NVIDIA AI Updates: July 18, 2026

1. NVIDIA Positions Vera Rubin Around Intelligence Per Dollar for Agentic Post-Training

NVIDIA. NVIDIA detailed how its Vera Rubin platform targets the continuous post-training loop that agentic AI systems require as environments shift and tools change, emphasizing a new metric it calls intelligence per dollar alongside cost per token. The company said Vera Rubin can train the largest models with one-fourth the GPUs of the Blackwell generation, and cited testing from Prime Intellect showing roughly 30% greater throughput per CPU for reinforcement learning workloads versus x86 alternatives. NVIDIA also pointed to its 550-billion-parameter Nemotron 3 Ultra model scoring 71.7% on the SWE-bench real-world coding benchmark as an example of the payoff from sustained post-training investment. Source

2. NVIDIA Details BlueField Co-Design for Agentic AI Factories

NVIDIA. NVIDIA described how BlueField turns data center infrastructure into part of the inference pipeline for agentic AI factories, offloading networking, storage, and security from host CPUs so those cores stay focused on agent execution. The BlueField-4 DPU pairs a 64-core Grace CPU with up to 800 Gb/s of Ethernet or InfiniBand connectivity and PCIe Gen6, delivering double the networking bandwidth, six times the compute, and four times the memory capacity of BlueField-3. NVIDIA also introduced the Vera BlueField-4 STX storage processor, which combines a Vera CPU and ConnectX-9 SuperNIC with up to 1.6 Tb/s of Spectrum-X Ethernet and is optimized for KV cache management, claiming gains in GPU utilization, latency predictability, and tokens per watt. Source