# Rent a GPU machine or cluster for a training or fine-tuning run

> Rent a GPU machine or cluster for a training or fine-tuning run: 1 companies fully do it, 15 with a limit, 26 carry the tag. Every listing in GPU & serverless compute with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/gpu-compute.training-run/
- Markdown: https://www.anchorterminal.com/niches/gpu-compute.training-run/index.md (~2,100 tokens)
- Slim: https://www.anchorterminal.com/niches/gpu-compute.training-run/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/gpu-compute.training-run/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

One company's docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 1 | 15 | 26 | 16 | 8 |

## What counts as done

An instance of the requested type is reachable over SSH after the launch call, runs to completion or restarts from the last checkpoint after interruption, and is terminated with no further charge.

Measured, for our own tests: SSH-reachable within 15 minutes of the launch call, and terminated within one call.

How agents ask for it: fine-tune this 8B model on my data; give me an H100 box with SSH for the afternoon; I need 8 H200s for a week

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Hyperbolic (Hyperbolic Labs, Inc.)**: 1x and 8x H100, H200, B200 and B300 VMs with InfiniBand or RoCE in listed regions, SSH root in under 5 minutes (10 to 15 in eu-north-9), billed hourly from ready, and terminated by one API call. Nothing is charged for an instance that never becomes ready.

## 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.

- **Baseten (Baseten Labs, Inc.)**: Training Jobs run a container on H100 or H200 with node_count for InfiniBand multi-node, checkpoints synced and resumable, and SSH into the running container. No other GPU types for training and no provisioning time published.
- **Beam (Smartshare, Inc.)**: beam machine reserve or beam pool scale leases H100 and other nodes by the hour and a function or Pod runs the script on them; VMs are CPU only. Reserved nodes renew hourly past the TTL until released, and --nodes 2 gives separate machines with no interconnect stated.
- **CoreWeave (CoreWeave, Inc.)**: A NodePool with instanceType and targetNodes adds HGX nodes to a CKS cluster, with SUNK for Slurm and spot pools giving a seven-minute preemption window. No SSH VM, and the organisation must be approved by sales and activated by email before any token exists.
- **Crusoe Cloud (Crusoe Technologies LLC)**: crusoe compute vms create launches types such as h100-80gb-sxm-ib.8x with 3,200 Gbps InfiniBand partitions, CMK and Managed Slurm give clusters, and a stopped VM is not billed for compute. Provisioning time is unpublished, quotas are not capacity, and compute needs a non-prepaid card.
- **GPUsMarket (Tyga.Cloud Ltd)**: POST /api/v1/gpu/deploy with instanceType docker or ssh gives a container with SSH in about 30 s on cloud GPUs, destroyed with DELETE /pods/:id. Consumer cards from individual hosts, no clusters, and a $10 to $100 card hold per rental.
- **Lambda Cloud (Lambda, Inc.)**: 1x to 8x H100 SXM, A100 and B200 VMs launch from the API with an SSH key and terminate in one call; a user fine-tuned Llama 3.1 8B this way (acc-28). 8x H200 nodes are almost never available (acc-27), capacity has run out for days (acc-30), and 1-Click Clusters are weekly reservations approved by email.
- **Massed Compute (Massed Compute, Inc.)**: POST /instance/launch starts a VM from gpu-inventory and images lists with SSH keys, debited per minute and ended by /instance/terminate. Single machines only (clusters by quote), SXM-only images do not run on PCIe cards, and no provisioning time is published.
- **Modal (Modal Labs, Inc.)**: A Function with gpu="H100:8" and a 24 hour timeout runs a training script, and @clustered gives up to 32 nodes with 3,200 Gbps RDMA. There is no SSH VM, so interactive work goes through modal shell, and checkpoints must go to a Volume.
- **Nebius AI Cloud (Nebius B.V.)**: 8-GPU H100, H200, B200 and B300 VMs join InfiniBand fabrics at 400 Gbps a GPU, with Soperator Slurm and Kubernetes on top, and Serverless AI jobs with a timeout and bucket mounts. Preemptible VMs get a 60 s SIGTERM, no provisioning time is published, and signup charges $25 to a card.
- **Northflank (Northflank Ltd)**: A job with gpu.enabled, gpuType and gpuCount up to 8 runs to completion on Northflank's cloud, and persistent volumes hold checkpoints. No SSH VM, no multi-node interconnect, and GPU deploys need at least $50 of credit.
- **Runpod (Runpod, Inc.)**: Pods give SSH on Secure or Community Cloud and Instant Clusters give multi-node capacity through the REST API; one user found it the best developer experience but saw smaller instances dropped (acc-01), another found any named card above 20 GB hard to get (acc-03).
- **Thunder Compute (Thunder GPU, Inc.)**: POST /instances/create gives an A6000, L40, A100 80 GB or H100 instance with SSH, billed per minute and deleted in one call. Single machines only, and there is no stop: pausing means snapshot, delete and restore, with snapshot storage billing meanwhile.
- **Together GPU Clusters (Together Computer, Inc.)**: Multi-node H100, H200, B200 and B300 clusters with 3.2 Tb/s InfiniBand, Kubernetes or Slurm, created and deleted headlessly with tg clusters. Provisioning takes 10 minutes to an hour, past the 15 minute mark, and preemptible nodes get a five-minute drain.
- **Vast.ai (Vast.ai Inc.)**: Interruptible and on-demand instances with SSH from consumer cards to H200 at per-second rates, and a $10 budget covered small runs (acc-22). Provisioning took 25 minutes (acc-20) or hung after a deposit (acc-19), RTX 5090s ran at 80 percent clock with nothing on the listing (acc-23), and there are no clusters (acc-01).
- **Verda (DataCrunch Oy)**: GPU instances with SSH and a startup script in prepaid 10-minute increments, and Instant Clusters of 16 to 128 H200, B200 or B300 GPUs over InfiniBand that take about 20 minutes to start. Spot instances are evicted without warning, cluster availability was moved to dedicated customers, and a zero balance deletes volumes.

