Track GPU and compute spend

GPU hours by cluster, project and run, what they cost and how much budget is left. A scheduling agent can check the budget before it launches more GPUs.

An example

@ADoor Import this month's cloud bill export, and match the GPU lines to runs by tag.

ADoor Added 1,284 entries to Compute usage

Compute usage

ProjectGPU hoursCostBudget left
Pretraining 1.3B18,40041,2008,800
Evals1,1502,6002,400

What you can ask later

  • “Which project is closest to its budget?”
  • “What did run r-0412 cost in total?”
  • “Which runs cost more than 5,000 last month?”
  • “Which team’s spend grew most this quarter?”
  • “Chart GPU spend per week by project for the last quarter.”

Who can see what

  • Finance reads everything and edits budgets.
  • Each team reads its own usage.
  • An import script writes usage through the HTTP API.
  • A scheduling agent reads budgets before it launches a run.

How it compares with your cloud’s cost tools

Your cloud’s cost tools, or a platform like Vantage, pull billing data automatically, with recommendations and anomaly alerts.

Tying spend to your own runs and projects is where ADoor helps. Join usage to the training log with plain SQL, and let an agent check the budget before it starts a run.

How to start

  1. Connect ADoor to your assistant. ChatGPT, Claude, Grok and Cursor all connect through MCP.
  2. Tell it: “Set up compute spend: Compute usage (date, cluster, project, run, GPU hours, cost) and Budgets (project, period, amount).”
  3. Give each team read access to its own usage, and finance edit access to budgets.

Get your account

Questions people ask

Does it pull my cloud bill automatically?

No. Import exports as CSV, or post them from a script through the HTTP API.

Can I join spend to runs?

Yes, with plain SQL across both lists.

Is ADoor free?

Yes. The free plan holds 3 tables and 100 rows. Pro is $10 a month per workspace, with 1 GB of storage and no per-GB charges.

Which assistants work with ADoor?

ChatGPT, Claude, Grok and Cursor connect through MCP. You sign in once per assistant, approve it, and choose what it can reach.

Updated October 20, 2026.

This page as Markdown