Memory tiers you can rent
| Memory you need | Setup | Total GPU memory | Strata on demand |
|---|---|---|---|
| Up to ~80 GB | 1× H100 80GB | 80 GB | $3.23/hr |
| Up to ~96 GB, single card | 1× RTX PRO 6000 | 96 GB | $2.03/hr |
| ~100–140 GB | 1× H200 | 141 GB | $4.31/hr |
| ~140–160 GB | 2× H100 (NVLink) | 160 GB | 2 × $3.23/hr |
| ~160–280 GB | 2× H200 | 282 GB | 2 × $4.31/hr |
| Up to ~640 GB | 8× H100 node | 640 GB | $18,863/month |
| Up to ~1.1 TB | 8× H200 node | 1,128 GB | $25,170/month |
| Up to ~1.4 TB | 8× B200 node | ≈ 1,440 GB usable | Reserved, on request |
Strata list prices per GPU-hour; node per month = price × 8 GPUs × 730 hours. Prices as of 2026-11-10.
How to work out how much VRAM you need
- Weights: parameters × bytes per parameter: 2 bytes in 16-bit, 1 in 8-bit, 0.5 in 4-bit. A 70B model is about 140 GB in 16-bit, 70 GB in 8-bit and 35–40 GB in 4-bit.
- KV cache and activations: add 20–50%. Long contexts and many concurrent users can need as much memory as the weights.
- Leave headroom. A model with 130 GB of weights technically loads on one 141 GB H200 but leaves little room for cache; two H100s or two H200s serve it more comfortably.
One big GPU or several smaller ones
- One card is simpler. No tensor parallelism to configure and no traffic between GPUs. That is why the H200 is popular for 70B serving and the 96 GB RTX PRO 6000 for 4-bit 70B and video models.
- Multi-GPU needs a fast link. H100, H200 and B200 nodes use NVLink, so splitting one model across GPUs costs little speed. The RTX PRO 6000 connects over PCIe 5.0, so it suits one model per card better than one model split across cards.
- Bandwidth matters for speed. Token generation is limited by memory bandwidth: 1.8 TB/s on the RTX PRO 6000, 3.35 TB/s on the H100, 4.8 TB/s on the H200 and 8 TB/s on the B200.
Very large models
Models of several hundred billion parameters need a full 8-GPU node. DeepSeek V3 and R1 (671B, about 700 GB in FP8) run on one 8×H200 or 8×B200 node; on 8×H100 they need 4-bit quantization or two nodes. B200 is allocated as full 8-GPU nodes on 3–36 month reserved terms. See how many GPUs DeepSeek needs.



