Niches · GPU & serverless compute

Let an agent rent and pay for GPU time under a budget

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

0companies, full fit
19companies, fit with a limit
23companies carry the tag
8share of the category's work
6first-hand accounts

What counts as done

The agent creates and destroys the resource with a key that cannot touch billing or other resources, total spend stays under the cap enforced by the platform or a hook rather than by the agent's own report, and nothing is left running after the task.

How agents ask for it: pick the right GPU for this job and rent it, stay under 20 dollars, spin up temporary inference for this task and shut it down after, the agent has its own wallet, make sure it cannot overspend

Does the job, with a limit 19

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Baseten B
Baseten Labs, Inc.
Team keys can be inference-only, metrics-only or scoped to one environment or model, and management keys cannot invoke models. The monthly budget only notifies; enforcement stops Model APIs, never dedicated deployments or training, and creates have no idempotency key. source
Beam C
Smartshare, Inc.
--max-spend caps what a reservation commits and Beam releases nodes before credit runs out, but max-spend is not a running budget and workspace tokens have no scopes or read-only mode (only the MCP ?readonly filter). No idempotency key on creates. source
Cerebrium C
Cerebrium Inc.
Service accounts carry an expiry and the API lists roles with their permissions, but no spend limit, budget or idempotency key appears in the docs, and max_replicas is the only platform-side cost cap. source
Cloud Run GPUs B
Google LLC
IAM roles such as Cloud Run Developer and Invoker scope a service account to deploy or call without billing rights, but a billing budget only notifies; stopping spend needs a user-built Pub/Sub function that disables billing, and there is no idempotency key on deploys. source
Crusoe Cloud B
Crusoe Technologies LLC
Project IAM has Admin, Editor and Reader roles and API keys are per user, so a Reader key cannot create. No spend cap was found, create has no idempotency key, and the account needs a non-prepaid credit card. source
GPUsMarket E
Tyga.Cloud Ltd
An agent can sign up with one POST and get a gpu_ key without a browser, and a per-rental token separates inference from account rights. The card hold ($10 default, selectable to $100) is the only cap; the account key has no scopes and no idempotency key exists. source
Hugging Face Inference Endpoints B
Hugging Face, Inc.
Fine-grained tokens scope access to specific resources and are the recommended production token, and endpoints are private by default. The only cost cap is maxReplica, the account needs a card and credits first, and the management API has no idempotency key. source
Koyeb D
Koyeb SAS
Billing alerts email at 80 and 100 percent of a budget threshold but do not stop usage, API tokens carry the user's full rights with no scopes, and only dry_run stands in for an idempotency key. source
Lambda Cloud C
Lambda, Inc.
Identities get API keys and workspace roles, with Instance Reader for read-only and Billing Admin kept separate, so the agent's key can be narrow. No spend cap or idempotency key exists, and an account needs a card before launching. source
Massed Compute E
Massed Compute, Inc.
Read-only API tokens exist and are recommended for AI assistants, and team Members never see billing. There is no spend cap beyond the deposit and auto-recharge threshold, launch has no idempotency key, and the owner must add a card first. source
Modal B
Modal Labs, Inc.
Workspace and Environment budgets are hard monthly caps and a spend limit stops workloads that would incur charges, enforced by Modal. Tokens are not scoped below a Workspace in the pages read, and a payment method must be on file before anything runs. source
Nebius AI Cloud BB
Nebius B.V.
Service accounts with group-based IAM and X-Idempotency-Key on every create give a narrow, retry-safe key. No spend cap was found (spot pricing policies cap a price, not spend), and the account needs a browser signup with a $25 card charge. source
Northflank C
Northflank Ltd
API tokens inherit RBAC roles that can be restricted to projects and denied delete, so the key can be narrow. No spend cap or idempotency key exists, and the 30 October 2026 incident told customers to assume environment variables were decrypted (acc-45). source
Replicate Deployments B
Replicate, LLC
Multiple named tokens can be created and disabled but carry full account rights with no scopes; max_instances is the only platform-side cost cap, there is no idempotency key, and prepaid credit or a card is needed first. source
Runpod C
Runpod, Inc.
Restricted keys scope access per Serverless endpoint and Read Only keys exist, and the account has a platform-side $80 an hour spend limit by default. The limit is per hour not per task, create has no idempotency key, and prepaid credit needs a card first. source
Thunder Compute D
Thunder GPU, Inc.
Named API tokens and workspace member roles exist and TNR_API_TOKEN runs headless, but no token scopes are documented, the credit alert only emails, threshold charges bill the card as usage grows, and create has no idempotency key. source
Together GPU Clusters C
Together Computer, Inc.
Running out of credits pauses an on-demand cluster at once and deletes it after a grace period, which is a platform-side stop, but there is no budget below the balance, the kubeconfig needs a project admin's key, and create has no idempotency key. source
Vast.ai C
Vast.ai Inc.
Scoped API keys take permission categories such as instance_write without billing_write, so the agent's key can be narrow. There is no platform-side budget, so an agent's own accounting is the only limit (acc-25), no idempotency key exists, and signup needs a browser and a $5 deposit. source
Verda C
DataCrunch Oy
Cloud API credentials carry one scope, cloud-api-v1, with no read-only form, so the agent's key can do everything the account can. The prepaid balance is the only cap and at zero instances are discontinued and volumes deleted; create has no idempotency key. source

Doesn't do this job 2

Listings in the category that do something else, with the reason.

ListingWhy
CoreWeave C
CoreWeave, Inc.
IAM has Viewer and Admin roles per service and expiring tokens, but an agent cannot open an account (sales approval), no spend cap or idempotency key was found, and the MCP kubectl_apply tool ignores dry_run and applies the manifest. source
Hyperbolic E
Hyperbolic Labs, Inc.
API keys have no scopes and the same key can call DELETE /v2/users/me; the only cap is the prepaid balance, which must cover an hour of all instances, and the account needs a $5 Stripe deposit before renting. No idempotency key on create. source

What stops agents finishing

  • only budget control is the agent's own accounting, no platform-side cap or approval hook
  • API key has full account rights including billing and delete-account, no read-only or per-endpoint scope
  • no idempotency key on create, so a retried launch bills twice
  • signup needs a browser, email verification and a card or 5 dollar deposit, so the agent cannot open an account
  • secrets stored on the platform decrypted in a breach and the customer learns from Hacker News

Around it

Before: agent-harnesses.scheduled-job, agent-harnesses.cloud-delegate, decision-models.gate-action. After: gpu-compute.find-and-provision, gpu-compute.batch-job.

Where this comes from

Weight 8 of 100 in GPU & serverless compute: 6 of 48 accounts touch an agent or automation holding the key: an agent skill picking RunPod hardware, a proposal for external spend authority because agents self-report budgets, a key leaked in deploy logs, a key left in an environment variable through a breach, a worker that drained a balance. Few, but every other use case runs through it when an agent is the operator.

The same as data: /api/v1/use-cases/gpu-compute.agent-provisioned-compute.json, or the MCP tool get_use_case with id gpu-compute.agent-provisioned-compute. Other jobs in this category: /api/v1/niches?category=gpu-compute.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.