NVIDIA AI Updates: September 22, 2026

1. TensorRT Multi-Device Inference Landed in Dynamo-Triton

NVIDIA added TensorRT multi-device inference to Dynamo-Triton, aimed at the case where a model’s compute and memory demands exceed what one GPU can supply. The integration handles serving across multiple GPUs rather than requiring the sharding and coordination to be built at the application layer. For teams currently hand-rolling multi-GPU serving around Triton, this collapses a meaningful chunk of that glue code. Source

2. NVIDIA Published a Guide to Evaluating Agents on Task Completion, Not Tool Calls

NVIDIA argues the shipping question for an agent is whether it can execute a chain of work across dozens of sequential tool calls against a live environment, which single-turn benchmarks do not measure. The guide walks from tool-call level metrics up to end-to-end task completion. It is a useful counterweight to eval suites that score an agent on whether it picked the right function signature. Source

3. NVIDIA Framed AI Security as an Engineering Problem With Named Owners

NVIDIA published a layer-by-layer treatment of agent stack security, defining it as an engineering discipline with security requirements, enforceable controls, named owners, and evidence that the protections actually work. The emphasis on evidence and ownership rather than policy documents is the operative part. It lands the same week a 0-day surfaced in a widely deployed consumer agent. Source

4. Physical AI Deployment Needs Safety at Every Layer, Not Just the Model

NVIDIA made the case for layered safety in physical AI, citing ABI Research’s projection of 49 million level 3-5 autonomous vehicles installed by 2035 and Omdia’s estimate of roughly 60 million industrial robots deployed between 2026 and 2035. The Halos framework covers safety across the stack rather than treating it as a model property. The scale numbers are the argument: at that installed base, per-deployment safety review stops being tractable. Source

5. DSX Ready Qualifies Power and Cooling Hardware Against AI Factory Designs

NVIDIA launched DSX Ready, a qualification program for power and cooling products used in AI factories. The premise is that power, cooling, water, site, and grid constraints now determine what builders can actually deploy, so the components need to be validated against the complete factory design rather than in isolation. Facilities constraints, not accelerator supply, are increasingly the thing that gates capacity. Source