What 'your own cloud tenant' actually means
Three options are usually on the table. A vendor API means calling the model maker directly, for example the OpenAI API or the Claude API from Anthropic, under the vendor's business terms. A model in your own cloud tenant means using the same or similar models through your existing cloud contract: Microsoft Foundry (which includes Azure OpenAI), Amazon Bedrock or Google Vertex AI. Open-weight models on your own infrastructure means downloading published model weights and running them on servers or GPUs you control.
The middle option is often misread. With Microsoft and Amazon, the model does not run on your own virtual machines. Microsoft says it hosts the models it sells in Microsoft's Azure environment and that they do not interact with services run by the model providers, such as the OpenAI API. Amazon says Bedrock runs each provider's model in model deployment accounts owned and operated by the Bedrock service team, and that model providers have no access to those accounts, to Bedrock logs, or to customer prompts and completions. What you gain is that the model sits inside your existing cloud agreement, region choices, identity and network controls, rather than a separate vendor relationship. These platforms also offer partner models, such as Claude, where the model maker's own retention terms can apply as well as the platform's; the Microsoft row below covers only models sold by Azure.
This page compares what each provider's own data-privacy documentation says. It is general information, not legal advice, and provider terms change often; every statement below was checked against the provider's page on 22 October 2026, and you should re-check the live page and your own contract before relying on it.
Provider by provider: what each says about your data
Each row summarises the provider's own documentation, named in the last column. 'Abuse monitoring' is the provider's process for detecting misuse of its service; it is the main reason prompts are retained even when they are not used for training. The 'Effort to run' column is Atlas Bio's assessment, not the provider's. Terms checked on 22 October 2026.
| Option | Training use | Retention | Abuse monitoring and who can see prompts | Data residency | Effort to run | Source (checked 22 Oct 2026) |
|---|---|---|---|---|---|---|
| Vendor API: OpenAI API | Data sent to the API is not used to train or improve OpenAI models unless you explicitly opt in (since 1 March 2023) | Responses API keeps application state for at least 30 days by default (or when store=true); files, vector stores and similar objects are kept until you delete them | By default, abuse monitoring logs are generated for API use and kept for up to 30 days; Modified Abuse Monitoring or Zero Data Retention exclude your content from those logs, subject to OpenAI's prior approval; OpenAI may keep logs longer where law requires or to protect against harm, keeps images flagged as potential child-abuse material for manual review even under these controls, and can, with advance written notice, make models ineligible for these controls and retain and human-review flagged content (Safety Retention) | Available to eligible customers via OpenAI sales; set per project; regional processing is listed for the US, Europe and the UAE, storage-only for several other regions; regions other than the US also require approval for abuse monitoring controls and a Modified Retention amendment | Lowest: an account, API keys and the business terms | OpenAI API docs: Data controls in the OpenAI platform |
| Vendor API: Claude API (Anthropic) | Inputs and outputs from commercial products are not used for training by default, unless you report feedback or opt in | API inputs and outputs are deleted from the backend within 30 days by default; a zero data retention agreement can change this for eligible APIs, subject to Anthropic's approval | If automated trust and safety systems flag content as violating the Usage Policy, inputs and outputs may be kept for up to 2 years and classification scores for up to 7 years; under zero data retention Anthropic still keeps safety classifier results; for designated 'covered models' (Mythos-class models and similar), prompts and outputs are kept for 30 days on every platform where they are offered, including Bedrock, Google Cloud and Microsoft Foundry, and organisations on zero data retention must turn retention on to use them unless Anthropic notifies them that they are eligible to use them with zero retention; human review of flagged content is limited to a small group of approved reviewers | Per-request inference geography: 'global' (default) or 'us', for Claude 4.6 and later models only; storage at rest (workspace geography) is currently US only | Lowest: an account, API keys and the commercial terms | Anthropic Privacy Center (training; retention; zero data retention; covered models) and Claude Platform docs: Data residency |
