Last updated: October 19, 2026
A subprocessor is a third party that Gradiated, a division of PostHog Inc., engages to process your request content: prompts, completions, captured traces and the training artifacts built from them. Our Data Processing Addendum authorizes the subprocessors below, and our Privacy Policy explains how we handle personal information, including the service providers we use for payments, email and analytics.
Current subprocessors
Amazon Web Services, Inc.
- Purpose: Cloud hosting for the Gradiated API, gateway, console, databases, storage, networking and secrets.
- Data processed: Inference requests and responses while they are processed, captured traces, training datasets, evaluation results and model adapters, account data (names, email addresses, IP addresses) and usage records.
- Location: United States (us-east-1)
- Security: aws.amazon.com/compliance
RunPod, Inc.
- Purpose: GPU servers that run hosted models and training jobs.
- Data processed: Inference requests and responses while they are processed, training datasets and evaluation examples during a training run, and the model adapters that the servers load.
- Location: RunPod data centers, which can be outside the United States, currently including the Netherlands
- Security: trust.runpod.io
OpenAI, L.L.C.
- Purpose: Evaluation of training runs. A comparison model answers held-out examples from your captured traces, and a judge model compares that answer with the answer of the model being trained. OpenAI models are the default for both.
- Data processed: Held-out evaluation examples from your captured traces (the request messages), the comparison model’s answers, and the answers of the model being trained.
- Location: United States
- Security: trust.openai.com
Anthropic, PBC
- Purpose: Evaluation of training runs, as for OpenAI, when a task uses an Anthropic model as its comparison or judge model.
- Data processed: Held-out evaluation examples from your captured traces (the request messages), the comparison model’s answers, and the answers of the model being trained.
- Location: United States
- Security: trust.anthropic.com
A training run sends data to OpenAI or Anthropic only when its comparison or judge model comes from that provider. If you choose models that Gradiated hosts for both, the run sends no data to either.
Services you choose
If you send a request to Gradiated with your own OpenAI or Anthropic API key, we forward it to that provider. The provider processes the request under your agreement with it, so it is not our subprocessor for that request.
If you sign in with Google or GitHub, that provider authenticates you under its own terms and shares your name and email address with us. Requests that reach us through OpenRouter are processed by OpenRouter under its own terms before they reach us.
Other vendors
We use other vendors that do not receive your request content, so they are not subprocessors. For example, Temporal Cloud runs our training workflows, but their payloads are encrypted with keys that Temporal does not hold. ClickHouse Cloud stores usage records keyed by internal identifiers: model, token counts, charges, timing and status. Resend delivers sign-in codes and invitations, and Stripe processes payments. Product analytics use PostHog, which is part of PostHog Inc., the same company as Gradiated.
How we store your data
Trace capture is on by default for each workspace so you can use post-training features. When it is on, we store the request and response content of requests that name a task in Amazon S3, encrypted with a dedicated AWS KMS key, for the retention period in effect when it was captured: 180 days by default, or 30 or 90 days. A change to the retention period applies only to content captured after the change. You can turn trace capture off at any time in workspace settings. When it is off, we do not persist prompts, messages, tool payloads, attachments, or generated outputs after request processing. Post-training features need captured traces, so no new traces are collected for them and automatic training stops. We do not write this content to the Gradiated product database or application logs.
Training artifacts are stored in Amazon S3 under the training run. Training datasets and intermediate artifacts, such as checkpoints, are deleted automatically 90 days after the run finishes. Evaluation results, including the compared answers and judgments, and published model adapters are kept until the workspace is deleted, or until you ask us to delete them by emailing team@gradiated.com. Deleted traces and artifacts can remain in versioned storage for up to 30 more days. A copy of an adapter that served requests can remain in a GPU server’s cache until the cache evicts it or the serving process is replaced.
We retain limited non-content data, including model identifiers, token counts, timing, request status, service identifiers, and billing data. We use this data for billing, security, reliability, support, and legal obligations.
We do not use your content to train models for anyone else. If trace capture is on (the default), your captured traces are used only to train and evaluate your own models; turn it off to stop capturing new traces.
How you learn of changes
We publish every new or replaced subprocessor on this page before it starts processing your request content, and record the change with its date in the change log below. The “Last updated” date at the top of the page changes at the same time. If you object to a change on reasonable data protection grounds, email team@gradiated.com within 30 days, as described in our DPA.
Change log
- October 19, 2026: Published this list as its own page, with the purpose, data processed, location and security link of each subprocessor. The list names the providers that process request content: Amazon Web Services, RunPod, OpenAI and Anthropic. ClickHouse Cloud and Stripe, which were on the earlier list, do not receive request content and are no longer listed.
Security measures
- Public traffic uses encrypted HTTPS connections.
- Inference services run in isolated infrastructure and are not exposed directly as public application endpoints.
- Access to secrets and operational systems is restricted through service credentials and access controls.
- We record non-content usage and operational events to support billing, reliability, and incident response.
- We review access and service behavior as part of normal operations.
Gradiated is an early-stage service. We do not currently claim SOC 2, ISO 27001, HIPAA, or similar certification. Contact team@gradiated.com for security questions or a customer security review.