Research · Cloud provider documentation
Which LLMs can run in Canada or Japan? Data residency tracker
As of October 14, 2026, none of the major clouds or first-party APIs documents a way to run the newest Claude, GPT or Gemini models with inference processed inside Canada. Canadian Regions on Amazon Bedrock route these models through a US geography profile, Azure offers in-country processing only up to GPT-5.2 (on reserved capacity), and Google's Montréal endpoint stops at Gemini 3.5 Flash. Japan has more options: Amazon Bedrock keeps Claude Opus 5.5 and Haiku 5.5 inside Japan through its Tokyo–Osaka (JP) profile. OpenAI's Canada and Japan regions store data at rest in-country but do not process inference there.
- Availability as of
- Update cadence
- Checked monthly against each provider's documentation; availability changes often
- First published
- Data
- AWS, Microsoft, Google, OpenAI and Anthropic documentation
- Compiled by
- Innovation Trigger Lab
Provider tables change often. Every cell on this page reflects the documentation as read on October 14, 2026. Check the linked page before you commit to an architecture.
Key facts
0
of the three first-party APIs (OpenAI, Anthropic, Google Gemini) process inference in Canada or Japan
OpenAI stores data at rest in both countries
US
the only Bedrock geography profile available from ca-central-1 and ca-west-1
Destinations: the Canadian source Region plus 3 US Regions
Tokyo + Osaka
Bedrock's JP profile keeps Claude Opus 5.5 and Haiku 5.5 inside Japan
Claude Opus 4.8/4.7 and Haiku 4.5 also In-Region in Tokyo
GPT-5.2 / 5.4
newest Azure OpenAI models with in-geography processing in Canada / Japan
Regional Provisioned (reserved PTU) only
In-country processing available? Canada and Japan by provider
The best documented option for keeping inference inside each country. Details, footnotes and exact quotes are in the per-cloud tables below.
| Provider | Canada | Japan |
|---|---|---|
| AWS Bedrock | Cross-border zoneNo current Claude, GPT or Nova model in-country; they route via the US geo profile (Canada + US). In-Region only for older open models (Llama 3, Mistral Large, Mixtral). | In-countryYes: Claude Opus 5.5, Haiku 5.5 and Sonnet 4.6 via the JP geo profile (Tokyo + Osaka); Claude Opus 4.8/4.7 and Haiku 4.5 In-Region in Tokyo; open-weight models In-Region in Tokyo. |
| Azure OpenAI / Foundry | In-countryLimited: GPT-4.1 mini and GPT-4o pay-per-token in Canada East; up to GPT-5.2 with Regional Provisioned (PTU). GPT-5.4 and newer are global only. No data zone covers Canada. | In-countryLimited: GPT-4.1 mini and GPT-4o pay-per-token in Japan East; up to GPT-5.4 with Regional Provisioned (PTU). GPT-5.6 Sol and GPT-6.1 Sol reach at most the APAC data zone; other newer models are global only. |
| Google Cloud (Gemini Enterprise Agent Platform / Vertex AI) | In-countryYes for Gemini 3.5 Flash, 2.5 Pro and 2.5 Flash at Montréal. Gemini 4 Argon, Gemini 3.6–3.8 Flash and Claude are global or US/EU only. | In-countryGemini 2.5 Flash at Tokyo. Gemini 3.5 Flash and 2.5 Pro are marked for Japan on one Google page but not listed at the Tokyo endpoint. Claude is not offered in Japan. |
| OpenAI API (direct) | Storage onlyData at rest in Canada (ca.api.openai.com); inference not regional. | Storage onlyData at rest in Japan (jp.api.openai.com); inference not regional. |
| Anthropic Claude API (direct) | Not offeredUS or global inference only. | Not offeredUS or global inference only. |
| Google Gemini API (direct) | Not offeredNo residency option documented. | Not offeredNo residency option documented. |
Tables scroll horizontally on small screens.
How to read the tables
- In-country
- The provider documents that inference stays inside the country (one Region, or several Regions all in that country).
