GPU Cloud
H200, dedicated.
A single NVIDIA H200 GPU, 141 GB HBM3e on the Hopper architecture, on an x86 host. $3.49 per hour, billed per minute, the same terms as every other GPUwerk tier.
Rented first come, first served: deploy an H200 when capacity is free, or join its waitlist in the console.
Specifications
Per GPU. The GPU and its memory are yours alone for the rental.
| GPU | H200 SXM, Hopper architecture |
|---|---|
| Memory | 141 GB HBM3e |
| Memory bandwidth | 4.8 TB/s |
| Max power | Up to 700 W |
| MIG | Up to 7 instances of 18 GB |
| Host | x86 host |
| Tenancy | The GPU is yours alone while you rent it, never shared |
GPU figures from NVIDIA H200. GPUwerk publishes tok/s and training throughput figures only for runs on its own hardware. See the DGX Spark fleet benchmarks for how they are reported.
What would you run on it?
The H100's compute with 76% more memory: larger models on one GPU and more room for the KV cache.
70B-class models with long context
A 70B model at FP8 fits with about 70 GB left for the KV cache, so longer prompts and more concurrent requests.
Models up to about 140B at FP8
Weights that need two H100s fit on one H200.
Fine-tuning
Full fine-tunes and LoRA runs on models that do not fit in 80 GB.
Poor fit: Native FP4
Hopper has no FP4; for 4-bit inference at Blackwell speed, rent a B200 or B300.
Need several GPUs linked by NVLink? See the 8× H200 server.
What fits in 141 GB
Weights only: memory divided by bytes per parameter. Leave headroom for the KV cache, activations and the runtime, often 10 to 30% more.
| Precision | Per parameter | Largest model, weights only |
|---|---|---|
| FP16 / BF16 | 2 bytes | about 70B parameters |
| FP8 | 1 byte | about 141B parameters |
| 4-bit quantized | 0.5 bytes | about 282B parameters (no native FP4 on Hopper; the runtime dequantizes) |
Billing
| Rate | $3.49 per hour, billed per minute |
|---|---|
| Billing | Per minute, against a prepaid credit balance |
| Fees | No egress fees |
| Commitment | None |
Prices are in USD, excluding applicable tax. Full billing mechanics, including what a stop versus a terminate does, are on the pricing page.
FAQ
Can I rent an H200 now?
Deploy it from the console when capacity is free; while none is, the console offers a waitlist and people in line are emailed one at a time, oldest first. Capacity is not held for anyone: whoever deploys first gets it.
How does billing work on this tier?
The same as every GPUwerk tier: per minute, against a prepaid credit balance, in USD, excluding applicable tax, with no egress fees and no commitment.
Need GPU time today?
A DGX Spark has 128 GB of unified memory and is rentable now, at $0.79/hour billed per minute. First top-up: pay $20, get $30 in credit. Or deploy an H200.