Journal — № 004
AI sovereignty is a spectrum.
A practical guide to where your AI runs, who can operate it, and who owns what it learns, with the UAE as a concrete example.
“In consuming intelligence, you are creating intelligence.”
That sentence changes the sovereignty question. When a team uses AI, it sends information to a model. At the same time, it creates new information: corrections, decisions, tests, memory, and a record of how the organization works.
We will use one example throughout. A customer service team in Dubai asks an AI system to review refund requests. At first, the system reads a public returns policy. Later, staff add customer messages, order histories, internal thresholds, and corrections from experienced agents. The system starts to learn how this company handles unusual cases.
The first question is usually, “Where are the documents stored?” That question is useful. Soon, more questions appear. Where is each prompt processed? Who can access the service? Who holds the encryption keys? Can the company keep its corrections, tests, and memory if it changes providers?
This article gives that broader idea a practical shape. It explains the three controls behind AI sovereignty, the difference between laws and assurance standards, and how cloud model routing changes the boundary. The UAE is our concrete example. Readers elsewhere can apply the same method using their own country-specific and sector rules.
What does AI sovereignty mean?
AI sovereignty is the degree of control an organization has over its AI systems, the data moving through them, and the knowledge created while people use them. The right level varies by workload. A tool that rewrites public marketing copy can sit in a different place on the spectrum from a system that reviews medical records or bank transactions.
Three controls make the idea concrete.
- Data boundary. Where prompts, files, outputs, logs, embeddings, backups, and model processing can go. An embedding is a numeric representation used for search. It can still reveal information about the source data, so it belongs inside the review.
- Operational control. Who can administer the system, access support data, hold encryption keys, respond to incidents, and keep the service running.
- Intelligence ownership. Who keeps the memory, corrections, evaluation cases, traces, tuned weights, and operating knowledge created around the model. An evaluation, often shortened to “eval,” is a repeatable test used to check whether the AI behaves as expected.
A datacenter inside your chosen country can strengthen the data boundary. Operational control and intelligence ownership still need their own decisions. This is why sovereignty works better as three spectra than as one badge.
Explore the spectrum
Move each control. The notes below will change with your choices.
01
Model processing is limited to one named cloud region.
02
Your organization controls key use, while the provider still operates the service.
03
You retain prompts, memory, corrections, tests, and traces outside the model provider.
This explorer is educational. It does not determine legal or regulatory compliance. Some notes may not apply to your organization. Your entity, sector, data, contracts, vendors, and technical configuration decide what needs attention.
What does “compliant” actually tell you?
Compliance language becomes confusing because several kinds of requirement appear in the same vendor questionnaire. They do different jobs.
- Law
- A binding rule passed by a government. The UAE Personal Data Protection Law is one example. Scope and exceptions matter.
- Sector rule
- A requirement for a regulated field such as banking or healthcare. It may be stricter than the general rule and may apply only to specific entities or data.
- Standard
- A structured way to manage risk. ISO standards and the NIST AI Risk Management Framework help teams organize controls and evidence. They do not replace applicable law.
- Audit or attestation
- Independent evidence about controls over a defined period or at a point in time. A SOC 2 report covers the service provider and scope named in that report. It does not make every customer workload compliant.
- Contract and configuration
- The provider’s promises and the settings your team enables. A regional option helps only when the contract, endpoint, logs, backups, support path, and connected services follow the intended boundary.
Your slider position is a starting point
A sovereignty assessment checks it against your real AI tools, data flows, vendor contracts, cloud settings, and country or sector requirements.
How do country-specific rules apply?
Every country has its own mix of general data law, sector rules, regulator guidance, and exceptions. Start with the legal entity and the data it handles. The UAE shows how quickly those layers can change the answer.
01
UAE Personal Data Protection Law
Federal Decree-Law No. 45 of 2021 regulates the processing of personal data. Articles 22 and 23 set conditions for transfers outside the UAE. The law does not create a blanket requirement to keep every piece of personal data in the country. Transfer conditions, safeguards, and the receiving jurisdiction need review.
Article 2 also lists exclusions, including government data and entities, health data, banking and credit data covered by sector rules, and companies in free zones with their own data-protection laws. Read the official PDPL text.
02
DIFC and ADGM data-protection regimes
An entity established in the Dubai International Financial Centre or Abu Dhabi Global Market may fall under that jurisdiction’s own data-protection regime. Both have rules for international transfers and their own regulators. Check the DIFC guidance and the ADGM transfer guidance for the relevant entity.
