Data Residency and Sovereign Cloud for AI in Saudi Arabia and the UAE

Cloud and model availability in this article is as of 29 September 2026 and changes often. INVISOM has no commercial relationship with any provider named here. This is general guidance, not legal advice.

Hosted in the region is not processed in the region

Ask a Gulf enterprise where its AI assistant runs and the answer is usually a cloud region: UAE North, me-central-1, Dammam. That answers where the data is stored. It rarely answers where the prompt is processed, and for large language models those are now different places.

The providers say so in their own documentation. Microsoft states that for every Foundry deployment type, data at rest stays in the chosen Azure geography, but that "Global" deployment types may process inference in any Azure region. Amazon Bedrock's global cross-Region inference sends requests to supported commercial AWS Regions worldwide, while CloudTrail keeps logging in your source Region. Google tells Vertex AI customers not to use the global endpoint if they have ML processing requirements, because they cannot control or know which region processes the request.

As of 29 September 2026, almost every frontier model reachable from a Gulf cloud region is offered through exactly that kind of global routing. A UAE bank that stores its documents in Abu Dhabi and calls Claude on Bedrock from me-central-1 is sending each prompt, with whatever customer data it contains, to wherever AWS chooses to run it.

Our argument is that Gulf AI residency should be decided per workload and per feature, by asking two separate questions of every component: where is it stored, and where is it processed?

The storage answer is almost always "in region", because all the major hyperscalers now store data in the UAE or Saudi Arabia. The inference answer is the one that decides whether a workload is compliant, and it is the one most architecture reviews skip.

The regulatory map: sector and data class decide

Neither country has a single rule that says "AI data must stay here". Localisation is triggered by a mix of privacy law, data classification, sector regulation and free-zone regimes. Several Saudi primary texts could not be retrieved for this article, so where we rely on law-firm analyses we say so.

Saudi Arabia. Law-firm analyses report that the Personal Data Protection Law (PDPL) became enforceable in September 2024 after a one-year transition. According to Mayer Brown's analysis, SDAIA's amended Regulation on Personal Data Transfer, issued in September 2024, allows transfers abroad through adequacy, "appropriate safeguards" (four modules of SDAIA standard contractual clauses that are invalid if modified, Binding Common Rules, or a certificate of accreditation) and listed exemptions. A transfer risk assessment is required when relying on safeguards, or for continuous or large-scale transfers of sensitive data. Clyde & Co reports that SDAIA issued a risk assessment guideline in early 2025.

The cybersecurity layer has shifted. According to Clyde & Co, the National Cybersecurity Authority's Essential Cybersecurity Controls (ECC-2:2024) removed the explicit in-country hosting requirements and moved localisation to the National Data Management Office (NDMO) at SDAIA, which entities should consult before acting. The same analysis notes that the ICT regulations for government entities still require government data to be hosted in the Kingdom. Security teams that still cite "NCA requires in-Kingdom hosting" are working from an old version.

We have not found a published NDMO instrument that maps each data classification level to a permitted hosting model, and we do not offer one here. Tables circulating online that do so should be checked against NDMO or the data-owning entity before they drive design. Our NDMO implementation guide covers the classification process itself.

For banks, the Saudi Central Bank (SAMA) is explicit. Its Cyber Security Framework says that, in principle, only cloud services located in Saudi Arabia should be used, and that using cloud services outside the Kingdom requires explicit SAMA approval.

UAE. The federal PDPL (Federal Decree-Law 45 of 2021) has been in force since 2 January 2022, with the UAE Data Office as regulator. The government portal, last updated December 2025, does not refer to issued Executive Regulations, and law-firm analyses in 2026 describe them as still pending, along with any adequacy list. In practice, sector rules often matter more than the PDPL. According to law-firm summaries of the Central Bank's Outsourcing Regulation, banks must keep their Master System of Record, including confidential data, in the UAE, and need CBUAE approval and customer consent to share customer confidential data abroad. Al Tamimi's analysis of the federal ICT in Health Fields law describes a similar in-country requirement for health data, with listed exceptions.

Free zones. The DIFC and ADGM run their own data protection regimes. The DIFC's Regulation 10, effective 1 September 2023, specifically addresses personal data processed through autonomous and semi-autonomous systems, and is one of the few AI-specific privacy rules in the region.

