gpu watch create

Create a watch rule

Synopsis

Creates a rule that alerts you when the market moves on a GPU model you care about.

--condition accepts one of three kinds:

• price-under — alert when the per-GPU hourly price drops below --threshold
• in-stock — alert when the model is available again (takes no threshold)
• count-at-least — alert when at least --threshold units are available

A rule targets a GPU model, not a specific listing. --region and --capacity-class narrow it; omit them and the rule watches every region and both capacity classes. Regions are canonical GPU.ai regions (us-west, us-east, …), not cloud-provider zone names — gpu pricing shows the set in use.

--gpu-model takes a model id, not a generation: h100_sxm, not H100. Alerts match the catalog exactly, so a rule for a model we do not list could never fire — the API refuses one rather than storing it, and names the closest ids when it can. Spelling is forgiving (H100-SXM and h100 sxm both resolve to h100_sxm); gpu gpu-types lists the models currently on offer — a model that is sold out right now is still watchable, which is rather the point.

Alerts fire on the edge: the moment the condition becomes true, not repeatedly for as long as it stays true. To alert a second time, three things must all be true — the condition has genuinely cleared, that clear held for at least five minutes, and a quiet period has passed since the last alert (6 hours for stock rules, 4 hours for price rules). A clear that is contradicted before those five minutes are up is discarded, so a single flaky poll cannot earn you a second email.

In practice that caps a single rule at four alerts a day for stock and six for price, and only when something real happened. If a rule is still noisier than you want, delete it — deletion takes effect within seconds, and a deleted rule can never mail you again.

Delivery goes to your organization's registered webhook endpoints and by email to you, the rule's creator. --webhook-endpoint routes this one rule to a specific endpoint instead. That flag takes the endpoint id printed by gpu webhooks list — there is deliberately no flag that accepts a URL, because a rule that carried its own destination would let anyone who can create a rule point our servers at an arbitrary address.

This creates a notification, not a purchase. Nothing here launches an instance or reserves capacity; see gpu autolaunch create for that, which is a separate resource with its own spend consent.

gpu watch create [flags]

Examples

# Alert me anywhere an H100 SXM drops under $4.00 per GPU-hour
gpu watch create --gpu-model h100_sxm --condition price-under --threshold 4.00

# Narrow it to one region and one capacity class
gpu watch create --gpu-model h100_sxm \
  --condition price-under --threshold 3.50 \
  --region us-west --capacity-class secure

# Tell me when B200 capacity comes back at all
gpu watch create --gpu-model b200 --condition in-stock

# Tell me when at least 8 units are free (enough for a full node)
gpu watch create --gpu-model h100_sxm --condition count-at-least --threshold 8

# Route this rule's alerts to one specific registered endpoint (an id, not a URL)
gpu watch create --gpu-model h100_sxm --condition in-stock \
  --webhook-endpoint 3f9c1c2a-5b6d-4a71-9c3e-2f8a1d0b7e45

Options

      --capacity-class string                Narrow the rule to one capacity class: secure or community (omit to watch both)
      --condition string                     What to alert on (required). One of: price-under, in-stock, count-at-least
      --disabled                             Create the rule parked: stored but not evaluated until you enable it
      --gpu-model string                     GPU model id to watch, e.g. h100_sxm — a model, not a generation (required)
  -h, --help                                 help for create
      --region string                        Canonical region to narrow the rule to, e.g. us-west (omit to watch every region)
      --threshold float                      The number the condition compares against: a per-GPU hourly price for price-under, a unit count for count-at-least. Not used by in-stock
      --webhook-endpoint endpoint-id         Route this rule's alerts to one specific webhook endpoint, given as the endpoint-id shown by 'gpu webhooks list'. Not a URL: endpoints are registered once with 'gpu webhooks create' and referenced by id thereafter

Limits

Your organization can have up to 50 enabled watch rules. Disabled rules do not count against that ceiling, so parking a rule with --disabled costs nothing. Exceeding the limit returns a validation error naming the number.

Options inherited from parent commands

      --api-base string   API base URL (override with GPUAI_API_BASE env) (default "https://api.gpu.ai/v1")
      --debug             Enable debug logging to stderr
  -o, --output string     Output format: table|json (default table on TTY, json otherwise)

SEE ALSO

• gpu watch - Watch GPU price and availability, and get alerted when it moves