| Own cloud tenant: Microsoft Foundry (models sold by Azure, including Azure OpenAI) | Prompts, completions, embeddings and training data are not available to OpenAI or other model providers and are not used to train generative AI foundation models without your permission or instruction | Models are stateless; data for optional features (files, Responses API, Assistants threads, stored completions) is kept in your Foundry resource in your geography and can be deleted by you | Flagged prompts and completions may be stored in an abuse-monitoring store in your resource's geography and reviewed by authorised Microsoft employees (located in the EEA for EEA deployments); customers approved for modified abuse monitoring are excluded from that storage and human review, though automated review may still run | Standard deployments process within the geography you choose; 'DataZone' deployments may process anywhere in the data zone (for example the EU); 'Global' deployments may process in any geography where the model is deployed; data at rest stays in your geography | Medium: an Azure subscription, choosing a deployment type, identity and network setup, and a Limited Access application for modified abuse monitoring | Microsoft Learn: Data, privacy, and security for Foundry Models sold by Azure; Microsoft Learn: Foundry Models sold by Azure abuse monitoring |
| Own cloud tenant: Amazon Bedrock | AWS states it will not use inputs or outputs to train Amazon or third-party models, and that they are not shared with model providers | Set by a per-Region data retention mode (none, default or aws_review) and each model's own requirement; 'none' means no request or response data is written to durable storage by AWS | Model-specific: for OpenAI models other than the gpt-oss family, classifier-flagged traffic is kept up to 30 days for automated abuse detection; for Claude Fable 5 and 5.1, all traffic is kept up to 30 days and flagged traffic may get human review by AWS; retained content stays with AWS and is not shared with model providers | With cross-Region inference, a geographic profile (for example US or EU) keeps processing in a Region within that geography, and a global profile may use any commercial AWS Region; check which inference options each model offers; content is stored at rest in the Region you use, but retained inputs and outputs from cross-Region inference are stored in the Region that processed the request | Medium: an AWS account, Region and retention-mode settings (API only at launch), IAM policies, and checking each model's retention requirement | AWS docs: Amazon Bedrock data protection; data retention; abuse detection; cross-Region inference; Amazon Bedrock FAQs |
| Own cloud tenant: Google Vertex AI (Google's docs now call it Gemini Enterprise Agent Platform) | Google states it won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction | Caching is on by default; cached data is deleted within 24 hours and caching can be turned off per project; the Interactions API stores prompts and responses unless you set store=false; Grounding with Google Search keeps derived query logs for up to 3 days, Grounding with Google Maps keeps data for 30 days, and storage for both cannot be turned off; some agent features retain session data and cannot reach zero retention | For Google models, Google may log prompts to detect abuse under its Acceptable Use and Prohibited Use policies; the page states no retention period for these logs and does not say who can review them; certain Advanced AI models or features may also log prompts and responses, and Google says zero retention may not be possible for them; request-response logging is off unless you enable it | Regional endpoints keep ML processing within the country or multi-region jurisdiction; the 'global' endpoint may process anywhere and gives no residency guarantee; data at rest stays in the location you select | Medium: a Google Cloud project, endpoint choice, cache settings and organisation policies | Google Cloud docs: Gemini Enterprise Agent Platform zero data retention; data residency |
| Open-weight model on your own infrastructure (for example gpt-oss) | None by the publisher: the weights run on your hardware, so prompts reach the publisher only if you send them | Whatever you configure in your own servers and logs | Your own administrators and anyone with access to your infrastructure and logs; no provider abuse monitoring | Wherever you run it | Highest: GPUs (OpenAI's Hugging Face model page says gpt-oss-120b, with its MXFP4 quantization, fits a single 80 GB GPU), serving software, patching, access control, logging and your own evaluation of the model; use is governed by the model's licence (Apache 2.0 for gpt-oss) | Hugging Face: openai/gpt-oss-120b model card (published by OpenAI) |
Where the options really differ
Training use is no longer the main differentiator. On business and API terms, every provider in the table says it does not train on your prompts by default. The bigger risk on that point is staff using personal or consumer accounts, which are governed by different terms; Atlas Bio's answer on keeping proprietary research data out of external AI models covers that first control.