The cloud region landscape

As of 29 September 2026, the in-country options look like this:

Two Saudi details matter for AI. Google Cloud says it holds a Communications, Space and Technology Commission (CST) Class C licence for Dammam, and that customers with a Saudi billing address must buy through CNTXT, which offers "Sovereign Controls" packages with CNTXT-managed external key management and Key Access Justifications. And Amazon has said that, with HUMAIN, it plans up to 50 MW of AI capacity in a Saudi "AI Zone" by 2028.

Neither Microsoft's nor Amazon's Saudi announcement says which AI services will be available at launch. Microsoft's own documentation explains why that matters: new models arrive on Global deployment types first, then Data Zone, and geography-based types last, with no guaranteed date. A new region is not the same as a new in-region model catalogue.

What you can actually run in-region today

The matrix below is the core of the problem. It shows where inference runs when you call each service from a Gulf region.

Four findings stand out:

• Azure UAE North is the only hyperscaler route to an OpenAI frontier model processed in the UAE, and only through Regional Provisioned (PTU) capacity. Microsoft lists eight OpenAI models there, including gpt-5.1 and the recently released gpt-6-sol. Pay-as-you-go Standard deployments in UAE North cover only embeddings and Whisper. Every other model, including 47 partner models, is Global. There is no Middle East Data Zone; Microsoft's zones cover the US, EU and Asia Pacific only.
• Amazon Bedrock offers no in-region frontier model in the Gulf. Claude models from Haiku 4.5 to Opus 5.5, GPT-5.6 and Grok are listed as Global cross-Region only from me-central-1 and me-south-1. Only two Amazon Nova models run in-Region in the UAE.
• Vertex AI in Doha and Dammam offers embeddings only. No Gemini or partner model has a regional endpoint in either region.
• Direct APIs differ. OpenAI announced UAE inference residency in August 2026 for eligible API, ChatGPT Enterprise and ChatGPT Edu customers, making the UAE one of three such locations alongside the US and Europe. Confirm model scope and eligibility in your contract. Anthropic's API accepts only "global" or "us" as inference geographies and stores workspace data in the US, so there is no Gulf option.

This pattern repeats a point we made about data platforms in our Snowflake and Databricks analysis: the platform being available in Dubai or Dammam is not a residency answer. Each AI feature has to be checked.

Four architecture patterns

Given that matrix, the practical question is how to get the benefit of frontier models without moving data that is not allowed to move.

1. In-region RAG, with the model chosen per data class. Keep the corpus, vector index, embeddings, retrieval, access control and logs in the region. In-region embedding models exist on Azure UAE North and Vertex AI Doha and Dammam, so the index never needs to leave. Only the retrieved context is sent to the model, and the gateway chooses the route by data class: an in-region provisioned model for restricted content, a global model for content cleared for transfer. Our comparison of RAG, fine-tuning and custom models explains why retrieval is the right default for enterprise knowledge in any case.

2. Open-weight models on in-region GPUs. For government data and sector-regulated data without a transfer approval, self-hosting an open-weight model on GPU capacity in the region keeps inference local. Note a trap: Azure lists Llama-3.3-70B and DeepSeek-V3.2 as available "in" UAE North, but as Global Provisioned deployments, which are processed globally. Self-hosting on in-region virtual machines or managed compute is a different path, and the operating burden (capacity, patching, evaluation) becomes yours.

3. Pseudonymise and minimise before transfer. Where a transfer basis exists, reduce what crosses the border. Replace names, account numbers and national IDs with tokens before the call, keep the mapping in the region, and re-identify the response locally. This is a risk-reduction measure, not a legal basis; pseudonymised data can still be personal data.

4. Block global routing by policy. Do not rely on developer discipline. Microsoft publishes an Azure Policy rule that denies a deployment SKU such as GlobalStandard. AWS documents a service control policy that denies Bedrock global inference profiles, and CloudTrail records an inferenceRegion field showing where each cross-Region request was actually processed. Anthropic's API lets administrators restrict a workspace's allowed inference geographies. Turn these on first, then grant explicit exceptions.

Key custody is a separate control. External key managers (AWS KMS External Key Store, Google Cloud EKM with Key Access Justifications, Azure Managed HSM) govern who can decrypt stored data. They do not change where a model processes a prompt.

A decision guide and a worked example

Run every AI use case through the same sequence and stop at the first trigger that applies.

