Short answer: yes, you can send personal data to OpenAI, Anthropic and other AI providers outside India under the DPDP Act, as the law stands on 20 November 2026. Section 16(1) is a negative list. Transfers are allowed unless the Central Government notifies a country or territory as restricted, and none has been notified. Section 16 starts on 13 May 2027. Three things can still stop you: a sector rule such as the RBI's payment data storage direction (Section 16(2) keeps it alive), a customer contract, and a future notification. The practical question is therefore not "is it legal?" but "where does this provider store and process my data, and can I prove it?" The table below answers that for the main providers, dated 20 November 2026.
This is an engineering guide, not legal advice. Whether a transfer is lawful for your business, your sector and your customers is a question for counsel.
Does DPDP allow personal data to go to AI APIs outside India?
Section 16(1) lets the Central Government, by notification, restrict a Data Fiduciary from transferring personal data to a country or territory outside India that it names. That is the whole of the transfer rule in the Act. There is no adequacy test, no list of approved countries and no standard contract clause scheme. If a country is not on the notified list, you may transfer to it.
Four related points complete the picture:
- Rule 15 says a transfer is allowed subject to any requirements the Government sets, by general or special order, for making personal data available to a foreign State or to a person or entity under its control. It is not the negative list itself. The negative list is Section 16(1). We found no such order as of 20 November 2026.
- Section 16(2) says the section does not restrict any Indian law that gives a higher degree of protection for personal data, or restricts its transfer. Sector rules therefore still apply.
- Rule 13(4) lets the Government name personal data that a Significant Data Fiduciary must not transfer outside India, along with the traffic data on its flow. No Significant Data Fiduciary has been notified, and we found no such list of data.
- Section 8(1) and 8(2) keep you responsible for what a Data Processor does for you, and require a valid contract with it. An AI provider that handles personal data on your behalf is a processor. Moving the data abroad does not move the responsibility.
DPDP does not require data to stay in India, and it does not require you to self-host a model. Hosting in India avoids a future Section 16 restriction and most sector localisation rules, which is a reason to prefer it where it costs little. It is not a legal default. For the wider AI architecture, including embeddings, logs and deletion, read our post on DPDP compliant AI architecture.
When does cross-border transfer still go wrong?

| Constraint | Where it comes from | What to do |
|---|---|---|
| Payment data | Where it comes from: RBI direction RBI/2017-18/153 of 6 April 2018: the "entire data relating to payment systems" must be stored "only in India". Kept alive by Section 16(2) | What to do: Keep payment system data in India. The RBI's FAQ allows processing abroad only if the data is deleted there and brought back to India within one business day or 24 hours, whichever is earlier. Do not send payment data to an AI API without checking this with counsel |
| Customer contracts | Where it comes from: Banks and large enterprises often fix the allowed regions in the master agreement or security schedule | What to do: Read the data location clause before you add an AI provider. A transfer can be lawful under DPDP and still breach the contract |
| Future notification | Where it comes from: Section 16(1) lets the Government restrict a destination at any time | What to do: Put the provider behind your own gateway so you can change region or vendor without rewriting the product |
| Significant Data Fiduciary duty | Where it comes from: Rule 13(4) localisation of data the Government specifies | What to do: Check whether you are likely to be notified. If so, plan for India-only handling of the named data |
| EU users | Where it comes from: GDPR Chapter V restricts transfers out of the EU by default. India has no EU adequacy decision (European Commission list, November 2026) | What to do: If you serve EU people, handle that under GDPR transfer tools with your counsel. DPDP compliance does not settle it |
Section 16 itself starts on 13 May 2027. Until then the transfer rule is not in force, but sector rules and contracts are. Build to the 2027 position now, so you do not redo it.
Where does each AI provider store and process your data?
This table is dated 20 November 2026 and is built from each provider's public documentation. These pages change often, and the answer depends on your plan, model, features and contract. Treat it as a map of what to ask for, and re-check each row against the live page and your signed terms before you rely on it.
