Nvidia acquires SchedMD and unveils Nemotron 3 open models

Summary

Nvidia has quietly deepened its software and model ambitions by acquiring SchedMD, the company behind the Slurm workload manager, while at the same time rolling out its Nemotron 3 family of open multimodal models for agentic AI. The SchedMD deal keeps Slurm open source and vendor-neutral but effectively brings a critical piece of AI cluster infrastructure under Nvidia’s wing, reinforcing how central its stack has become to training and serving large models at scale. On the model side, Nemotron 3 introduces Nano, Super and Ultra variants using a hybrid mixture-of-experts architecture, with Nvidia claiming up to 4x higher token throughput and 60% lower reasoning token generation costs versus its previous generation, plus a 1M token context window optimized for multi-agent workflows. The company is releasing not just weights but also massive pretraining, post‑training and RL datasets, along with NeMo Gym and NeMo RL libraries, positioning Nemotron as an “open but enterprise-ready” counterweight to both frontier proprietary models and the wave of Chinese open-weight competitors. Strategically, this move cements Nvidia as not just the dominant chip supplier but a major model and tooling provider in the emerging agent ecosystem, while still leaning heavily on openness to keep demand flowing to its hardware.