NVIDIA announced NVFP4 4-bit precision format for Blackwell GPUs integrated into MaxText.

NVFP4 achieves 1.8x training speedup over FP8 on GPT-3 175B model with no accuracy loss claimed for models up to 70B.

Qualified HBM4 from Samsung, SK Hynix, and Micron for Vera Rubin

xAI pivots Colossus cluster to rental compute, targeting 30%+ margins as a neo-hyperscaler

Nvidia launches Cosmos 3 at CVPR

Launched physical AI agent skills, Alpamayo 2 Super, Cosmos 3, and other tools at CVPR

Nvidia acquired Kumo AI for over $400 million

xAI drops JAX for GPU training, builds custom C training framework using Grok Build

Integrated SuperGrok subscriptions into Hermes Agent with single OAuth login

Colossus 2 unified Blackwell-only cluster under preparation for training and IPO support