| Provider and product | India storage at rest | India processing (inference) | Training and retention, per the vendor |
|---|---|---|---|
| OpenAI API | India storage at rest: Yes, through in.api.openai.com | India processing (inference): Yes, but only for gpt-6-luna, on Standard pay-as-you-go, and without hosted tools. Data residency is for eligible organisations, arranged through OpenAI sales. OpenAI charges a 10% uplift on data residency endpoints for eligible models released on or after 5 March 2026 (confirm whether it applies to your model). System data, such as account and usage data, may be stored and processed outside the region | Training and retention, per the vendor: Not trained on your data by default (OpenAI enterprise privacy page, updated 8 January 2026). Abuse logs are kept up to 30 days. Zero data retention needs approval |
| ChatGPT Enterprise and Edu | India storage at rest: Yes, India is available for storage at rest | India processing (inference): No. Inference residency covers only the EEA and Switzerland, the US and the UAE | Training and retention, per the vendor: Not trained on your data by default. Retention is set in workspace controls and your contract |
| Claude API (Anthropic, first-party) | India storage at rest: No India option. Workspace geography is US only | India processing (inference): No India option. The inference_geo setting is "global" or "us". US-only inference costs 1.1x | Training and retention, per the vendor: Not used for training without express permission. Inputs and outputs are deleted within 30 days by default (Anthropic page updated 1 July 2026) |
| Claude on Amazon Bedrock | India storage at rest: Depends on the features you use. Check where any stored data sits for your account and Region | India processing (inference): Since 29 September 2026, Claude Opus 5, Sonnet 5 and Haiku 4.5 are available through an India inference profile that routes only between ap-south-1 (Mumbai) and ap-south-2 (Hyderabad) | Training and retention, per the vendor: Governed by AWS terms. Fable 5.x models need an aws_review setting that allows up to 30 days' retention, so check before you use them for personal data |
| Azure OpenAI and Foundry | India storage at rest: Stored data stays in the resource's geography | India processing (inference): Standard deployments process in the chosen geography. Global deployments may process anywhere. DataZone deployments process within the US or EU zone. Which models are offered in India regions varies, so check the Azure portal | Training and retention, per the vendor: Data is not available to OpenAI and is not used for training (Microsoft documentation). Check the abuse monitoring terms for your deployment |
| Microsoft 365 Copilot | India storage at rest: Stored with your other Microsoft 365 content, under the data residency commitments in Microsoft's Product Terms | India processing (inference): Outside the EU, queries "may be processed in the US, EU, or other regions" (Microsoft page, updated 4 November 2026). In-country processing for India has been announced. Confirm it for your tenant | Training and retention, per the vendor: Prompts, responses and Graph data "aren't used to train foundation LLMs" |
Google Gemini and other providers are not in this table because we did not verify their pages for this post. Use the same questions for them. For which of these products suits staff use, product features or self-hosting, see our comparison of ChatGPT Enterprise, Azure OpenAI, Copilot and self-hosted models.
What is the difference between storage, processing and logs?
The word "residency" hides three separate places your data can sit. Each one needs its own answer.
- Storage at rest. Files, conversation history, batch inputs and fine-tuning data held by the provider. India storage means these sit in an Indian region.
- Processing (inference). Where the model actually reads your prompt and writes the answer. A provider can store in India and still run the model in the US or Europe. ChatGPT Enterprise is an example: India storage is available, but inference residency is not.
- Logs, caches and stored features. Abuse monitoring logs, prompt caches, conversation state and tool outputs. These can keep data longer, and sometimes elsewhere, than a plain request. OpenAI, for example, says data residency does not apply to system data such as account, metadata and usage data. Ask where each of these lives and for how long.
A provider that offers India storage but not India processing still sends your prompts out of India. That is allowed under Section 16 as of 20 November 2026. It is not fine if your customer contract or sector rule says "processed in India". Write down which of the three your promise covers.
Does choosing an India cloud region solve it?
Only if the model you use is deployed there, and the way you call it keeps traffic there. Three traps come up often:
- Global routing. Some deployment types and inference profiles send requests to whichever region has capacity. On Azure, a Global deployment may process anywhere. On Bedrock, a cross-region profile can route outside your region. Use the India profile or a Standard deployment pinned to your geography, and record which one you picked.
- Model availability. The newest model often arrives in the US or EU first. If the model you want is not in an India region, the choice is a slower model in India or a faster one outside. Decide this per use case, based on the data in the prompt.
- Features outside the region. On the OpenAI API, India processing excludes hosted tools. Tools, file search and web search can send data to separate services. Read the feature list for your plan before you turn them on.
Regional processing can also cost more. OpenAI charges a 10% uplift on data residency endpoints for eligible models released on or after 5 March 2026, and Anthropic charges 1.1x for US-only inference on Claude 4.6 and later models. Include that in your model-choice decision, not as a surprise on the invoice.
How do you design AI features for transfer rules you cannot yet see?

The rules are open on three sides: the Government may notify a country, a regulator may add a localisation rule, and a customer may ask for India-only handling. The way to cope is to make the region a setting, not a hard-coded fact.
- Put a gateway in front of every provider. One internal service calls the model. Product code never holds a provider URL or key. Changing region or vendor then becomes a configuration change.
- Route by data class. Decide, per use case, whether the prompt can contain personal data. Send prompts with personal data to an India-resident route. Send low-risk prompts, such as public text or synthetic test data, to the cheaper or faster global route. Enforce this in the gateway, not in a policy document.
