Compute
Run Python, AI, and machine learning jobs on managed GPU compute.
Bring existing code and a model, or ask your coding agent to recommend a starting point. Orchard handles the estimate, approved run, logs, results, and optional model deployment.
From estimate to result
- Plan: use your code and settings, or compare public model candidates before choosing.
- Estimate: compare Cheapest by estimated total cost, Balanced as the default recommendation, and Fastest by estimated time.
- Approve: choose an option and set maximum cost and runtime before Orchard creates a machine.
- Run: monitor setup and training logs, then download outputs or review reported metrics.
Managed GPU environment
You do not need Python, CUDA, PyTorch, TensorFlow, or JAX on your computer. Orchard creates an isolated environment, installs the declared packages, and checks the assigned GPU before running your command.
Explicit package versions are not silently replaced. Dependency, CUDA, driver, or CPU-only framework problems stop before training and appear in the logs and environment report.
Encryption throughout 🔒
Compute encrypts job data throughout storage and transfer, from the private input copy to logs, results, and model files. The original Storage file stays unchanged, and Orchard opens the encrypted copy only inside the active job.
A validated model bundle stays encrypted until Orchard opens it inside the managed deployment service and transfers it directly. Encryption keys and raw storage URLs are never exposed. Inference inputs and results pass through without being saved to Orchard Storage.
Deploy without retraining or duplicate packaging charges
When a completed model loads and passes its sample inference locally, the agent uploads that ready bundle directly. Orchard does not create another paid Compute machine just to package it again. The managed service still builds the deployment and Orchard repeats the sample request remotely before reporting Ready. That managed build and first smoke test require no speed-mode choice and are not charged to Builder usage.
If the actual saved model cannot load on CPU, Orchard uses an available local GPU first. Separate paid managed GPU validation is only a fallback when local validation is unavailable. That fallback requires measured local RAM or GPU-memory evidence, omits duplicate choices and price premiums with no meaningful time advantage, then recommends the lowest-cost timely option. After Ready, metering starts with the first real API request and covers its cold start, active time, and scale-down under the Builder spending cap. Zero replicas cost nothing, and private access is the default.
Current boundaries
- Compute runs project archives and command-line workflows; it is not an interactive notebook or general-purpose server.
- Paid cloud jobs require a Builder plan, spending cap, a fresh estimate, and explicit maximum-cost approval.
- Managed model deployment requires a Builder plan and spending cap; local packaging, managed build, and the first smoke test do not add a Builder usage charge.
- Recommendations and reports do not guarantee model quality, safety, or license suitability.
- Do not include passwords, tokens, or other secrets in an archive or run command.
Ask your agent
- "Train a model for this dataset. Recommend the model and evaluation plan first."
- "Use the training code in this project and show me Cheapest, Balanced, and Fastest estimates."
- "Review this completed run, validate its sample locally, then deploy it as a private API without retraining."