To show how this plays out, consider an illustrative UAE retail bank building a customer-service assistant that answers from product documents and, when authenticated, from the customer's account data.

The design runs two model routes behind one gateway. That costs more than a single global endpoint. It is also the only version a bank's regulator is likely to accept.

What auditors will ask for

Residency claims fail on evidence more often than on design. A short weak-versus-strong comparison helps teams prepare.

What to do in the next 90 days

1. Inventory every AI call. For each workload, list the model, endpoint, deployment type or inference profile, and the actual processing region.
2. Classify the data each call carries, using the national scheme for government data and the sector rule for banks and health, not only the PDPL.
3. Turn on policy guardrails that deny global deployment types and inference profiles by default, with a logged exception process.
4. Build or adopt a model gateway that routes by data class, so the in-region and global routes sit behind one interface.
5. Document transfer bases per vendor, including Saudi SCCs and risk assessments where you rely on them, and confirm UAE positions with counsel given the pending Executive Regulations.
6. Re-test region availability monthly. Azure Saudi Arabia East and the AWS Saudi region launch in Q4 2026; plan a review of placement decisions in early 2027 rather than assuming launch-day AI coverage.
7. Assign ownership for residency decisions in your AI governance forum, alongside the controls described in our enterprise AI operating model.

Frequently asked questions

Does hosting an AI application in a UAE or Saudi cloud region keep the data in-country?

It keeps stored data in-country. It does not necessarily keep inference in-country. As of 29 September 2026, most frontier models reachable from Gulf regions on Azure, Amazon Bedrock and Vertex AI use global routing, which can process prompts outside the region.

Can I run a frontier LLM entirely inside the UAE or Saudi Arabia today?

In the UAE, yes, with limits: Azure UAE North offers eight OpenAI models through provisioned capacity, and OpenAI announced UAE inference residency for eligible customers in August 2026. In Saudi Arabia, no hyperscaler offered an in-region frontier model endpoint on that date; self-hosting open-weight models on local GPUs is the main option.

Does the NCA still require Saudi data to be hosted in the Kingdom?

According to law-firm analyses of the 2024 control updates, the NCA removed its explicit hosting controls and referred localisation to NDMO. Government data hosting rules and sector rules such as SAMA's still apply, so many workloads must stay in the Kingdom.

How do I stop developers from using global AI endpoints by accident?

Use platform policy. Azure Policy can deny Global deployment SKUs, AWS service control policies can deny Bedrock global inference profiles, and Anthropic workspaces can restrict allowed inference geographies. Combine this with logs that record the actual processing region.

Sources and further reading

• Region availability for Foundry Models sold by Azure, Microsoft Learn, 3 September 2026. UAE North model lists by deployment type.
• Understanding deployment types in Microsoft Foundry Models, Microsoft Learn, 6 August 2026. Processing locations, Data Zones, launch order and Azure Policy example.
• Global cross-Region inference and cross-Region inference, AWS documentation. Worldwide routing, SCP controls and CloudTrail inferenceRegion.
• Regional availability by model and Claude Sonnet 4.6 model card, AWS documentation.
• Generative AI locations, Google Cloud, updated 29 September 2026. Doha and Dammam model availability and global endpoint guidance.
• Dammam region access, Google Cloud, updated 24 September 2026. CNTXT, Sovereign Controls and CST licence.
• Data residency, Anthropic documentation.
• OpenAI expands inference residency to the United Arab Emirates, OpenAI press release via Zawya, August 2026.
• Microsoft Saudi Arabia East region announcement, Microsoft, 31 August 2026.
• AWS cloud region in Saudi Arabia, Amazon.
• Cyber Security Framework, 3.4.3 Cloud Computing, SAMA Rulebook.
• Updates to Saudi Arabia's personal data protection regulations, Mayer Brown, 9 October 2024.
• Saudi Arabia: Essential Cybersecurity Controls, Clyde & Co, 16 January 2025; Update on Saudi Arabia risk assessment guidelines, Clyde & Co, March 2025.
• Data protection laws, UAE Government portal, updated 4 December 2025.
• Regulation 10: autonomous and semi-autonomous systems, DIFC.

This article is general guidance, not legal advice. Confirm specific obligations with counsel and against current SDAIA, NDMO, NCA, SAMA, UAE Data Office, CBUAE, DIFC and ADGM publications.