GPU pricing
What Tensorant compute charges for each GPU an hour, for training and for serving, what is included, and when charging starts and ends.
Tensorant compute runs your training runs, test sessions and published models on GPUs Tensorant provides, so there is no GPU account to set up. It is charged per second from credits you buy in the console. This page lists the price of every GPU and says exactly what is counted.
Prices
Tensorant compute lists an hourly price for each GPU: one for training runs and test sessions, and one for published models while their GPUs serve. CPU and memory are included in the GPU price. The prices below are for one GPU, and are read again from the current price list every hour.
| GPU | Memory per GPU | Training per GPU an hour | Serving per GPU an hour |
|---|---|---|---|
| L4 | 24 GB | $1.76 | $1.76 |
| A10 | 24 GB | $2.12 | $2.12 |
| L40S | 48 GB | $3.94 | $3.94 |
| A100 40GB | 40 GB | $4.12 | $4.12 |
| A100 80GB | 80 GB | $5.06 | $5.06 |
| RTX PRO 6000 | 96 GB | $5.24 | $5.24 |
| H100 80GB | 80 GB | $6.80 | $6.80 |
| H200 | 141 GB | $7.51 | $7.51 |
| B200 | 180 GB | $9.56 | $9.56 |
A machine with more than one GPU costs the number of GPUs times the price of one: a 4-GPU machine of one type costs four times its price per GPU. A machine has 1, 2, 4 or 8 GPUs, up to the most its GPU comes in; the console offers only those counts. The console shows the price an hour for the GPU and count you choose before you start anything, and only the price that applies: a training run's or test session's when you train or test, a published model's when you publish. A published model that wakes on request is charged for each worker while it runs, at the price the console shows when you publish it.
How it is charged
- When charging starts and ends. Charging starts when a GPU is assigned to your machine, including the time to load the image, and ends when the machine stops. It is counted per second. Time spent loading the model is part of it.
- Not charged for our failures. If a machine fails to start or is taken away for reasons on our side, it is not charged.
- No GPU free. If no GPU of the kind is free, nothing is charged: the start is refused, or the record says Waiting for a GPU until one is assigned. Its time limit starts when the machine runs. If no GPU becomes free in time, usually about 20 minutes, it ends with nothing charged.
- Tax. Prices here are for compute. Tax is added when you buy credits, not when you use them. See Credits.
- Your hourly price limit still applies to every run and deployment, and you approve the estimate before anything paid starts.
The training-token fee
GPU time is what a machine costs. Every training run, trials included, also pays a fee for the tokens it trains on, by the size of the base model, charged from your credits as it trains. The launch confirmation shows about what it will be and the most it can be. See Training-token fee.
Limits
- A Tensorant compute run, test session or machine lasts at most 24 hours. A training run longer than that is refused when you start it. A published model that stays on is moved to a fresh machine before then without a gap you need to handle.
- If the infrastructure interrupts a machine, Tensorant restarts the work and refunds the time that was lost to credits.
- Until Tensorant reviews your organization, usually within a working day, Tensorant compute runs 1 GPU at a time across it. See One GPU at a time until review. Your own GPU accounts are not affected.
- Once reviewed, an organization can have 8 GPUs running at once by default. Tensorant can change this for your organization: write to support.
Your own GPU account instead
With the Own compute subscription, you can also connect your own RunPod, Verda, Lambda, AWS, Azure or Google Cloud account and pay that provider directly for its GPUs at its prices. The training-token fee of runs there still comes from your credits. See Connections.
Need a hand? Visit troubleshooting