How Valkyrie’s prepaid wallet, per-use pricing, and monthly spend limits work.
Valkyrie is prepaid and billed per use. Each workspace
has a wallet with a balance, and the workspace’s fine-tuning jobs and deployments
draw against it. There are no seats. You manage all of this on the
Wallet screen, and the active workspace’s balance is always
shown in the Balance chip in the top bar.
The balance goes down as your models run and back up when credit is added. The
Wallet screen shows the current balance and the full history of
charges and top-ups, so you can always see where the money went.Every charge goes to the wallet of the workspace that owns the fine-tune or
endpoint. Switching workspaces switches the wallet you see. When a workspace’s
balance runs out, its running endpoints pause, the same as when a
spend limit is reached.
Owners, Admins, and Billing members can top up the workspace’s wallet by card:
1
Open Wallet, then Add Funds
In the dashboard, open Wallet in the sidebar and click Add Funds.
2
Pick or enter an amount
Choose $10, $25, $50, or $100, or enter a custom amount in USD. A single top-up
is between $5 and $500.
Add funds: preset amounts or a custom amount.
3
Pay on Stripe's secure checkout
Click Continue to payment. You’re redirected to Stripe’s hosted checkout
page to pay by card. Valkyrie never sees or stores your card details.
4
Your balance updates automatically
Once Stripe confirms the payment, Valkyrie credits the wallet and you’re returned
to Wallet with Payment successful. This usually takes a few seconds, with no
need to refresh or contact support. If you cancel on Stripe’s page, you’re
returned with Checkout canceled and nothing is charged.
Confirmation shown after a successful top-up.
If you closed the tab or your connection dropped right after paying, don’t worry:
the credit isn’t tied to that browser session. It lands as soon as Stripe confirms
the payment, whether or not you’re still looking at the page.
Need a larger amount, invoicing, or another custom arrangement? Use Request more
credit on the same screen.
To try the top-up flow without a real payment, use Valkyrie’s development
environment at valkyrie.dev.azumo.com, where
Stripe runs in test mode. Test payments add test credit to that environment only.
1
Sign in to the development environment
Sign in at valkyrie.dev.azumo.com as an Owner,
Admin, or Billing member of the workspace you want to top up.
2
Start a top-up
Open Wallet, click Add Funds, pick an amount, and click Continue to
payment. Stripe’s checkout page is marked as a test or sandbox page.
3
Pay with a test card
Enter one of the test cards below, any future expiry date (for example 12/34),
any 3-digit CVC, and any name and postal code.
4
Check the result
Back on Wallet, the balance goes up by the amount you paid, and the top-up
appears as a credit under Resource Charges.
Card number
What happens
4242 4242 4242 4242
The payment succeeds.
4000 0025 0000 3155
Stripe asks for authentication; complete it to succeed.
4000 0000 0000 9995
Declined for insufficient funds. Nothing is credited.
4000 0000 0000 0002
Declined. Nothing is credited.
These are Stripe’s public test card numbers. See Stripe’s
testing guide for more.
Test cards work only in the development environment. On
valkyrieapp.azumo.com, Stripe charges real
cards, and test card numbers are declined. Never enter a real card in the
development environment.
Fine-tuning jobs: GPU time for the training run, billed per minute while it
trains.
Running endpoints: GPU time while an endpoint is up, whether or not it serves
requests.
Paused endpoints: the kept disk, while the endpoint is paused. See
Paused endpoints.
Inference requests themselves aren’t charged per token or per request.Before you deploy, the deploy form shows an estimate for the machine Valkyrie would
pick: the hourly cost split into GPU and storage, how long it should take to get
ready and what that costs, and what the disk costs per day while paused. The deploy
page for a fine-tuned version also shows the Cost while running per hour.
A running endpoint holds a GPU even when it isn’t receiving requests. Endpoints
that serve no requests for 2 hours are stopped automatically, unless they’re
always on. Use stop/resume or a schedule to avoid
paying for idle time in between. See Deploy a model.
New fine-tune estimates the
run’s time and cost as a range, and shows the Balance after the run. To start,
the wallet must hold the estimate’s high end plus a 20% buffer. When it
doesn’t, Start fine-tune is disabled with the amount needed and an Add
funds link. The AI assistant’s card does the same, and the API and MCP refuse
the job with HTTP 402 and the error balance_below_estimate.Runs without an estimate, such as a model outside Ready to fine-tune, only need
a positive balance.
Pausing an endpoint stops its GPU but keeps the machine and its disk, so Resume
is fast. The kept disk is billed while paused, usually a small fraction of the
running price. The endpoint’s page shows the daily disk cost, how much has been
billed so far, and when the pause ends.An endpoint paused for more than 72 hours is released to stop its storage
charges, and shows as Terminated. It keeps its URL and settings, and resuming
it rents a fresh machine.
Interruptible capacity is cheaper GPU time that can be taken back. Valkyrie bids
just above the lowest accepted price, and never above the regular on-demand price.
When someone outbids it, the machine is stopped.
Endpoints: where interruptible capacity is offered, tick Interruptible
(cheaper bid box) under the deploy form’s advanced settings, or pass
capacity: "interruptible" from the API or MCP. The endpoint’s page shows the bid
it was rented at. When it’s outbid, the endpoint shows Interrupted (outbid): it
didn’t fail, and Restart on a new instance brings it back on the same URL and
API keys. An always-on endpoint is brought back
automatically.
Fine-tuning jobs: tick Interruptible (cheaper bid box) on
New fine-tune or on the
AI assistant’s fine-tune card, or
pass capacity: "interruptible" from the API or MCP. When outbid, the job
resumes from its last saved checkpoint on a new machine, up to 3 times, and the
steps since that checkpoint are trained again. After that it fails, and
Retry continues from the last checkpoint. The time and cost estimate and the
wallet check use on-demand prices.
Use interruptible capacity for work that can wait or restart, not for an endpoint
that must always answer.
Owners, Admins, and Billing members can set monthly spend caps on the
Wallet screen: one for the whole workspace and one
per team. Caps work on top of your balance: Valkyrie never spends more than the
wallet holds, even when a cap is higher.
A team’s cap can’t be higher than the workspace’s. If you lower the workspace
cap below an existing team cap, Valkyrie doesn’t apply it or shrink the team cap
silently. It lists the conflicting teams, you edit their caps, and all the new
values save together.
Caps reset on the 1st of each month (UTC).
Owners and Billing members get an email when spend reaches 80% and 100% of a cap.
When spend reaches a cap, the running fine-tunes and endpoints it covers pause,
and nothing new can start under it. Starting one is refused before any GPU is
rented, with a message giving the cap and this month’s spend. A paused endpoint
can be resumed once there is room under the cap again.
Work owned by a team counts against the team’s cap and the workspace’s. Work
owned by a person or by the whole workspace counts against the workspace’s cap
only. See Who owns new work.
Being on a team doesn’t let you spend; only your workspace role does.
