Subprocessors
The third parties that process personal data on our behalf.
This page lists the third parties that process personal data on our behalf. It is referenced by Section 8 of our Data Processing Addendum and Section 8.2 of our Terms of Service.
We give at least thirty days' notice before adding or replacing a subprocessor that processes customer personal data. To be notified, email legal@round3.ai and we will add you to the list.
Customer Traffic is the column that matters. It answers the question enterprise procurement actually asks: does my production data leave your building, and who gets it.
Model providers, on your account
When you run an evaluation or experiment against a model, three.dev calls that provider as you, using the API key you supply, encrypted per organization with AES-256-GCM. Your existing agreement with that provider governs their use and retention of the traffic, including any no-training terms you negotiated. We do not substitute our own account for yours, and these providers are not our subprocessors on this path.
| Provider | Purpose | Customer Traffic | Whose account |
|---|---|---|---|
| OpenAI | Model inference | Yes | Yours |
| Anthropic | Model inference | Yes | Yours |
| Google (Gemini) | Model inference | Yes | Yours |
| Microsoft Azure OpenAI | Model inference | Yes | Yours |
| Amazon Bedrock | Model inference | Yes | Yours |
Model providers, on our account
Parts of the Service run on our own provider account rather than yours. These providers are our subprocessors and we are responsible for their terms.
| Provider | Purpose | Customer Traffic | Whose account |
|---|---|---|---|
| Anthropic | AI judges we operate, and generating failure cluster titles, descriptions, and taxonomy labels | Yes, content derived from your traffic | Ours |
| OpenAI | AI judges we operate, when a judge is configured to run on this provider | Yes, content derived from your traffic | Ours |
| Google (Gemini) | AI judges we operate, when a judge is configured to run on this provider | Yes, content derived from your traffic | Ours |
Content sent on this path is still your Customer Traffic. It is never used to improve any model, feature, or dataset made available to another customer or to the public.
Specialized models that three.dev builds and serves for you run on our own infrastructure, described below. They are not a third-party provider.
Infrastructure
| Provider | Purpose | Customer Traffic | Location |
|---|---|---|---|
| Amazon Web Services | Application compute, primary database, container registry, load balancing, key management, secrets, logs | Yes | us-east-1, United States |
| Google Cloud Platform | Storage for the specialized model weights we build for customers, and infrastructure state | No | United States |
| ClickHouse Cloud | Analytics store for evaluation results, verdicts, and telemetry | Yes | United States |
| RunPod | GPU hosting for the specialized models we build and serve for customers | Yes | United States |
| Cloudflare | DNS, edge security, and hosting for three.dev, including the web app, the management API, and the endpoints that serve the specialized models we build for you. | Yes, in transit | Global edge |
| OpenRouter | Model routing service, on your account. Forwards your request to an upstream model provider it selects | Yes, in transit | Varies by upstream |
OpenRouter is a routing service rather than a model provider, so the provider that ultimately receives your traffic depends on the model you select and on OpenRouter's own routing. We cannot enumerate those upstream providers, and their terms are governed by your agreement with OpenRouter. If you need a closed subprocessor list, ask us to disable OpenRouter for your organization.
Cloudflare sits in front of the web app, the management API, and the endpoints that serve the specialized models we build for you, so that traffic reaches its edge before it reaches us and is decrypted there, then re-encrypted on the way to our infrastructure. Cloudflare is a transit point rather than a destination; we do not store Customer Traffic there.
Product and business operations
| Provider | Purpose | Customer Traffic | Location |
|---|---|---|---|
| Clerk | Authentication and user management for the three.dev app | No | United States |
| Datadog | Application monitoring, logs, traces, and on-call routing | Yes, in limited cases. See below | United States |
| Dash0 | Application monitoring, logs, traces, and metrics, in parallel with Datadog | Yes, in limited cases. See below | United States |
| Stripe | Payment processing and invoicing for customer accounts. Listed ahead of first use so the notice period in Section 8.3 of the DPA runs now | No | United States |
| PostHog | Product and website analytics, including session replay of the app | No. Replay and autocapture mask all text and element attributes by default | United States |
How Customer Traffic can reach Datadog and Dash0. Our monitoring pipeline sends application traces, metrics, and logs to both. Ordinary telemetry carries identifiers, timing, and token counts, not content. One path can carry content: error responses returned by model providers, which are logged verbatim for debugging and can echo part of the request that caused them. Request and response bodies are never written to logs. This path is subject to each vendor's retention on our account and to the same obligations as every other subprocessor in this list.
Changes to this list
We publish changes here before they take effect. See Section 8.3 of our Data Processing Addendum for the notice period and your right to object.