AWS, Microsoft and Google all have South African cloud regions. Having a region is not the same as running the model there. This page shows where each vendor actually processes your prompts, checked against their own documentation on 11 October 2026.
“In country” here means the GPU that runs the model is in South Africa, not just the API endpoint, the logs or the stored data.
| Vendor · SA region | Models you can call | Where inference runs | In country? |
|---|---|---|---|
| AWS Bedrockaf-south-1 · Cape Town | Claude, current line (Sonnet 5.5, Opus 5.5) and earlier | Global cross-Region only: any AWS region with capacity. The endpoint and CloudWatch/CloudTrail logs stay in Cape Town. | No |
| Azure OpenAIsouthafricanorth · Johannesburg | Current GPT line (GPT-5.x, GPT-6) on Global Standard and Global Provisioned | Global types run in any Azure region. Regional (in-country) inference covers gpt-4o 2024-05-13 on Provisioned only, plus two embedding models. | gpt-4o PTU only, and that version is scheduled to retire on 9 Dec 2026 |
| Google Vertex AIafrica-south1 · Johannesburg | Gemini is not listed for africa-south1 | Gemini is served from other regions or the global endpoint. The region does support custom model inference. | No for Gemini; yes for models you deploy yourself |
| Open-weights, self-hostedany SA region or on-prem | Llama, Qwen, Gemma, GLM, Mistral and others | On the compute you choose, served with e.g. vLLM or Ollama | Yes |
As of October 2026, no major vendor runs a current frontier model in South Africa. The only managed in-country option is Azure’s gpt-4o (2024-05-13) on Regional Provisioned Throughput, and that version is scheduled to retire on 9 December 2026. If prompts must stay in the country, plan on open-weights models running on South African compute. Otherwise, get a written regional-capacity commitment from Microsoft before you build on the PTU path.
The vendor question comes second. First decide whether prompts and grounding data may cross the border. That’s a legal call for your information officer, not an architecture preference.
This covers most workloads where a POPIA section 72 transfer basis applies, for example an adequate level of protection at the destination or the data subject’s consent.
af-south-1, southafricanorth) still helps: logs and data at rest stay local, and latency to the endpoint is lower.Typical cases are bank offshoring rules, public-sector and sovereign data, client PII in FAIS-regulated advice, and contracts that name the country.
Bedrock went GA in Cape Town in November 2025. You can call Claude from af-south-1, but only through global cross-Region inference. AWS’s own model cards for the current Claude line list af-south-1 as Global only: In-Region and Geo are both marked not supported. There is no Africa geo profile, so global routing is the only option from Cape Town.
“You can now invoke models from the Cape Town Region while Amazon Bedrock automatically routes requests to Regions with available capacity… your Amazon CloudWatch and AWS CloudTrail logs stay centralized in af-south-1.” — AWS Machine Learning Blog, Jan 2026
In plain terms, the API endpoint and the audit logs are in Cape Town. The GPU running Claude is wherever AWS has capacity at that moment.
Every Azure OpenAI deployment keeps data at rest in your Azure geography. Where the prompt is processed depends on the deployment type you choose.
| Deployment type | Where inference runs | Available in southafricanorth |
|---|---|---|
| Global Standard / Provisioned | Any Azure region where the model is deployed | Current GPT line, incl. GPT-5.x and GPT-6 |
| Data Zone | Within the US, EU or APAC zone | None. There is no Africa data zone |
| Standard (Regional) | Only in the deployment region | text-embedding-3-large, text-embedding-ada-002. No chat models |
| Provisioned (Regional PTU) | Only in the deployment region | gpt-4o 2024-05-13 only |
“Global types: Might be processed in any Azure region where the model is deployed… Standard/Regional types: Processed in the region associated with your deployment.” — Microsoft Learn · Region availability for Foundry Models, Sep 2026
Microsoft lists gpt-4o 2024-05-13 for retirement on 9 December 2026. Standard deployments auto-upgrade to the replacement (gpt-5.6-sol), but provisioned deployments are not auto-upgraded, and Microsoft’s documentation doesn’t yet show a successor in Regional PTU for South Africa North. Until it does, the in-country Azure path has an end date.
PTU also carries a meaningful minimum commitment, which rules it out for smaller workloads regardless of the model question.
Google’s generative AI locations page (updated 10 October 2026) does not list africa-south1 for any Gemini model. Calls from South Africa go to another region or to the global endpoint, which Google says “doesn’t support data residency requirements.”
Johannesburg does support Vertex AI custom model inference, along with Model Registry and Pipelines. So open-weights models you deploy yourself, such as Gemma or Llama on a Vertex endpoint, can serve from inside the country. That’s the in-country route on GCP.
This page started as an audit of our own overstated residency claims. We keep the record here because the facts move and the tree is held to the same standard.
agents/claude, gemini, gemini-omni, google-adk, llama, gpt, openai-agents-sdk, the agents and tech hubs, tech/aws, and the Microsoft hub, copilot-sa and azure-openai-workers pages.We corrected agents/gpt (Claude on Bedrock “for SA residency”) and tech/microsoft (“the only major LLM vendor with a South Africa inference region”; “fully in country”). Applied in PR #32, commit ba8725a.
agents/gpt — comparison table, Claude rowaf-south-1 for SA residencyaf-south-1 endpoint with in-country logs (Claude 4.5 inference itself routes globally via cross-Region inference — not strict residency)tech/microsoft — hero copysouthafricanorth. Google Vertex AI's africa-south1 and AWS Bedrock's af-south-1 are also present in country, but for current frontier closed models neither matches Azure's Regional PTU residency guarantee. For regulated workloads that have to keep prompts and grounding data in country — banking, healthcare, public sector, FAIS-covered advisory — that asymmetry is the deciding architectural input.tech/microsoft — § 02 residency footnotesouthafricanorth is fully in country — both the resource and the inference compute are in Joburg.southafricanorth keeps inference in country only on Regional PTU deployments — and currently only for the GPT model versions with local capacity. The default Global Standard SKU keeps data at rest in country but processes prompts and completions in any Microsoft Foundry region globally. The strict-residency path is Regional PTU; the pragmatic path is Global Standard with the M365 Geo data-at-rest commitment. Confirm specific model availability with Microsoft before architectural decisions — current GPT Standard SKU is documented in eight US/Sweden regions only, with SA North coverage via PTU.