- Cross-border zone
- Processing stays inside a defined zone that also includes other countries (for example a US geography profile, or Azure's Asia Pacific data zone).
- Global only
- Available in that Region, but only through global routing, which may process the request anywhere.
- Storage only
- Data at rest stays in the country, but inference may run elsewhere.
- Docs disagree
- Two pages from the same provider disagree. We show both readings in the footnote.
- Not offered
- Not listed for that Region or country.
Amazon Bedrock
Bedrock offers three routing options per model: In-Region (one Region), Geographic cross-Region inference (a named geography such as US or JP) and Global. Canada has no Canada-only geography, so current models called from Canadian Regions either use the US profile or global routing.
| Model family | ca-central-1 (Canada Central) | ca-west-1 (Calgary) | ap-northeast-1 (Tokyo) | ap-northeast-3 (Osaka) |
|---|---|---|---|---|
| Claude Opus 5.5, Claude Haiku 5.5 [A2, A3] | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Cross-border zoneUS geo profile (ca-west-1 + 3 US Regions) or Global | In-countryJP geo profile (Tokyo + Osaka) or Global | In-countryJP geo profile (Tokyo + Osaka) or Global |
| Claude Sonnet 5.5, Opus 5, Sonnet 5, Fable 5, Fable 5.1 [A2, A4] | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Cross-border zoneUS geo profile (ca-west-1 + 3 US Regions) or Global | Global onlyGlobal cross-Region only | Global onlyGlobal cross-Region only |
| Claude Opus 4.8, Claude Opus 4.7 [A1, A3] | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Cross-border zoneUS geo profile (ca-west-1 + 3 US Regions) or Global | In-countryIn-Region (bedrock-mantle endpoint) or JP geo profile | In-countryJP geo profile (Tokyo + Osaka) or Global |
| Claude Haiku 4.5 [A1] | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Global onlyGlobal cross-Region only | In-countryIn-Region (bedrock-mantle endpoint) or JP geo profile | In-countryJP geo profile (Tokyo + Osaka) or Global |
| Claude Sonnet 4.6 | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Cross-border zoneUS geo profile (ca-west-1 + 3 US Regions) or Global | In-countryJP geo profile (Tokyo + Osaka) or Global | In-countryJP geo profile (Tokyo + Osaka) or Global |
| OpenAI GPT-6.1 Sol, GPT-6 Sol / Astra / Luna (on Bedrock) | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Cross-border zoneUS geo profile (ca-west-1 + 3 US Regions) or Global | Global onlyGlobal cross-Region only | Global onlyGlobal cross-Region only |
| Amazon Nova 2 Lite | Cross-border zoneUS geo profile (ca-central-1 + 3 US Regions) or Global | Cross-border zoneUS geo profile (ca-west-1 + 3 US Regions) or Global | In-countryJP geo profile (Tokyo + Osaka) | Not offeredNot listed |
| Open-weight: gpt-oss-120b/20b, DeepSeek V3.2, Qwen3 235B, Mistral Large 3, Kimi K2.5, GLM 5 | Not offeredNot listed | Not offeredNot listed | In-countryIn-Region (single Region) | Not offeredNot listed |
| Older open models: Llama 3 8B/70B Instruct, Mistral Large, Mixtral 8x7B, Mistral 7B | In-countryIn-Region (single Region) | Not offeredNot listed | Not offeredNot listed | Not offeredNot listed |
Tables scroll horizontally on small screens.
Notes and exact wording
- A1
In-Region is the only Bedrock option that pins processing to one Region. For Claude in Tokyo it is listed only on the bedrock-mantle endpoint (Anthropic Messages API on Bedrock), not on bedrock-runtime. Anthropic's Bedrock page says regional endpoints carry a 10% premium over global.
Source wording:
Your requests never leave the AWS Region you specify.
— Regional availability by models (Amazon Bedrock User Guide) - A2
From both Canadian Regions the only geographic profile is US, whose destinations include three US Regions. Anthropic's Bedrock Regions table lists ca-central-1 as 'Global, US' and ca-west-1 as 'Global'. No Canada-only profile is documented.
