Costs, limits, and billing

What remote compute costs on each provider, who pays, how hi asks before spending, and how every machine is guaranteed to stop.

What things cost

Some real numbers from testing hi:

Colab’s rates for other GPUs are shown as unmeasured; hi compute billing shows your balance and the rate you are using right now.

hi asks before spending

Starting paid hardware prints the price, the limit, and on Hugging Face the account that pays, then asks:

Run train.py on hf a10g-small ($1.00/h, billed to hifinab), stopping after 1h.
Start it? [y/N]

Colab CPU and Hugging Face hardware under $0.10 per hour start without asking. In scripts, where there is no terminal to ask in, hi stops with confirmation needed; rerun with --yes until you add --yes.

Every machine stops by itself

Every run and instance has a maximum lifetime set with --max. Give it in hours (2, 1.5) or with a unit (30m, 1h30m, 2d):

How the limit is enforced differs by provider:

• Hugging Face stops the job itself: --max becomes the job’s timeout. It stops even if your laptop is off.
• Colab cannot stop a runtime from inside it, so hi compute up starts a small background process on your laptop that stops the machine at --max. It survives closed terminals and sleep. It does not survive a reboot, so every hi compute ls also stops machines past their limit, and Colab ends sessions after 24 hours regardless.
• Colab runs pass --max to the Colab CLI as the run’s timeout, and the runtime is released when the run ends.
• RunPod has no built-in limit either. hi installs a watchdog on the pod that terminates it at --max, and runs the local watcher as well.

Choose the shortest --max that fits the work. You can always start another.

No time limit

--max none (or an empty answer to the menu’s lifetime question) starts a machine without a limit. It keeps running, and costing money, until you stop it, so hi shows a warning and asks you to confirm; in scripts, add --yes.

• On Hugging Face, the job gets a one-year timeout.
• On Colab, no watcher runs, and Colab still ends the session after 24 hours.
• hi compute ls shows no limit in the STOPS IN column.

Check what is running

hi compute ls

NAME   PROVIDER  HARDWARE    UP   STOPS IN
box    colab     G4          42m  3h18m
sweep  hf        a10g-small  5m   55m

Colab balance: 1795.89 compute units, currently using 8.90 units/h

Stop everything you started with hi compute stop --all.

Who pays on Hugging Face

Hugging Face Jobs need pre-paid credits on the account that pays. You may belong to organizations; hi picks the payer in this order:

1. --namespace <account> on the command.
2. The HI_HF_NAMESPACE environment variable.
3. The account saved with hi compute billing <account>.
4. Your own account, if it can pay.
5. Your only organization that can pay.

hi compute billing shows your accounts and which one hi bills:

hi compute billing

Hugging Face:
ACCOUNT    TYPE  CAN PAY  PLAN  BILLED
quantbert  user  no
hifinab    org   yes      team  <- hi bills this

Chosen by: automatic: your own account if it can pay, else your only organization that can

Choose one explicitly, or go back to automatic:

hi compute billing hifinab
hi compute billing --clear

Hugging Face reports whether an account can pay, but not its credit balance; check that at huggingface.co/settings/billing. If a job is refused with Hugging Face Jobs need pre-paid credits, add credits or choose an account that can pay.