x86 server with eight Rubin GPUs, two 64-core Intel Xeon 6 CPUs and 4TB of DDR5. OCI Compute with Vera CPUs is positioned for CPU-heavy agent work such as tool calls and code execution, with OCI Kubernetes Engine scaling CPU and GPU resources independently. OpenAI is named as an early-access Vera CPU customer. Source

NVIDIA. CUDA 13.4 can now enforce a per-process GPU memory boundary that is transparent to the application: a workload can request an exact memory size or a guaranteed minimum with a burstable maximum, and allocations above the limit fail as they would on a smaller GPU. Workloads can also choose between time-slicing and sm-sharing compute modes. To use this from Kubernetes, NVIDIA released GPU Fractioning Operator, an open-source operator under the KAI Scheduler GitHub organization that runs next to the GPU Operator, supervises Multi-Process Service on each node, applies per-container limits from Pod annotations and reports usage per fraction rather than per GPU. KAI Scheduler v0.18.0 is the first scheduler integrated with it, through a new NvFractions sharing mode enabled with a Helm flag. Source

3. Three Optimizations Raise MoE Decode Throughput 42.5% on DGX Spark

NVIDIA. NVIDIA profiled Inclusion AI’s Ling-3.0-flash (124B total parameters, 5.1B active, 512 routed experts) at batch size 1 on DGX Spark and found decoding was memory-bound. Swapping the FlashInfer MXFP4 MoE backend for the Humming kernel library cut time per output token from 27.71 ms to 25.35 ms (9.3%). Adding online FP8 quantization of the LM head, enabled with SGLANG_ENABLE_FP8_LM_HEAD=1, halved that layer’s weight reads and lifted the gain to 21.7%. Multi-token prediction speculative decoding using the model’s NEXTN layer accepted 3.90 tokens per round on a synthetic workload. On the full GSM8K test set, the three changes together raised decode throughput 42.5% and cut mean latency 24.5% with no statistically detectable accuracy change. Source

4. Nemotron-Personas Datasets for France and Belgium Target Locale-Specific Fine-Tuning

NVIDIA. NVIDIA released Nemotron-Personas-France (built with Pleias) and Nemotron-Personas-Belgium (with Pleias and KU Leuven), synthetic resident datasets grounded in regional census, geographic, linguistic and cultural distributions. The open Nemotron data program now covers persona datasets for 12 countries. A published workflow pairs personas with a behavior contract and official procedure text to generate synthetic conversations, fine-tunes Nemotron 3.5 Lightning with LoRA on NeMo, merges the adapter and evaluates against a frozen test set. The Nemotron repository includes reusable steps for each stage plus agent skills that help coding agents assemble the pipeline. Source

5. RTX Spark Developer Stack Adds Nemotron 3.5 Bolt for On-Device Game Characters and a Faster DLSS 5 Model

NVIDIA. NVIDIA detailed developer tooling for RTX Spark, its Windows PC platform pairing a Blackwell RTX GPU and Grace CPU with up to 128 GB of unified memory. NVIDIA ACE adds Nemotron 3.5 Bolt E4B 4GB for on-device AI characters, which NVIDIA says scores about 40% higher on average than the prior model on relevant benchmarks and outputs up to about 80% fewer tokens than peers. An October DLSS 5 model update runs 2x faster than the September version, halves GPU overhead and cuts 4K VRAM use by 20%. The DLSS 4.5 Unreal Engine plugin 8.9.0 adds Windows on Arm support for UE 5.8, and Nsight tools support native Arm64, emulated x64 and Arm64EC applications. Source

6. C-BRAIN Alzheimer’s Consortium Builds Open-Source Agentic Research Tools With NVIDIA

NVIDIA. NVIDIA described its work with the Consortium for Biomedical Research and AI in Neurodegeneration (C-BRAIN), an 18-institution group led by Washington University’s Dr. Randall Bateman. C-BRAIN runs 64 NVIDIA data center GPUs on premises plus Azure nodes, uses a hybrid centralized and federated approach to train on unpublished and proprietary data without moving it, and is incorporating BioNeMo and MONAI. It has built two open-source tools: OpenScientist, an agentic co-scientist developed with Lawrence Berkeley National Laboratory that C-BRAIN says completed analyses estimated at weeks of expert work in minutes for under $10 per run, and InsightEngine, which checks claims against literature and a biomedical knowledge graph. Source