SwarmDoGPU — rent GPUs by the second

How it works

Aggregated supply, scored for reliability, billed by the second.

Instances are provider-agnostic — the catalog pools capacity from multiple sources and presents it as one list with one bill, so you pick on price and reliability rather than on who happens to own the rack.

01 · pick

One catalog

Filter by GPU model, or by on-demand only, and sort by price. Every offer shows VRAM, vCPU, RAM, SSD and a measured uptime score before you commit to anything.

02 · launch

One click

Instances come up provisioned with SSH and Jupyter ready. No image plumbing, no per-provider console to learn.

03 · pay

One bill

Per-second billing across every provider in the pool, with a hard spend cap set on each instance. No commitments and no idle burn.

Interruptible or on-demand. Cheaper interruptible capacity sits alongside on-demand in the same catalog, labelled per offer. Take the discount when your job can be restarted; take on-demand when it can’t.

What’s in the pool

Datacentre cards and workstation cards, side by side.

Single cards through to multi-GPU nodes, across nine regions in North America, Europe and Asia. Availability and price move with the spot market — the catalog is the source of truth.

ClassTypical useSizing

H100SXM and 80GB HBM3 — large-model training and high-throughput servingup to ×8 nodes

A100 80GBPCIe and SXM4 — training, fine-tuning, and serving models that want the full 80GB×1 to ×4

L40SInference, rendering, and mixed graphics-plus-compute work×1 to ×2

RTX 4090Cost-efficient fine-tuning, batch jobs and sweeps×1 to ×2

RTX 3090The cheap end — experiments, dev boxes, long-running background work×1

Not sure which? The catalog sorts by price and shows an uptime score per offer, so the cheapest box that clears your reliability bar is usually one sort away. See what’s live now →

Bring your own workload

It doesn’t have to be ours, and it doesn’t have to be an LLM.

SwarmDoGPU is a general-purpose GPU service, not a SwarmDo accessory. Your code, your framework, your data. We happen to be a demanding customer of it ourselves, which is the useful part — the awkward edges get found by us first.

▲Training & fine-tuning

Jobs that want a big card for hours or days. Take interruptible capacity and checkpoint, or on-demand if a restart would hurt.

◉Inference & serving

Stand an endpoint up on the card that fits it — whatever the model is and whoever trained it. Open weights, your own, or off the shelf.

⇉Batch & parallel work

Sweeps, evals, offline processing, render-and-verify loops — the jobs that want throughput more than one very fast card.

⛁Anything else GPU-bound

It needn’t be AI at all. If it needs a GPU and you’d rather not own one, that’s the use case.

Our own models, as a worked example. SwarmDo-A1 and A2 are open-weight coding models that fit a single 80GB card, and their execution-verified patch selection is exactly the bursty, parallel, restartable job this pool is good at. They’re one tenant among many — not the boundary of what it runs.

The fine print

What to know before you rent.

SwarmDo would rather say the awkward part up front than have you find it at 2am.

Interruptible means interruptible. Discounted capacity can be reclaimed. Checkpoint your work, or pick an on-demand offer — both sit in the same catalog, labelled per offer.
Spot prices move. Rates and availability shift with the market and with the region. Any number you see quoted elsewhere — including on this page — is a snapshot; the live catalog is the source of truth.
Uptime scores are measured, not promised. The percentage on each offer describes how that capacity has behaved. It’s a signal to sort on, not an SLA.
You’ll need an account to rent. Browsing the catalog and prices is open; provisioning an instance requires signing in.

Open SwarmDoGPU → Our models The plugin

Questions, or a workload that doesn’t fit the catalog? maintainers@swarmdo.com — a human replies.