Source wording:
ca-central-1 (Canada) | ca-central-1 (Canada), us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon)
— Amazon Bedrock model cards: Claude Opus 5.5, Opus 4.8, Opus 4.7, Haiku 4.5 (Geo inference details) - A3
The JP geo profile (model IDs prefixed jp.) routes only between Tokyo and Osaka, so processing stays in Japan. AWS also states that prompts and outputs may be stored in the destination Region for abuse detection.
Source wording:
ap-northeast-1 (Tokyo) | ap-northeast-1 (Tokyo), ap-northeast-3 (Osaka)
— Amazon Bedrock model cards: Claude Opus 5.5, Opus 4.8, Opus 4.7, Haiku 4.5 (Geo inference details) - A4
Global cross-Region inference can send a request to any supported commercial AWS Region. AWS recommends Geographic profiles when you have data residency requirements.
Source wording:
Routed worldwide
— Route model inference requests across AWS Regions with cross-Region inference
Azure OpenAI and Microsoft Foundry
Azure separates where data is stored (always the resource's geography) from where it is processed (set by the deployment type). Only Standard and Regional Provisioned deployments keep processing in the Canadian or Japanese geography. Microsoft says geography-based types arrive last for new models, with no guaranteed date.
| Model family | Canada East | Canada Central | Japan East | Japan West |
|---|---|---|---|---|
| GPT-6.1 Sol [B2] | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | Cross-border zoneData Zone Standard (APAC) or Global | Not offeredAzure OpenAI not listed |
| GPT-6 Sol, GPT-6 Astra, GPT-6 Luna | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | Not offeredAzure OpenAI not listed |
| GPT-5.6 Sol [B2] | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | Cross-border zoneData Zone Standard (APAC) or Global | Not offeredAzure OpenAI not listed |
| GPT-5.5 | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | Not offeredAzure OpenAI not listed |
| GPT-5.4, GPT-5.4 mini [B1] | Global onlyGlobal Standard / Global Provisioned only | Global onlyGlobal Standard / Global Provisioned only | In-countryRegional Provisioned (PTU) only; also Data Zone (APAC), Global | Not offeredAzure OpenAI not listed |
| GPT-5.2 [B1] | In-countryRegional Provisioned (PTU) only | In-countryRegional Provisioned (PTU) only | In-countryRegional Provisioned (PTU) only | Not offeredAzure OpenAI not listed |
| GPT-5.1 [B1] | In-countryRegional Provisioned (PTU) only | Global onlyGlobal Standard / Global Provisioned only | In-countryRegional Provisioned (PTU) only | Not offeredAzure OpenAI not listed |
| GPT-4.1 mini [B1] | In-countryStandard (pay-per-token) | Global onlyGlobal Provisioned only | In-countryStandard (pay-per-token); Regional Provisioned | Not offeredAzure OpenAI not listed |
| GPT-4o (2024-11-20) [B1] | In-countryStandard (pay-per-token); Regional Provisioned | In-countryRegional Provisioned, older 2024-05-13 version only | In-countryStandard (pay-per-token); Regional Provisioned | Not offeredAzure OpenAI not listed |
| Other Foundry Models sold by Azure: DeepSeek V3.2/V4, Llama 3.3 70B, Mistral Large 3, Grok 4.7, Kimi K2.6, Cohere Command A, MAI-Thinking-1 [B3] | Global onlyGlobal Standard only | Global onlyGlobal Standard only | Global onlyGlobal Standard only | Global onlyGlobal Standard only |
Tables scroll horizontally on small screens.
Notes and exact wording
- B1
Applies to Standard and Regional Provisioned deployments only. Microsoft adds that processing 'might be' moved between regions within that geography for operational purposes. Regional Provisioned means buying reserved throughput units (PTUs); it is not pay-per-token.
Source wording:
Prompts and responses are processed within the customer-specified Azure geography
— Deployment types for Microsoft Foundry Models - B2
Data Zone deployments from Japan East use the Asia Pacific zone, not a Japan-only zone. Microsoft defines data zones for the US, EU and APAC only; Canada is in none of them.
