Deploy your Docker image on a dedicated GPU. Your application gets its own HTTPS address, your data stays on your disk, and you only pay for the hours you use.
Pick an image, launch, then stop the instance: the demo below follows the portal flow.
ghcr.io/open-webui/open-webui:ollamaLive GPU
Application output
Loading WebUI… ollama: GPU detected, CUDA available ollama: model loaded in 6.2 s Uvicorn running on http://0.0.0.0:8080 GET /api/chat 200
A powerful GPU is a heavy investment, and it often sits idle between projects.
Networking, images, permissions, options billed separately: it takes time before an application answers.
A training run, a calculation campaign, an LLM to try: a few hours or days, not a machine for the year.
Enter your Docker image, your application port, the vCPUs, memory and disk.
A GPU is reserved for you, the machine starts and pulls your image.
Your application answers on its HTTPS address. Follow the GPU and logs from the portal.
Each instance runs in its own virtual machine, with the GPU passed directly to the VM. No sharing, no noisy neighbour.
An HTTPS address with a valid certificate is created for every instance.
Stop the instance at night, restart it in the morning: images, models and files are still there. Only deletion erases them.
Hourly rate on quote by GPU type, counted per hour slot, never counted twice, with a CSV export.
Cheaper, with a grace period to shut down cleanly if the GPU is needed.
GPU temperature, power, utilisation and memory, plus your application output, without logging into the machine.
Run an open-source model (Qwen, Gemma, Llama…) behind a chat interface, without sending your data to a third party.
Host a vision or language model behind an HTTPS API.
Fine-tune a model on your data, then stop the machine to stop paying.
Parallel computing, simulation, image or point-cloud processing.
Tested on our bench: Open WebUI and Ollama deployed from their Docker image, reachable on their own HTTPS address, on a 24 GB RTX 3090.
| On demand | Spot | |
|---|---|---|
| The GPU | Stays yours until you stop it | Can be reclaimed if an on-demand customer needs it |
| Interruption | None | After a grace period to shut down cleanly |
| Price | On quote, hourly, by GPU | On quote, cheaper than on demand |
| For | Live applications, LLMs, APIs | Training and jobs that can resume |
The path of a request, from your browser to the GPU.
GPU servers are not exposed directly to the Internet. Every action is logged: creation, GPU reservation, start, stop.
Each customer in their own virtual machine, with a dedicated GPU.
Sign-in with your TriDyme account (Keycloak), with organisations, members, roles, quotas and API keys.
Only the images allowed for your organisation can be launched.
It is in early access, on request. Tell us about your workload: we open an access for you and help you deploy your first instance.
Prices are set on quote, depending on the GPU type and the offer. Usage is counted per GPU hour: a stopped instance releases the GPU and is no longer billed.
Yes, on the instance disk, as long as the instance is not deleted.
Yes, once it is added to the images allowed for your organisation. Today: public images and the registry.tridyme.ai registry.
If an on-demand customer needs the GPU, your instance gets a grace period to shut down cleanly, then the GPU is released.
Tell us about your workload: we open an access for you and help you deploy your first instance.