NVIDIA AI Updates: September 17, 2026

1. Vera Rubin NVL72 Debuted at 3.7x the Throughput of GB300 NVL72

NVIDIA. MLPerf Inference v6.1 results landed with Vera Rubin NVL72’s first preview submission delivering up to 3.7x better throughput than GB300 NVL72. NVIDIA also submitted a 288-GPU configuration spanning four GB300 NVL72 racks that held 99% scaling efficiency against a single-rack baseline, meaning throughput grew close to linearly with added hardware. Software optimizations alone accounted for up to 1.6x higher performance versus the v6.0 submissions. Source

2. NVIDIA, Google, and Emerald AI Launched an Alliance for Grid-Responsive Data Centers

NVIDIA. The three companies announced the AI Energy Management Alliance, a coalition for data centers that adjust electricity draw in response to grid conditions rather than presenting flat, static demand. The argument is that interconnection processes were designed for inflexible loads, and that a facility able to shift workloads, discharge storage, or lean on paired generation can be treated as a controllable resource — which in turn unlocks faster and larger grid connections. Power has become a binding constraint on US AI infrastructure expansion, so the goal is extracting more useful capacity from existing grid headroom. Source

3. TensorRT Edge-LLM Cleared the MLPerf Edge Agentic Benchmark 6.4x Faster Than llama.cpp

NVIDIA. Running Qwen3.6-27B on a single Jetson AGX Thor Developer Kit, TensorRT Edge-LLM hit 52.33 tokens per second and finished all 1,007 turns of the MLPerf Inference v6.1 Edge Agentic performance workload in 24 minutes 36 seconds, against 2 hours 37 minutes for the llama.cpp reference. The gains come from NVFP4 quantization on weights and activations, an FP8 KV cache, and KV cache reuse across agent turns that served roughly 96% of prompt tokens from hot cache. Tree-based multi-token prediction with an 8-step, top-2, 16-node verification tree added about 40% decoding throughput over linear MTP on function-calling workloads. Source

4. An Agent Skill Ported All 24 TileGym Operators From Python to cuTile Rust

NVIDIA. NVIDIA described cuTile Rust, which extends Rust’s ownership model to tile-based GPU kernels by splitting mutable outputs into disjoint pieces while preserving host-side ownership contracts across kernel launches. An AI agent skill in the TileGym repository translated cuTile Python and Triton-TileIR kernels into cuTile Rust, converting all 24 public TileGym operators at 99.5% of cuTile Python performance on average when benchmarked on a DGX B200. The pipeline is a bounded multi-agent workflow with machine-checkable verdicts at each stage, including IR diffing against reference Tile IR to confirm structural equivalence before any functional testing runs. Source

5. Earth-2 CorrDiff Was Retrained to Forecast UK Air Pollution

NVIDIA. University of Manchester professor David Topping worked with the Earth-2 team to apply the generative downscaling model CorrDiff to pollution fields rather than weather, training it on data generated from existing chemistry-climate simulations using Isambard-AI, the UK’s national AI supercomputer. Chemistry-based air quality models are slow enough that resolution and run frequency both suffer, which is the constraint the approach targets. The team has since added Earth-2 StormCast for time-dependent forecasts that consume air quality observations directly, and demonstrated the test-training and inference workflows on a DGX Spark. Source