Source wording:
data processed within any Asia Pacific nation
— Region availability for Foundry Models sold by Azure - B3
Storage at rest stays in the resource's geography for every deployment type, but Global types 'may be processed in any Azure region'. The Standard/Regional table lists no Foundry Models other than Azure OpenAI in any region.
Source wording:
Data stored at rest remains in the designated Azure geography.
— Deployment types for Microsoft Foundry Models
Google Cloud (Gemini Enterprise Agent Platform, formerly Vertex AI)
Google publishes a per-model table of where ML processing happens. Canada and Japan each have one generative AI endpoint location: Montréal and Tokyo. Claude and the newest Gemini models are offered on the global endpoint and US/EU multi-regions only.
| Model family | Montréal (northamerica-northeast1) | Tokyo (asia-northeast1) |
|---|---|---|
| Gemini 4 Argon, Gemini 3.8 / 3.7 / 3.6 Flash, Gemini 3.5 Flash-Lite [C1] | Global onlyGlobal endpoint or US/EU multi-region | Global onlyGlobal endpoint or US/EU multi-region |
| Gemini 3.5 Flash [C2, C3] | In-countryMontréal endpoint | Docs disagreeJapan listed in residency table; not listed as a Tokyo endpoint |
| Gemini 2.5 Pro [C2, C3] | In-countryMontréal endpoint (64k and 1M context) | Docs disagreeJapan listed for 64k context only; not listed as a Tokyo endpoint |
| Gemini 2.5 Flash [C2] | In-countryMontréal endpoint (128k and 1M context) | In-countryTokyo endpoint (128k context; 1M not listed) |
| Claude on Google Cloud (Opus 5.5, Sonnet 5.5, Haiku 5.5 and earlier) [C1] | Global onlyGlobal endpoint or US/EU multi-region | Global onlyGlobal endpoint or US/EU multi-region |
| Open models (gpt-oss, DeepSeek, Qwen3, Llama, GLM, Kimi) | Not offeredNot listed at Montréal | Not offeredNot listed at Tokyo |
Tables scroll horizontally on small screens.
Notes and exact wording
- C1
Said of the global endpoint (aiplatform.googleapis.com). US and EU multi-region endpoints keep processing inside those jurisdictions, which does not help a Canada- or Japan-only requirement.
Source wording:
don't provide any data residency guarantees
— Data residency (Gemini Enterprise Agent Platform, formerly Vertex AI) - C2
Said of locational endpoints such as northamerica-northeast1. Data at rest stays in the chosen location regardless of endpoint. Toronto (northamerica-northeast2) and Osaka (asia-northeast2) are not listed as generative AI endpoints; asia-northeast3 is Seoul, not Japan.
Source wording:
ML processing remains entirely within the broader multi-regional or country jurisdiction
— Data residency (Gemini Enterprise Agent Platform, formerly Vertex AI) - C3
Google's endpoint list shows Gemini 2.5 Flash as the only Gemini model at Tokyo (plus embedding models), while the data residency page also marks Japan for Gemini 3.5 Flash and Gemini 2.5 Pro (64k). We report both and treat the Japan cell as unconfirmed until the pages agree.
Source wording:
Tokyo (asia-northeast1)
— Deployments and endpoints (Gemini Enterprise Agent Platform, formerly Vertex AI)
First-party APIs (OpenAI, Anthropic, Google)
Calling a model provider directly is the simplest setup, but none of the three documents in-country inference for Canada or Japan. OpenAI is the only one with Canadian and Japanese storage regions.
| Model family | Canada | Japan |
|---|---|---|
| OpenAI API [D1] | Storage onlyca.api.openai.com: regional storage yes, regional processing no | Storage onlyjp.api.openai.com: regional storage yes, regional processing no |
| Anthropic Claude API [D2] | Not offeredinference_geo is US or global only | Not offeredinference_geo is US or global only |
| Google Gemini API (AI Studio) [D3] | Not offeredNo residency option documented | Not offeredNo residency option documented |
Tables scroll horizontally on small screens.
