Skip to main content
Applies to: every Valar customer using Valar’s inference services
Effective date: 2026-05-01
Contact: [email protected]
This document describes how Valar handles customer data, including personal data, while delivering its AI inference services.

Definitions

  • Customer Content - the prompts, inputs, files, and related data submitted to run an inference job, together with the outputs the service returns.
  • Job metadata - the operational metadata required to run and track jobs, such as job identifiers, timestamps, status, and routing or state information.
  • Compute subprocessor - a third-party compute provider Valar may draw on to run inference.

1. Controller and processor roles

The Customer is the controller of the data it sends to Valar and decides which inference jobs run. Valar is the processor: it handles that data solely to carry out the requested inference job and return the results to the Customer.

2. Scope of data processed

Valar may handle two categories of data, defined above: Customer Content and Job metadata.

3. Permitted and prohibited uses

Valar uses Customer data for two purposes only:
  1. running the inference job the Customer has requested, and
  2. operating the service capabilities that job depends on — routing, scheduling, and job-state tracking.
Valar will not:
  • use Customer Content to train, fine-tune, or otherwise improve any machine learning model, whether Valar’s own or a third party’s;
  • use Customer Content for marketing or advertising; or
  • collect telemetry from GPU machines that would capture Customer Content.

4. Storage and retention

AspectHow Valar handles it
Persistent storageCustomer Content is stored persistently only in Amazon S3 buckets, unless the Customer opts for the customer-owned bucket arrangement below.
Transient processingEverything else happens transiently and in-memory only, lasting no longer than the job itself.
Automatic deletionValar’s production S3 buckets carry an automated deletion rule (an S3 bucket lifecycle policy) that removes Customer Content soon after processing. Timing can shift due to job retries, failures, or other operational factors. In no event will Valar hold Customer Content for more than 48 hours.
Customer-owned buckets (optional)Customers wanting tighter control over storage configuration and retention may elect to use their own Customer-owned S3 buckets.

5. Security and access controls

Valar protects Customer data with controls that include the following.
  • Authenticated GPU machines. Inference GPU machines connect to Valar infrastructure over TLS (certificate transparency required) and authenticate with a high-entropy bearer token. They are never publicly reachable.
  • Restricted connectivity. Inference machines reach only the services they need:
    • container registry
    • routing service
    • object storage (e.g. Amazon S3)
  • Data center assurance. Valar relies on cloud and data center providers that maintain SOC 2 and ISO 27001 assurance programs.
  • Compute provider restrictions. Valar may draw on several compute subprocessors, subject to two conditions:
    • they are barred from accessing Customer Content, and
    • they must never store Customer Content unencrypted.

6. Subprocessors

Valar engages third-party providers for infrastructure such as compute, container distribution, routing, and job-state tracking. With the sole exception of Amazon S3 as described in Section 4, subprocessors never store Customer data. Valar may update its subprocessors from time to time.

7. Security incident notification

If Valar becomes aware of a security incident involving unauthorized access to Customer Personal Data, it will notify the Customer within 72 hours of becoming aware.

8. Questions

Direct any questions about this document or about how Valar handles Customer data to [email protected].