According to AndrewYNg, Marin 535B-A23B starts fully open training with code, data, and logs, targeting 18.75T tokens on GB200 NVL72 over ~3 months.

According to StanfordAI Lab, Marin 535B-A23B began open training on 11 GB200 NVL72 for 3 months using 18.75T tokens, with post-training to follow.

According to SawyerMerritt, SpaceX will use Nvidia Rubin GPUs and Vera CPUs in Starmind AI1, boosting peak power 67% to ~250kW for NVL72-class compute.

According to SawyerMerritt, SpaceX raised AI1 peak power to ~250kW and average to 160kW to run an Nvidia Rubin NVL72 rack, enabling compute interchangeability.

According to Satya Nadella on X, Microsoft Azure is the first cloud to bring up an NVIDIA Vera Rubin NVL72 system for validation, marking a major step toward next‑generation AI infrastructure with NVIDIA. As reported by Satya Nadella, early validation of the NVL72 cluster signals imminent availability of rack‑scale accelerated computing optimized for large model training and inference workloads. According to public NVIDIA briefings referenced by industry coverage, Vera Rubin integrates NVL72 nodes built around Blackwell‑class GPUs with advanced NVLink networking, which can increase training throughput and reduce interconnect bottlenecks for frontier models. For enterprises, this suggests shorter time‑to‑train, higher job density, and improved cost efficiency on Azure’s AI supercomputing platform, as indicated by Microsoft’s ongoing partnerships with NVIDIA mentioned by Nadella.