Notes and exact wording
- D1
Applies to regions without regional processing, which includes Canada and Japan (only the US, Europe and the UAE are marked for regional processing). Non-US regions also require approval for abuse-monitoring controls and a Modified Retention amendment; eligible models released on or after March 5, 2026 carry a 10% uplift.
Source wording:
OpenAI may also process and temporarily store Customer Content outside of the Region
— Data controls in the OpenAI platform: Data residency controls - D2
Workspace geo (where data is stored at rest) is also US-only. US-only inference on Claude 4.6 and later models is priced at 1.1x.
Source wording:
Only "us" and "global" are available.
— Data residency (Claude API) - D3
The Gemini API page lists countries where the service can be used (Canada and Japan are included) but no processing-location control. Google points enterprise customers to its cloud platform, covered in the Google Cloud table.
Source wording:
try the Gemini API in Gemini Enterprise Agent Platform
— Available regions for Google AI Studio and Gemini API
Methodology and definitions
- Processing vs data at rest
- This tracker classifies each option by where inference runs (the prompt and response are processed), not where files, logs or fine-tuned weights are stored. Several providers keep data at rest in your chosen location while processing elsewhere, and some store prompts in the processing Region for abuse monitoring.
- What counts as in-country
- A cell is "In-country" only when the provider's documentation names a single Region or a set of Regions that are all in that country (for example Bedrock's JP profile, which routes between Tokyo and Osaka). Anything that can route to another country is "Cross-border zone" or "Global only", even if the request starts in Canada or Japan.
- Scope
- Text and multimodal LLMs that the provider itself serves (pay-per-token or reserved capacity). Self-hosting open-weight models on your own Canadian or Japanese instances is always possible and is not tracked here. Embedding, speech and image models are mentioned only where they help explain a Region's coverage. Claude in Microsoft Foundry and sovereign or government clouds are not covered.
- How cells were read
- Every cell was read from the provider pages in the sources list on the date shown. For Bedrock, Canadian and Japanese geo-profile destinations were taken from each model's card. Where two pages from one provider disagree, both readings are shown. Model lists are grouped by family; a group row applies only to the models it names.
- Not legal advice
- This page reports what providers document about processing locations. It does not say whether any option satisfies PIPEDA, provincial law, Japan's APPI or a contract. Ask your privacy or legal adviser, and check the regulators' own guidance linked below.
Regulatory context
Canada's federal private-sector privacy law (PIPEDA) and Japan's Act on the Protection of Personal Information (APPI) both address transfers of personal information outside the country. Many organizations also face sector rules, provincial laws or customer contracts that require in-country processing. The regulators' own pages:
Sources
Primary documentation only. Titles are given as published. Several providers do not print a last-updated date; for those, the date shown is when we read the page.
- Regional availability by models (Amazon Bedrock User Guide)
Amazon Web Services · No date shown on page; checked
- Amazon Bedrock model cards: Claude Opus 5.5, Opus 4.8, Opus 4.7, Haiku 4.5 (Geo inference details)
Amazon Web Services · No date shown on page; checked
- Route model inference requests across AWS Regions with cross-Region inference
Amazon Web Services · No date shown on page; checked
- Geographic cross-Region inference
Amazon Web Services · No date shown on page; checked
- Supported Regions and models for inference profiles
Amazon Web Services · No date shown on page; checked
- Claude in Amazon Bedrock (Opus 4.7 and later): Regions
Anthropic · No date shown on page; checked
- Region availability for Foundry Models sold by Azure
Microsoft · Page last updated
- Deployment types for Microsoft Foundry Models
Microsoft · Page last updated
- Data residency (Gemini Enterprise Agent Platform, formerly Vertex AI)
Google Cloud · Page last updated
- Deployments and endpoints (Gemini Enterprise Agent Platform, formerly Vertex AI)
Google Cloud · Page last updated
- Data controls in the OpenAI platform: Data residency controls
OpenAI · No date shown on page; checked
- Data residency (Claude API)
Anthropic · No date shown on page; checked
- Available regions for Google AI Studio and Gemini API
Google · Page last updated
Frequently asked questions
Can I use Claude on Amazon Bedrock with data kept in Canada?
