Metered API
The metered API will charge per token for open-weight models shared by all of its customers. Prompts and outputs will never be logged or stored, the same as on a private endpoint.
Private endpoints
Private endpoints will serve the models you choose from dedicated GPUs. No other customer’s workload will share those GPUs, and their capacity will never go to anyone else.
Wholesale capacity
Wholesale customers will rent our servers or individual GPUs in US datacenters by the hour, or reserve them for a term.
Agent hosting
Agent hosting will follow the launch.
Data handling
The metered API and private endpoints will be return-to-source: each response goes back to the system that sent the request, and no prompt or output is stored on our side. Prompts will not be logged, and customer data will never be used for training.
Every contract will include a signed data processing agreement. Private endpoint contracts will also include a business associate agreement where HIPAA applies. Zero retention will be written into the contract, and we use no foreign subprocessors.
The Compliance page lists our privacy commitments and our SOC 2 and FedRAMP status.
Models and versions
You will choose the open-weight models you run. On a private endpoint, each model version will stay pinned until you decide to change it, so when an examiner asks which model produced a decision last March, you can name the version.
Support
Metered API and wholesale customers will get email support, with replies the same business day. Private endpoints will add a named engineer in the United States who tunes the deployment to your workload, maintains it and answers the phone.
Pricing
The metered API will be priced per token, at rates published on this site. Each private endpoint will have a fixed price for its term, so its cost does not move with usage. Wholesale capacity will be priced by the hour or by the term.
Who it is for
ArcVault is for health systems, banks, financial services firms, law firms, federal agencies and federal contractors, whose data cannot go to public inference.