---
title: Compute
description: GPU inference pods and workers on ARGUS, the read surfaces for pods, planner decisions, pricing, strain, and cost, and the audited write surface for launching and stopping pods.
---

# Compute



Compute in ARGUS means the GPU inference pods and workers that run camera streams through the vision pipeline. The Operations Console COMPUTE page and the `argusctl compute` command family are two faces of the same ops API: both reflect live provider reality (running pods, drift, orphans, spend) so untracked GPU burn cannot hide.

## Read surfaces

All of these are normal authenticated reads. No operator gate, no tailnet.

| Surface                                                                  | Console                           | CLI                                                 |
| ------------------------------------------------------------------------ | --------------------------------- | --------------------------------------------------- |
| Worker list (provider, GPU, region, status, load, $/hr, pinned, drift)   | COMPUTE page, ALL WORKERS table   | `argusctl compute list`                             |
| Running pods plus the stores and cameras each serves                     | COMPUTE page, RUNNING PODS cards  | `argusctl compute served`                           |
| Planner decision log (A to F graded allocation decisions with rationale) | Planner decisions page            | `argusctl compute planner`                          |
| Live GPU pricing and availability catalog                                | COMPUTE page, PRICING REFERENCE   | `argusctl compute catalog`                          |
| Per-worker utilization, headroom, idle burn, recommendations             | (derived view)                    | `argusctl compute strain`                           |
| Spend: month to date, instantaneous burn, forecast, idle waste           | COMPUTE page, monthly spend block | `argusctl cost month` / `now` / `forecast` / `idle` |

The COMPUTE page also runs a live cross-check against the provider: pods that exist at the provider but have no worker row are flagged as orphans, and running orphans are rendered as danger rows because they bill while staying out of the tracked monthly total. The cross-check requires `RUNPOD_API_KEY` to be set on the ops API; without it a banner says drift and orphans are not shown.

## Write surface

All compute mutations go through the `/compute/ops/*` endpoints and share one privilege model:

* A personal access token (PAT) with the **`compute:write`** scope. The `admin-full` preset includes it. No tailnet is required, so writes work headlessly from anywhere.
* A required **`--reason`** (3 to 500 characters) on every CLI write; the console collects the equivalent through its step-up confirmation modal.
* A **per-operator rate limit** of 20 calls per 60 seconds. Exceeding it returns `429`.
* An **audit row** for every call: the platform-wide `argus.argus_audit_log` plus the compute domain audit table.
* **Step-up confirmation**: launches require a server-side `confirm: "LAUNCH"` value (the CLI sends it for you), and stop, resume, and terminate confirm the specific worker. Terminate is destructive and additionally requires `--yes` on the CLI.

Every CLI write also supports `--dry-run`, which prints the planned request (with header values redacted) without sending it.

## Guides

* [Launch a pod](/docs/compute/launch)
* [Stop, resume, or terminate a pod](/docs/compute/stop-resume-terminate)
* [Planner decisions and the pricing catalog](/docs/compute/planner-catalog)
* [GPU spend visibility](/docs/compute/costs)

Full command reference: [argusctl compute](/docs/cli/compute) and [argusctl cost](/docs/cli/cost).


---

For a semantic overview of all documentation, see [/sitemap.md](/sitemap.md)

For an index of all available documentation, see [/llms.txt](/llms.txt)

For agent-facing discovery, including API and MCP surfaces, see [/agents.md](/agents.md)