- not this job: Cerebrium (Cerebrium Inc.): Serverless apps and cerebrium run containers only; no SSH-reachable machine, no multi-node, and runs are bounded by response_grace_period up to 12 hours.
- not this job: Cloud Run GPUs (Google LLC): Containers only, one GPU per instance, no SSH and no multi-node; GPU job tasks are capped at 12 hours in preview.
- not this job: Hugging Face Inference Endpoints (Hugging Face, Inc.): Inference endpoints only; no VM, SSH, job or cluster product.
- not this job: Koyeb (Koyeb SAS): Services and workers run containers from git or an image; there is no SSH-reachable VM, no run-to-completion job, and no multi-node cluster.
- not this job: Replicate Deployments (Replicate, LLC): No VM, SSH or cluster; the guides send users to Lambda or Brev for a GPU machine to package a model before pushing it.

## What stops agents finishing

- no multi-node clusters, so a run bigger than one box means syncing between separate machines
- interruptible instance pre-empted without notice and no checkpoint
- the GPU type is out of capacity for days and the only remedy is retrying the launch call
- stopped instance keeps billing for storage, and a zero balance deletes the volume

## Around it

Before: gpu-compute.find-and-provision, open-weights.fine-tune-base. After: gpu-compute.self-hosted-llm-server, gpu-compute.serverless-model-endpoint, open-weights.big-model-on-big-box.

## Where this comes from

Weight 16 of 100 in GPU & serverless compute: 8 of 48 accounts describe a training or fine-tuning run on a rented machine; most are on marketplaces or VM clouds, one on a cluster-less marketplace that made multi-node work not worth it.

As data: https://www.anchorterminal.com/api/v1/use-cases/gpu-compute.training-run.json, or the MCP tool get_use_case with id gpu-compute.training-run. 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
