vGPU runs work locally or in the cloud, groups GPUs for larger workloads and divides enterprise GPUs so smaller jobs use only the capacity they need. One platform for GPU jobs, model services and private, encrypted inference.
vGPU is in early access. Tell us about your workload to join.
Run models and workflows across multiple GPUs when they exceed one card's memory. vGPU uses software that splits the work across suitably connected GPUs; separate cards do not automatically become one larger memory pool.
Partition enterprise GPUs that support it, including NVIDIA B200, into smaller GPU instances. Smaller workloads use an appropriately sized partition instead of reserving an entire card.
Encrypt requests on your device and receive encrypted results. Models run inside a verified, protected CPU and GPU environment, designed to keep prompts and outputs private from infrastructure operators.
Your client verifies the environment and its encryption key before sending data. Plaintext exists only inside the protected execution boundary; ordinary routing services cannot decrypt it, and confidential requests never fall back to an ordinary endpoint.
Discover your GPUs and installed engines, then run workflows from a local page. No cloud account or Kubernetes required. Jobs keep running after the browser closes. Connect a vGPU.com account when you want to manage your machines remotely.
Submit a recipe with your inputs and budget, follow its progress and cancel it at any time. Finished jobs release their capacity, and interrupted jobs resume from committed checkpoints where the workload supports them.
Choose dedicated capacity across our own fleet and partner clouds. Each offer states memory per GPU, runtime isolation, storage and interruption terms.
Use shared model APIs or a dedicated model endpoint, with compatibility for Chat Completions, Responses and Messages clients.
Bring your own weights and training data. Run training and fine-tuning recipes, keep versioned outputs and deploy the resulting models with project-scoped access.
Run agents in isolated workspaces with tools, files, task history and budgets. Continue or cancel tasks, review their changes and use your existing coding clients.
Receive a notice, acknowledge it and report when your workload is ready to release capacity. Notice time depends on the offer, and some interruptions arrive without warning.
Keep large datasets, model weights, checkpoints and results separate from compute. Recovery uses committed checkpoints and approved destinations; moving work between clouds uses portable data and a compatible recipe.
Use your local GPUs first, with optional overflow to cloud capacity when they are full or too small. You set the spending and data-transfer limits; connecting an account never turns on paid rentals or exports your data by itself.
Release finished allocations and fit work using measured memory needs. Model caches, reuse, batching and bounded warm pools balance startup latency against the cost of idle GPUs.
Project permissions, quotes, spending caps, usage views and invoices keep costs predictable. More expensive recovery destinations require an approved policy.
Use consistent job and resource contracts across providers, with SDKs, operation status, cancellation and signed webhooks. Placement respects each workload's hardware and data requirements.
Bring home, office and private-cloud GPUs into one fleet for separate jobs or replicas. Pooling a fleet is different from splitting one model across its machines.
Reuse established engines for execution, queueing, workflows and model serving. Keep the local starter small: engines, models and optional cluster tools are installed only when you choose them.
A dedicated GPU can remain billable while idle because it is still assigned to you. Sleeping a model or setting replicas to zero does not necessarily release the rented machine.
Elastic deployments return unused compute, with a cold start when demand returns. Compute charges end at the defined release event; storage and other agreed charges may continue.
Prices and release terms are shown before purchase.
Share your workload, the GPUs you use today and what needs to improve.
Email [email protected]We work with data centres and GPU owners on capacity partnerships. Contributing capacity is always separate from connecting your own local runtime.
Email [email protected]