# Find GPU capacity at a price and get a machine running

> Find GPU capacity at a price and get a machine running: 0 companies fully do it, 13 with a limit, 22 carry the tag. Every listing in GPU & serverless compute with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/gpu-compute.find-and-provision/
- Markdown: https://www.anchorterminal.com/niches/gpu-compute.find-and-provision/index.md (~2,050 tokens)
- Slim: https://www.anchorterminal.com/niches/gpu-compute.find-and-provision/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/gpu-compute.find-and-provision/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

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

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 13 | 22 | 14 | 10 |

## What counts as done

A list or availability call returns the complete set of has or regions with capacity, the launch call either succeeds or returns a structured capacity error, and the instance becomes SSH-reachable or Ready.

Measured, for our own tests: the set comes from one list or availability call, the launch result within 30 seconds, and the instance is SSH-reachable or Ready within 15 minutes.

How agents ask for it: find me the cheapest available A100 80GB right now; is an H100 available anywhere, launch one if so; keep trying until a GPU frees up

## Does the job, with a limit

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.

- **Beam (Smartshare, Inc.)**: beam machine list shows live serverless availability and beam pool has lists provider, region, resources and cost with a --max-spend cap; beam machine reserve --yes takes the cheapest offer. No provisioning time or capacity error format is documented.
- **CoreWeave (CoreWeave, Inc.)**: The beta Capacity Finder API (POST /v1beta1/capacity/availability:get) ranks zones for spot node pools only, from cached data that neither reserves nor guarantees; on-demand capacity is a quota question with sales. Node pool conditions report Sufficient or partial capacity after the fact.
- **Crusoe Cloud (Crusoe Technologies LLC)**: GET /v1/capacities?product_name=a100.8x&location=... reports availability per type and location before a create, and creates return an operation to poll. Requests are HMAC-signed, no idempotency key exists, and the docs say quotas do not guarantee capacity.
- **GPUsMarket (Tyga.Cloud Ltd)**: GET /marketplace/search takes gpu, maxPrice and minVram, /gpu/cheapest?vram=24 picks the cheapest, and /gpu/deploy claims about 30 s provisioning. 502 means the host is offline and the agent rents another; completeness of search results is not documented.
- **Hyperbolic (Hyperbolic Labs, Inc.)**: GET /v2/on-demand/rentals/options lists rentable configurations and the docs say capacity changes hour to hour; sold-out alerts are email only and personal accounts only. Create has no idempotency key, and an instance can sit pending up to the 3 hour provisioning timeout before auto-termination.
- **Lambda Cloud (Lambda, Inc.)**: GET /instance-types returns regions_with_capacity_available and launch fails with the structured code instance-operations/launch/insufficient-capacity. Launch has no idempotency key, 429s arrive after five launches a minute, and users report capacity gone for days with scripting retries the only remedy (acc-30).
- **Massed Compute (Massed Compute, Inc.)**: GET /gpu-inventory lists available GPU configurations before launch, and launch returns 402 until credit and auto-recharge are set. No idempotency key, no documented capacity error shape, and no timing for reaching SSH.
- **Nebius AI Cloud (Nebius B.V.)**: list-platforms shows the platforms and presets a project can use and X-Idempotency-Key makes a retried create safe, but there is no capacity count per region, quota overruns fail with QuotaExceeded, and creates return an operation to poll.
- **Runpod (Runpod, Inc.)**: List GPU types and data centres return pricing with availability (HIGH to NONE) before a Pod create, and 80 GB cards now arrive in minutes (acc-02). A Pod takes one GPU type so fallback is the caller's loop, and there is no idempotency key.
- **Thunder Compute (Thunder GPU, Inc.)**: GET availability and GET /v2/specs list GPU types and current availability before create, and creation fails when a type is sold out. No idempotency key, no capacity error shape documented, and no timing to SSH.
- **Together GPU Clusters (Together Computer, Inc.)**: GET /v1/compute/regions lists regions, instance types and driver versions but no prices or capacity counts, create fails with a structured error when credits are short, and there is no idempotency key; Ready arrives in 10 minutes to an hour.
- **Vast.ai (Vast.ai Inc.)**: POST /api/v0/bundles filters has by gpu_name, num_gpus, reliability and price, but repeated identical queries returned a random 10 to 30 percent of matching has (acc-24), unknown fields silently dropped the gpu_name filter (acc-26), and a reservation hung at provisioning four times (acc-19).
- **Verda (DataCrunch Oy)**: GET /v1/instance-availability lists what each location can launch and POST /v1/instances returns 503 service_unavailable when it cannot, which is a structured capacity error. There is no idempotency key, so a retried launch can start a second billed instance.

- not this job: Baseten (Baseten Labs, Inc.): Fixed instance-type catalogue with no availability call, offer search or price comparison; Baseten places replicas across its own multi-cloud capacity.
- not this job: Cerebrium (Cerebrium Inc.): No availability call or offer search; capacity is a static GPU-by-region table and a preference list of up to five GPU types the platform falls back through.
- not this job: Cloud Run GPUs (Google LLC): Two GPU types in a fixed region list with a default quota of 3 GPUs a region; no availability call, offer search or price comparison.
- not this job: Hugging Face Inference Endpoints (Hugging Face, Inc.): Fixed instance tiles per cloud and region, greyed out when unavailable, and a provider list with status; no cross-vendor search, price cap or capacity error for a VM.
- not this job: Koyeb (Koyeb SAS): Fixed GPU instance types in specific regions with no availability call or price search.
- not this job: Modal (Modal Labs, Inc.): No availability call or offer search; the agent passes a GPU priority list and Modal schedules on whatever is free, with region selection as the only placement knob.
- not this job: Northflank (Northflank Ltd): GPU models are a dropdown per region with VRAM and price; no availability call, cross-region search or capacity error is documented.
- not this job: Replicate Deployments (Replicate, LLC): Hardware is a fixed SKU list chosen per deployment; no availability call, offer search or capacity error.

## What stops agents finishing

- search endpoint returns a random 10 to 30 percent of matching has, so the cheapest offer is often missing
- provisioning hangs with no error and the deposit is spent on nothing
- sold out for days and no queue or notification, only retrying
- launch has no idempotency key, so a retried call after a timeout starts a second billed machine
- availability claimed in marketing but on-demand refused after sign-up

## Around it

Before: gpu-compute.agent-provisioned-compute. After: gpu-compute.training-run, gpu-compute.self-hosted-llm-server, gpu-compute.batch-job.

## Where this comes from

Weight 14 of 100 in GPU & serverless compute: 10 of 48 accounts are about getting a machine at all: capacity gone for days, provisioning that hangs, search results that are a random subset, availability that changed between list and launch.

As data: https://www.anchorterminal.com/api/v1/use-cases/gpu-compute.find-and-provision.json, or the MCP tool get_use_case with id gpu-compute.find-and-provision. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