Not for current models, as of October 14, 2026. From ca-central-1 and ca-west-1, Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, Opus 4.8 and other current versions are offered through the US geographic profile, which can route to us-east-1, us-east-2 and us-west-2 as well as the Canadian source Region, or through global routing. AWS documents no Canada-only profile and no In-Region option for these models in Canada.
Can I run Claude on Bedrock in the Tokyo Region with processing kept in Japan?
Yes for some versions. Claude Opus 5.5, Haiku 5.5 and Sonnet 4.6 can use the JP geographic profile (model IDs starting with jp.), which routes only between Tokyo (ap-northeast-1) and Osaka (ap-northeast-3). Claude Opus 4.8, Opus 4.7 and Haiku 4.5 are also In-Region in Tokyo through the bedrock-mantle endpoint. Claude Sonnet 5.5, Opus 5, Sonnet 5 and Fable 5/5.1 are global only from Japan.
Which Azure OpenAI models are available in Canada East or Japan East with in-country processing?
Pay-per-token Standard deployments are listed for GPT-4.1 mini and GPT-4o (2024-11-20) in both Canada East and Japan East. With Regional Provisioned (reserved PTU) capacity, Canada East goes up to GPT-5.2 and Japan East up to GPT-5.4. GPT-5.5, GPT-5.6 and the GPT-6 family are Global only in Canada; in Japan East, GPT-5.6 Sol and GPT-6.1 Sol also have Data Zone deployments, which process anywhere in Asia Pacific.
Does OpenAI's Canada or Japan data residency keep processing in the country?
No. OpenAI's table lists Canada (ca.api.openai.com) and Japan (jp.api.openai.com) with regional storage but without regional processing, and states that for such regions it may process and temporarily store customer content outside the region. Only the United States, Europe and the UAE are marked for regional processing.
Can I use Gemini on Google Cloud with processing in Canada or Japan?
In Canada, Gemini 3.5 Flash, Gemini 2.5 Pro and Gemini 2.5 Flash are listed with ML processing in Canada at the Montréal (northamerica-northeast1) endpoint. In Japan, Gemini 2.5 Flash is available at the Tokyo (asia-northeast1) endpoint; Google's data residency page also marks Japan for Gemini 3.5 Flash and Gemini 2.5 Pro, but its endpoint list does not, so confirm before relying on them. Newer Gemini models and Claude on Google Cloud are global or US/EU only.
Does in-country processing cost more?
Often. Anthropic's Bedrock page says regional endpoints cost 10% more than global ones, and AWS says global cross-Region inference is about 10% cheaper than geographic for some models. OpenAI charges a 10% uplift on data residency endpoints for eligible models released on or after March 5, 2026. On Azure, the newest models reach in-country processing only through reserved PTU capacity rather than pay-per-token.
How often is this tracker updated, and can I reuse it?
We re-check every source monthly and when a provider announces a regional launch; it was last checked on October 14, 2026. You are welcome to quote it with a link to this page. Always confirm against the provider's current documentation, which may have changed since.
Cite this page
Writing about AI data residency in Canada or Japan? Copy a citation or a link below, and include the as-of date: availability changes often.
Citation
Innovation Trigger Lab. (October 14, 2026). Which LLMs can run in Canada or Japan? Data residency tracker [Data set]. Based on AWS, Microsoft, Google, OpenAI and Anthropic documentation. https://www.itl-inc.ca/en/research/llm-data-residency-canada-japanHTML link
<a href="https://www.itl-inc.ca/en/research/llm-data-residency-canada-japan">LLM data residency tracker for Canada and Japan (Innovation Trigger Lab)</a>Need AI that keeps data in-country?
Innovation Trigger Lab designs and builds AI systems around residency constraints: choosing the cloud, Region and routing for each workload, retrieval over your own documents, and self-hosted open-weight models where no managed option fits.