5 offers at spot / preemptible rates across 3 providers that publish this card — cheapest first, in USD (non-USD marked “≈”). Rates are per GPU-hour; the total is what the whole configuration costs.
288 GB HBM3e is NVIDIA's own figure for the B300, and it is what seven providers state (Spheron, fal, GPU.ai, Latitude, RunPod, packet.ai, Enverge). This entry held 268 GB from 2026-08-20 to 2026-09-14 on the strength of DataCrunch's API, which reports usable rather than raw memory (CoreWeave's 270 is the same kind of figure) — those strings stay as aliases of the same chip. Decision: Rahul, 2026-09-14.
Verdaformerly DataCrunchno comparable on-demand rate
Nebius AI Cloudno comparable on-demand rate
DigitalOcean GPU Dropletsno comparable on-demand rate