NVIDIA AI Updates: September 11, 2026

1. NIM Optimizations Serve 2.5x More Concurrent Users on Nemotron 3 Ultra

NVIDIA. A developer blog walks through the full-stack serving changes that let Nemotron 3 Ultra hold 2.5x more concurrent users on the same GPU footprint without giving up interactivity. The framing is aimed squarely at agentic workloads, where prompts run long and context gets reused across steps, which changes the throughput-versus-latency tradeoff compared with single-turn chat serving. Source

2. NVIDIA Is Running Its Own Million-Part Supply Chain on Nemotron and Palantir Foundry

NVIDIA. NVIDIA and Palantir detailed a joint system that encodes supply chain expertise into agents running on Nemotron models over Palantir Foundry, measured end to end from “wafer-out to first token.” The metric splits into time-to-rack, covering silicon leaving the fab through an assembled system arriving on a data center floor, and time-to-token, covering power, cooling, networking, and the software stack. NVIDIA is the first deployment. Source

3. d-Matrix Will Connect Its Raptor Inference XPUs Through NVLink Fusion

NVIDIA. Inference chipmaker d-Matrix announced it will use NVLink Fusion to attach its next-generation Raptor XPUs to NVIDIA’s scale-up networking, Spectrum-X scale-out, and the MGX rack architecture. CEO Sid Sheth framed it as a shortcut past the non-silicon work: “With NVLink Fusion and MGX, we can integrate our Raptor XPUs into a broadly deployed, liquid-cooled architecture.” It is another data point that NVLink Fusion is becoming the on-ramp competitors take rather than route around. Source

4. Skild AI’s S1 Learns New Robot Tasks From a Single Video With No Weight Updates

NVIDIA. Skild AI’s S1 robot foundation model, trained on NVIDIA infrastructure with Isaac Lab and Cosmos for synthetic data and simulation, learns previously unseen long-horizon tasks from one video demonstration using in-context learning, with no task-specific post-training or weight updates. Skild reported a $100 million annual revenue run rate ten months after its first commercial deployment, across more than 60 deployment partnerships in manufacturing, logistics, inspection, security, and food preparation. Source

5. NVIDIA Claims Every Commercial-Scale Robotaxi Program Runs on Its Stack

NVIDIA. A platform post argues that every major robotaxi program operating at commercial scale today uses NVIDIA for training, simulation, in-vehicle compute, or some combination, against a market NVIDIA projects at $400 billion by 2035 with over 6 million commercial vehicles. The pitch is modularity: developers take the training, simulation and safety validation, or in-vehicle pieces separately alongside their own stacks rather than adopting all three. Source