GPU capacity and pricing
We run your models on private GPU capacity and negotiate the price for each deal. The prices below are a reference point. When we talk, we can usually beat them.
GPUs
| GPU | Memory | Good for | Reference price per GPU-hour |
|---|---|---|---|
| NVIDIA L40S | 48 GB | Small and mid-size models, embedding models and vision models. | from $1.10 |
| NVIDIA RTX PRO 6000 | 96 GB | Mid-size models that fit on one GPU, at a lower cost than HGX systems. | $1.25 |
| NVIDIA H100 | 80 GB | The most common GPU for production LLM serving. | $3.15 |
| NVIDIA H200 | 141 GB | Larger models and longer contexts that need more memory per GPU. | $3.80 |
| NVIDIA B200 | 180 GB | High-throughput serving of large models. | $5.95 |
| NVIDIA B300 | 288 GB | The largest open models and long-context workloads. | $6.65 |
| NVIDIA GB300 NVL72 | 288 GB per GPU | Very large models served across 72 GPUs in one NVLink domain. | Priced per deal |
We negotiate the price of dedicated capacity with our partners for each deal, and it is usually lower than these reference prices.
How it works
Which model you want to serve, how much traffic you expect, and any region or compliance requirements.
We find private GPU capacity that fits the workload and negotiate the price with you.
We profile your model on that hardware, deploy it and commit to the performance we measured.
FAQ