Skip to content
Credits

Credits

Add credits to turn on Tensorant compute and pay the training-token fee, watch your balance, and know what stops at $0 and what starts again.

Tensorant is pay as you go: credits pay for what you use. They pay for Tensorant compute, GPUs that Tensorant provides for your training runs, test sessions and published models, charged per second while a machine runs. They pay the training-token fee of every training run, trials included. And they pay for storage and egress past what is included, when your data is in Tensorant storage. See Pricing and billing.

If you use your own GPU account, with the Own compute subscription, that provider bills you directly for its GPUs. The training-token fee of runs there still comes from your credits.

Turn on Tensorant compute

There is one step: add credits. Your first payment turns Tensorant compute on, and you get an email saying it is ready. There is no request to send and nothing to wait for.

Until the first payment you cannot start work on Tensorant compute. Until Tensorant reviews your organization, usually within a working day, you can add up to $100 in credits in total, and Tensorant compute runs 1 GPU at a time. The Credits card says how much more you can add. Tensorant can set a higher limit for your organization: contact support. If Tensorant compute was declined or revoked for your organization, it stays off and you can still add credits for the training-token fee, storage and egress; contact support to discuss it.

Add credits

An owner adds credits under Plan and billing.

  1. Choose Add credits.
  2. Choose $10, $25, $50, $100 or $250, or enter your own amount from $10 to $2,000 for one payment.
  3. Pay at the secure checkout of our payment provider, Polar. Tax is added where it applies.

You receive the amount you chose in credits: tax does not buy credits. When the payment succeeds, the balance goes up and you get a confirmation email. Polar emails the receipt.

Credits never expire. They are not refundable, except where the law requires it. See the Terms of Service.

Your balance

The Credits card shows your balance, what your running compute costs each hour (the burn), and a history. The sidebar shows the balance too. Each row of the history is one of:

  • what ran: a training run, a test session or a published model, for each hour it used, with its GPU and count and the amount charged,
  • a training run's training-token fee, under the run's name, such as "Training-token fee, 4.2M tokens",
  • a top-up you paid for,
  • a grant or adjustment made by Tensorant.

Each GPU has an hourly price: one for training runs and test sessions, and one for published models while their GPUs serve. A machine costs that price times its number of GPUs, charged per second. Charging starts when a GPU is assigned to your machine, including the time to load the image, and ends when the machine stops. If a machine fails to start or is taken away for reasons on our side, it is not charged. Before a machine starts, Tensorant checks that your balance covers the first minutes of it, and of what is already running. A training run's start also needs credits for its training-token fee's estimate, on your own GPU account too. If the balance does not cover them, the start is refused and the message links to Add credits.

Warnings

  • When your balance covers about 1 hour at your current burn, you get an email and the console shows a banner. An organization with nothing running gets the same warning when the balance is low.
  • When compute stops because the balance reached $0, you get an email and the banner says so.

At $0

When the balance reaches $0:

  • Training runs and test sessions stop, on Tensorant compute, and training runs on your own GPU account too, since their training-token fee is charged from credits. They do not start again after a top-up. A run that is already exporting its adapter on your own GPU account finishes.
  • Published models on Tensorant compute answer 402 (insufficient_credits) until credits are added. Calls to them do not wake them.
  • After a top-up, an always-on published model starts again by itself. A wake-on-request model wakes on the next request.

Your data, projects, examples, adapters and settings are kept. Add credits to continue.

Who can do what

Owners add credits. Everyone in the organization can open Plan and billing to see the balance and history. Adding credits is done in the console only. An agent over MCP can read the balance with get_credits, and the CLI can start work with --provider tensorant.

Where your data lives

An organization that has not connected its own storage uses Tensorant storage: Tensorant keeps its data, separately for each organization. Storage past the 5 GB included, and egress past the monthly allowance, are charged from your credits: see Storage and egress. Every organization can connect its own S3 bucket instead, free, without a subscription, which Tensorant never meters or charges. Volume storage in your own GPU account is part of that account, so it comes with the Own compute subscription. You can switch to your own storage, but only while the organization has no projects. With your own GPU accounts and your own storage, everything stays in your cloud. See Connections.

Need a hand? Visit troubleshooting