Retention for abuse monitoring is where the options separate. Vendor APIs keep prompts for a period by default (up to 30 days for both OpenAI and Anthropic) and keep flagged content longer in Anthropic's case. Microsoft Foundry reviews flagged content automatically by default and stores flagged content, or content that is part of a suspected abuse pattern, for possible human review unless you are approved for modified abuse monitoring. Amazon Bedrock makes retention a setting you control per Region, but some models (currently Claude Fable 5 and 5.1) will not run unless you accept retention and possible human review by AWS. Google says it may log prompts for abuse detection on its own models without stating a period on the page we checked, and that some advanced models also log responses.
Data residency depends on configuration more than on the provider. Every cloud platform offers a 'global' option that may process your prompts in any geography for availability, capacity or cost reasons, and each documents a narrower option that keeps processing in a chosen geography. The default is not always the narrow one: Anthropic's default inference geography is 'global', and Microsoft's Global and DataZone deployment types widen where processing can happen. If residency matters for a dataset (for example data subject to data-transfer law, or technical data that may be export-controlled), the deployment type or endpoint has to be chosen and recorded, not assumed, and whether the data is controlled at all is a question for your export-control or legal adviser.
Optional features change the picture. Stateful features such as stored conversation state, file stores, caching and web grounding keep data that a plain request would not. OpenAI's Responses API keeps application state for at least 30 days by default; Google's Grounding with Google Search keeps derived query logs for up to 3 days with no way to turn that off. A zero-retention arrangement usually means turning such features off as well as signing the agreement.
Who can see your prompts is the question most often left unanswered. The answer is narrow but rarely 'nobody', and it differs by provider. Microsoft says authorised Microsoft employees may review flagged content unless you are approved for modified abuse monitoring. AWS says it reviews content only for models whose provider requires it, plus possible child-abuse material. OpenAI keeps the right to retain and review some flagged content even under its zero-retention controls, and Anthropic still keeps safety classifier results under zero data retention. Google's page we checked describes prompt logging but does not say who can review it. With open-weight models the answer becomes your own administrators, which is only better if your own access control and logging are good.
When open-weight models on your own infrastructure make sense
Running an open-weight model yourself is the only option in the table where no model provider receives prompts; if your servers or GPUs are rented, the hosting provider's terms still apply. That makes it the natural home for material that should never leave your boundary, such as unpublished sequences, partner-confidential data or assay results that are central to a programme.
The cost is that every control the providers run for you becomes yours: hardware and capacity, patching the serving software, authenticating users and agents, keeping logs that show what was asked and answered, and evaluating whether the model is good enough for the task. Smaller open-weight models are usually less capable than the frontier models offered through vendor APIs and cloud platforms, so some tasks may need a split: a local model for the sensitive step and a hosted model for the rest. The split only protects the material if the hosted step receives none of it, including summaries, extracted values, annotations or embeddings derived from it; anything passed onward should be treated as the same data class as its source.
Open-weight models are also offered through the cloud platforms. Amazon's abuse-detection page, for example, excludes the gpt-oss models from the flagged-traffic retention it applies to other OpenAI models on Bedrock. That can be a middle path, but the cloud platform's terms then apply, not the self-hosted row of the table.
Choosing by data class
The practical answer for most biotechs is not one option but a rule per class of data, decided in advance and enforced at the point where a request leaves your systems. Atlas Bio's answer on how to secure AI agents in life-science R&D sets out the wider controls, and the Secure Agentic AI for Life Sciences page describes Atlas Bio's approach.
