Skip to content

Private AI inference for regulated work.

ArcVault will run open-weight models on hardware we own in the United States. Your prompts will return to you and nothing will be kept. You will choose the models and the budget.

Service summary

Platform details
Deployment
Private endpoint or metered API
Hardware
Owned by ArcVault, in US datacenters
Data
Returned to you, not stored
Your prompts
Never logged or used for training
Models
Your choice of open-weight models
Model versions
Pinned on private endpoints
Capacity
Reserved for private endpoints
Support
A named US engineer for private endpoints

For organizations that answer to a regulator

Health systems, banks, financial services firms, law firms, federal agencies and federal contractors hold data they cannot send to public inference.

Security
Public inference providers set their own terms for retention, logging and subprocessors, and they can change those terms.
Privacy
PII, PHI and privileged material carry legal duties. Every vendor that receives them needs terms your counsel will sign.
Provenance
Providers update and retire hosted models on their own schedule. An examiner who asks which model produced a decision last March expects a version number.
Cost
Spend follows usage, which is hard to forecast, and prices change at the provider’s discretion. Budget owners need a figure that holds for the year.

How it works

What you will get

Owned and run in the US
ArcVault is a US-owned company with US-based staff. Every request will run on servers we own in US datacenters, and private endpoints will get dedicated GPUs that no other customer shares.
A named US engineer
Each private endpoint will have a named engineer in the United States who tunes it to your workload and answers the phone.
Evidence for your reviewers
Every contract will include a signed data processing agreement, and private endpoints will add a business associate agreement where HIPAA applies. Before you sign, your reviewers will get the agreement templates, our security policies, an architecture diagram and a completed security questionnaire.
Model choice and pinning
You will choose the open-weight models you run. On a private endpoint, each version will stay pinned until you change it, so you can say which one produced any past output.

What we will offer

Metered API rates will be published per token, and each private endpoint will have a fixed price for its term.

Metered API
Pay per token for open-weight models shared across customers. Prompts and outputs will never be logged or stored.
Private endpoints
Run the models you choose on dedicated GPUs, with pinned versions and a named US engineer.
Wholesale capacity
Rent our servers or individual GPUs in US datacenters by the hour, or reserve them for a term.
Agent hosting
We will add it after launch.

Compliance

Inference is a commodity. Privacy is not.

Many providers run the same open-weight models, and the price keeps falling. The difference is who keeps your prompts and outputs. ArcVault will keep none of them, and our contracts will say so.

Preparing

SOC 2 Type II

An independent CPA firm tests how a company’s security controls operate over time, then issues a report. We are preparing for our SOC 2 Type II audit.

Pursuing

FedRAMP 20x Class C

FedRAMP is the federal government’s program for certifying cloud services for agency use. We are pursuing FedRAMP 20x Class C Certification.

  • Contractual zero retention
  • Signed DPA, BAA on private endpoints
  • US-only staff and hardware
  • No foreign subprocessors
arcvault Systems

Tell us what you are trying to run.