From Blackwell Ultra to Ada, every class is metered, indexed and deliverable on Quantum-2 InfiniBand. Pick a SKU to see specs and forward pricing.
| GPU | Architecture | Memory | Spot /hr | Best for | Availability |
|---|---|---|---|---|---|
| GB300 NVL72 | Blackwell Ultra | 288 GB | $4.30 | Frontier training | Low |
| B200 | Blackwell | 192 GB | $3.60 | Training + inference | Medium |
| H200 | Hopper | 141 GB | $2.45 | High-throughput inference | High |
| H100 | Hopper | 80 GB | $2.05 | Training workhorse | High |
| L40S | Ada | 48 GB | $1.05 | Inference + graphics | High |
| RTX 5090 | Blackwell | 32 GB | $0.62 | Dev + fine-tune | Medium |
| RTX 4090 | Ada | 24 GB | $0.41 | Dev + batch | High |
Available SKUs (H100/H200/L40S) provision in days. Frontier classes (GB300/B200) are delivered against forward contracts from the build-out, typically Q3 to Q4 2026.
Non-blocking Quantum-2 InfiniBand with 8-way NVLink nodes, tuned for large-scale distributed training.
Yes. Burst on spot, anchor your baseline on reserved or a forward contract indexed to WATT-CPI.