- Minimise before sending. Strip or replace names, phone numbers, account numbers and IDs before the prompt leaves your system, and map them back afterwards. Our guide to PII redaction for LLMs covers how.
- Pin the region in code and log it. Set the regional endpoint or profile explicitly. Log the provider, model, endpoint and region for each request class, without storing the prompt, so you can show where data went.
- Keep a vendor register. For each provider, record the contract, the data location terms, training and retention settings, sub-processors and the date you checked. Re-check on a schedule, because these terms change. Our AI tool settings guide lists what to record.
- Test the failure mode. If the India route is down, should the gateway fail closed or fall back to a global route? For personal data, fail closed unless your contract and counsel say otherwise.
When should you run the model in India yourself?
If the provider you need has no India processing, or your customer or regulator needs proof that data never left India, the alternative is to run an open-weight model on infrastructure you control, in India. That brings its own work: GPUs, serving, monitoring, updates and security. It makes sense when the data is sensitive, the volume is steady, or a contract demands it. It is a poor fit for light, occasional use, where a regional API is simpler.
Our post on VPC, on-premises and GPU cloud compares the hosting options, and our Own Your AI in India page covers how we set up private AI on a client's own servers or cloud. A hybrid is common: a regional API for general work and a private model for the prompts that carry the most sensitive data.
What should you check before you send personal data to an AI provider?
- The data in the prompt: does it include personal data, and which kinds?
- The provider's role: is it your Data Processor, with a signed contract that meets Section 8(2) and Rule 6 safeguards?
- The destination: storage region, processing region, and where logs and caches sit, for your plan and model.
- Sector and customer rules: payment data, other regulator rules and contract location clauses.
- Training and retention: training off by default or by written opt-out, and retention that fits your deletion promises.
- Sub-processors: who else the provider uses, and where they are.
- Exit and change: a gateway, a tested fallback, and a re-check date.
- Evidence: a dated vendor register entry and request logs that show the route taken.
Sources: Digital Personal Data Protection Act, 2023, Sections 8 and 16 (meity.gov.in). Digital Personal Data Protection Rules, 2025, Rules 6, 13 and 15 (dpbi.gov.in). RBI notification RBI/2017-18/153 of 6 April 2018 and RBI FAQ on storage of payment system data (rbi.org.in). OpenAI: data residency guide (developers.openai.com), data residency for ChatGPT (help.openai.com), enterprise privacy page and API data controls. Anthropic: data residency page (platform.claude.com) and privacy centre retention article. AWS: Amazon Bedrock Claude India cross-region inference announcement and Bedrock data retention documentation. Microsoft Learn: Azure OpenAI data privacy and Microsoft 365 Copilot privacy. European Commission: adequacy decisions. Last reviewed 20 November 2026.
FAQ
Can I use OpenAI or Claude under the DPDP Act?
Yes, as the law stands on 20 November 2026. Section 16(1) of the DPDP Act only lets the Government restrict transfers to countries it notifies, and none is notified. You still need a processor contract, safeguards, and a check of sector rules and customer contracts. The Act does not require you to keep data in India.
Does DPDP require data to be stored in India?
No. The Act has no general data localisation rule. Section 16(1) allows the Government to restrict transfers to notified countries, and Section 16(2) keeps stricter sector laws, such as the RBI's rule that payment system data be stored only in India. Rule 13(4) can also require named data of a Significant Data Fiduciary to stay in India.
Does OpenAI process data in India?
For the API, yes, in a limited way. As of 20 November 2026, India supports storage and processing on in.api.openai.com, but only for the gpt-6-luna model on Standard pay-as-you-go, without hosted tools. For ChatGPT Enterprise and Edu, India is available for storage at rest, but inference residency covers only the EEA and Switzerland, the US and the UAE.
Does the Claude API have an India region?
Not on Anthropic's own API as of 20 November 2026. The inference_geo setting offers "global" or "us" only, and workspace geography is US only. Claude Opus 5, Sonnet 5 and Haiku 4.5 are available through an India inference profile on Amazon Bedrock, which routes only between ap-south-1 and ap-south-2.
What is the difference between data residency and inference residency?
Data residency is where the provider stores your data at rest, such as files and conversation history. Inference residency is where the model actually processes your prompt. A provider can offer one without the other. Ask about both, and about where logs and caches are kept.
How do I stay ready if India restricts a country later?
Call every model through your own gateway, set the provider region in configuration, keep a dated vendor register, and route prompts with personal data to a region you can defend. That lets you change region or vendor without rewriting the product. Section 16 starts on 13 May 2027, so build it before then.
