osFoundry / Platform
Cloud GPU
Rent A100 / H100 / L40 instances directly from the platform. Pay per hour, no contracts.
What it is
Cloud GPU lets you rent NVIDIA accelerator instances — A100, H100, and L40-class hardware — directly inside osFoundry and pay for them by the hour, with no contracts and no reserved-capacity commitments. Instead of opening a cloud account, requesting GPU quota, and wiring up drivers and CUDA yourself, you open the instance picker, choose a GPU by VRAM and region, and the platform provisions it for you. When you are done, you release it and billing stops.
Inside osFoundry the same rented GPU backs three workflows without any extra setup. You can run managed fine-tuning jobs against it, spin up a GPU-powered notebook for experimentation, or deploy a model as an auto-scaling endpoint that provisions cloud GPU on demand. The instance picker pulls prices live from the server rather than from a hardcoded table, so the per-hour rate you see — alongside GPU model, memory, vCPU, RAM, and an estimated cold-start time — is the rate you are actually charged. If real prices cannot be loaded, the platform shows nothing to select rather than quoting a number it will not honor.
This matters because GPU time is expensive and easy to leave running. Per-hour billing that begins at provisioning and ends at release, with no minimum term, keeps a five-minute experiment cheap and a forgotten instance's cost visible. Because osFoundry is local-first, small jobs stay free on your own device GPU, and you reach for a rented A100 or H100 only when a model genuinely needs more VRAM than your machine has. It is the same posture as the rest of the workspace: your keys, your data where possible, and paid infrastructure only when you choose it.
Key capabilities
- Rent A100, H100, and L40-class NVIDIA GPU instances on demand — no quota requests, no procurement, no reserved-capacity contracts
- Pay by the hour with billing that starts when the instance is provisioned and stops when you release it; nothing to cancel and no minimum term
- Pick your instance from a live, server-priced list showing GPU model, VRAM, vCPU, RAM, region, and estimated cold-start time before you commit
- Use the same rented GPU for fine-tuning jobs, Jupyter-style notebooks, and auto-scaling model endpoints, all from one workspace
- Keep small jobs free on your own hardware — local device GPU runs at no cost, and you reach for cloud GPU only when a model needs more VRAM than you have
Use cases
- ML engineer: fine-tune a 13B model that won't fit on a laptop by renting an H100 for the length of the training run, then releasing it the moment the job finishes
- Researcher: open a GPU notebook on an A100 for an afternoon of experimentation and pay only for the hours the instance is up, with no account or quota to set up first
- Product team: deploy a model as an auto-scaling endpoint on L40-class GPUs, letting the platform provision and bill capacity per hour as request volume rises and falls
Frequently asked questions
Which GPUs can I rent and how much do they cost?
You can rent A100, H100, and L40-class NVIDIA instances. Prices are per hour and are served live inside the instance picker, which shows each GPU's model, VRAM, vCPU, RAM, region, and estimated cold-start time. osFoundry deliberately does not publish a fixed price table in the app, because the rate you charge is the rate the server quotes at the moment you pick — so check the live picker for the current per-hour cost in your region.
How does per-hour billing work, and are there contracts?
There are no contracts, reservations, or minimum terms. Billing starts when the instance is provisioned and stops when you release it, so a short experiment costs only a fraction of an hour and there is nothing to cancel. Note that deploying an endpoint provisions and bills a GPU as soon as it comes up, even before the model is fully ready, so release instances you are no longer using.
Do I have to use cloud GPU, or can I run models locally?
You can run models on your own device GPU for free — osFoundry is local-first, and local training and inference use no cloud credits. Cloud GPU is there for when a model needs more VRAM than your machine has, or when you want dedicated accelerator capacity for a fine-tuning job or a served endpoint. You choose per workload whether to stay local or rent.