NVIDIA B200 (cloud, 180 GB)

Rented NVIDIA B200 SXM GPU with 180 GB of HBM3e at 8 TB/s, 2.25 PFLOPS of FP16 dense compute and native FP4, and fifth-generation NVLink at 1.8 TB/s when rented as an 8-GPU node.

Designed for bursty training and inference with no capital outlay, billed per GPU-hour from $2.12 per GPU-hour on marketplace clouds to $14–16 on hyperscalers; Spheron spot $5.34, and scalable from one GPU to hundreds in minutes with CUDA, PyTorch, and the NVIDIA container stack preinstalled by every provider.

Lives in someone else's data center: nothing to power, cool, or secure, and the trade is per-hour cost and data leaving your premises. Runs a 180-billion-parameter model at 8-bit on one GPU; 1 trillion at FP4 on eight.

$2.12/hr floor, $5.34 spot, $6.04 boutique, $14–16 hyperscaler. Allocation, not price, is the constraint.