03
Health data
Article 13 of Federal Law No. 2 of 2019 restricts storing, processing, generating, or transforming health data related to health services provided in the UAE outside the country, unless an authorized resolution allows it. That can make inference location, logs, backups, and support access especially important. See the official health data law.
04
Bank outsourcing
The Central Bank’s outsourcing standard requires a bank’s master system of record to stay in the UAE. Storing confidential customer data outside the UAE requires Central Bank approval and prior written customer consent under the stated conditions. This is a rule for regulated banks, not a general rule for every UAE business. Read Article 6 of the CBUAE standard.
05
Critical and government systems
The UAE Information Assurance Regulation applies to designated critical entities and provides a national control baseline. The National Cloud Security Policy adds guidance on cloud governance, data location, portability, resilience, and exit. Scope should be confirmed before treating either document as a direct requirement for a private organization.
Four assurance frameworks also appear often. ISO/IEC 27001 covers an information security management system. ISO/IEC 42001 covers an AI management system. SOC 2 is an independent service-auditor report against selected trust services criteria. The NIST AI Risk Management Framework is a voluntary framework for organizing AI risk work.
These frameworks can give a team a common control language and better evidence. Each one has a defined scope. Ask which entity, product, region, system, and time period the certificate or report covers.
Where do AWS, Azure, and Google Cloud fit?
A provider logo tells you little about the actual boundary. Region, service, model, deployment type, endpoint, and failover path provide the useful detail. The table uses the UAE as an example. The same questions apply elsewhere: does the provider operate a region in your country, which models are available there, and which deployment types can route requests somewhere else?
| Provider | UAE footprint | What to check for model inference |
|---|---|---|
| AWS |
AWS Middle East (UAE), me-central-1, has three Availability Zones.
| Amazon Bedrock offers in-region, geographic, and global inference profiles. In-region inference keeps requests inside one region. Geographic and global profiles can route them more widely. Availability varies by model. |
| Microsoft Azure | Azure lists UAE North and UAE Central. | Microsoft Foundry offers regional, data-zone, and global deployment types. Regional processing stays in the deployment region. Global processing may use any Azure region. Model and deployment availability varies by region. |
| Google Cloud | Google lists UAE network edge locations in Dubai and Fujairah. Its published Middle East cloud regions are in Doha and Dammam. | A network edge location is a point of presence, not a cloud compute region. Check the exact Vertex AI model location and endpoint. Regional endpoints can help enforce a processing boundary where the service supports them. |
Sources: AWS UAE Region, Amazon Bedrock inference options, Azure geographies, Microsoft Foundry deployment types, Google Cloud locations, and Google network edge locations.
How can inference leave the region?
Inference is the moment a trained model receives an input and generates an output. In our refund example, the customer message and order details form the input. The proposed refund decision is the output.
In our UAE example, the application server, database, and log store can all sit in the country while the inference request is processed elsewhere. Global or cross-region routing gives the provider a larger pool of model capacity. This can improve throughput, model availability, and resilience. A single-region option gives a tighter processing boundary, with possible limits on model choice, capacity, or failover.
Check every connected path: the model endpoint, content-safety service, vector database, tracing platform, cache, backup, telemetry, human support process, and disaster-recovery route. “Deployed in UAE” should be translated into a data-flow diagram that names each one.
How much sovereignty is enough?
More control usually brings a trade-off. A locally operated model can provide a tight data boundary and more independence. It also asks the organization to manage hardware, model updates, security patches, monitoring, capacity, and specialist skills. A managed global API can offer new models and higher capacity sooner. It also creates more provider dependence and a wider processing boundary.
Match the position to the workload. Public marketing copy may be comfortable with global processing. An internal policy assistant may need a private enterprise tenant, short retention, and a provider commitment that customer data is not used to train shared models. A system using personal data may need a transfer assessment and stronger region controls. A bank or healthcare workload may bring sector-specific location and approval requirements.
The three controls do not have to move together. A team might use a regional managed model, keep its own encryption keys, and store all memory and evals in an independent database. That is a valid position on the spectrum. The important work is making the position visible and defensible.
What should you ask before you buy?
Where should you start?
Most organizations already use several AI products. The useful first step is to map the real workloads against the three spectra. You can then see which data flows are well controlled, which vendor claims need evidence, and which country-specific or sector requirements may need a closer review.
This article provides general information. It is not legal advice. Confirm the rules that apply to your organization and workload with qualified legal and compliance advisers.