Blackwell B200 and GB200 NVL72 ramp across clouds
NVIDIA's Blackwell data-centre parts are now bookable by the hour, with 192 GB of HBM3e per GPU pushing single-node model sizes higher.
Blackwell has moved from launch slides to something you can actually rent. B200 instances are appearing on the on-demand menus at Lambda and CoreWeave, and the rack-scale GB200 NVL72 is shipping to the largest buyers — 72 GPUs wired together as one big accelerator over NVLink.
The headline number for most people is memory: 192 GB of HBM3e per GPU means far larger models fit on a single device before you have to split them across cards. That lowers the engineering tax on inference and shortens the path from a research checkpoint to a served endpoint.
If you're pricing this up, compare hourly rates on the hosting board rather than buying — data-centre Blackwell is sold through partners, not retail.