A common pattern, offered here as a starting point rather than a rule, is: public or low-sensitivity material can go to a vendor API on business terms; internal and proprietary material goes to a model in your own cloud tenant with the region pinned, the retention setting confirmed and stateful features off unless needed; and the most sensitive material is kept to a local open-weight model or held for a person's approval. Personal data and health information need their own legal basis and provider agreement under the law that applies to you, which is outside the scope of this page. Material received under a confidentiality, material-transfer or collaboration agreement is governed by that agreement first; some prohibit disclosure to any third-party processor, including a cloud AI service, whatever the provider's terms. These tiers address confidentiality only. If an AI output will form part of a GxP record or decision (for example in manufacturing, quality or a regulated study), that use also needs computerised-system validation, supplier assessment, audit-trail and record controls under the regulations that apply to you (such as 21 CFR Part 11 and EU GMP Annex 11, and any AI-specific GMP guidance in force), and a fixed, recorded model version; that is outside the scope of this page.
Whichever options you use, record for each AI service: the provider page and contract that govern it, the date you last checked them, the model name and version, the deployment type or endpoint and region, the retention or abuse-monitoring setting and whether an exemption was approved, and which optional features are switched on. Re-check whenever the provider changes its terms or you adopt a new model, since several of the settings above are model-specific.
Choosing between these options starts with knowing where research data goes today. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report). In development: a customer-controlled AI environment in your own Azure tenant, an AI gateway and agent permissions. Each engagement states which controls are built.
Do AI vendors train on data sent through their APIs?
Not by default, according to their own pages checked on 22 October 2026. OpenAI says data sent to its API is not used to train its models unless you opt in, and Anthropic says it does not use inputs or outputs from its commercial products for training by default unless you report feedback or opt in. Consumer accounts can be governed by different terms.
Is Azure OpenAI the same as calling OpenAI directly?
No. Microsoft says the models it sells, including Azure OpenAI models, are hosted in Microsoft's Azure environment, do not interact with services operated by OpenAI, and that your prompts and completions are not available to OpenAI. Microsoft's own terms, abuse monitoring and residency options apply instead of OpenAI's.
Can cloud provider staff see our prompts?
In limited cases. Microsoft says authorised employees may review flagged content unless you are approved for modified abuse monitoring. AWS says it reviews content for models whose provider requires human review, which those models only run with if your retention mode allows it, and it may also store and review suspected child-abuse material. Google says it may log prompts to detect abuse on its own models; the page we checked does not say who can review those logs.
Are open-weight models automatically safer?
They remove the model provider, not the risk. No model provider receives your prompts, but your own team must secure the servers, control access, keep logs and evaluate the model. A poorly run local model can expose more than a well-configured cloud service.
Does choosing an EU or US region keep all processing there?
Not necessarily. A regional option usually keeps processing within a geography or jurisdiction, not one data centre region, and some regions offer storage only: OpenAI lists regional processing for the US, Europe and the UAE but storage-only for several others. Global deployment types and endpoints on Microsoft Foundry and Google Vertex AI may process prompts in other geographies, Anthropic's default inference geography is global, and Bedrock's cross-Region inference may process requests in another Region within a geography, or any commercial Region with a global profile.
How often should we re-check provider terms?
At least whenever you adopt a new model or feature and whenever a provider updates its data page. Several pages in this comparison changed during 2025 and 2026, and some retention rules now apply to specific models only.
- OpenAI API docs: Data controls in the OpenAI platform
- Anthropic Privacy Center: Is my data used for model training?
- Anthropic Privacy Center: How long do you store my organization's data?
- Anthropic Privacy Center: Zero data retention - which products it applies to
- Anthropic Privacy Center: Data retention practices for covered models
- Claude Platform docs: Data residency
- Microsoft Learn: Data, privacy, and security for Foundry Models sold by Azure
- Microsoft Learn: Foundry Models sold by Azure abuse monitoring
- AWS docs: Data protection in Amazon Bedrock
- AWS docs: Amazon Bedrock data retention
- AWS docs: Amazon Bedrock abuse detection
- AWS docs: Cross-Region inference in Amazon Bedrock
- AWS: Amazon Bedrock FAQs
- Google Cloud docs: Gemini Enterprise Agent Platform (formerly Vertex AI) and zero data retention
- Google Cloud docs: Gemini Enterprise Agent Platform (formerly Vertex AI) data residency
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (revision 1, 2011)
- Hugging Face: openai/gpt-oss-120b model card
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →