Building AI in the UAE: where should your models actually run?

What's changed on the ground

• In-country cloud regions. Hyperscalers operate in the UAE. AWS, for example, opened its Middle East (UAE) region in 2022 with three Availability Zones⁴⁵. Check each provider's current service list, since AI services don't always launch everywhere at once.
• Large-scale AI compute. Stargate UAE, a 1GW AI cluster in Abu Dhabi announced in May 2025 and developed by G42's Khazna with partners including OpenAI, Oracle and Nvidia, is under construction. Reports put the first 200MW phase on track for the third quarter of 2026, while timelines for the rest are not yet confirmed¹²³.
• Regulation still settling. The federal PDPL applies, the free zones run their own regimes, and DIFC already has rules aimed at AI systems, as I set out in UAE PDPL and AI in 2026.

A placement framework by data sensitivity

The gateway does the heavy lifting

Everything above depends on one control point. Every model call goes through an AI gateway that authenticates the caller, classifies or tags the data, redacts what shouldn't leave, routes to the right deployment and logs the decision. I describe the pattern in more detail in my notes on private LLMs, and the agent side in securing MCP servers.

With that in place, adding capacity from new UAE compute becomes a routing change rather than a rebuild. Hosting is configuration, not architecture.

Private models: smaller than you think

For the most sensitive tier, teams often assume they need frontier-scale models. In practice, many enterprise tasks such as extraction, classification, summarisation and grounded question answering work well with mid-sized open-weight models, especially when retrieval is good. My own experiments on an NVIDIA DGX Spark are about exactly this question: what can run privately, and how well. Measure on your own evaluation set before buying hardware or committing to a provider; the evaluation checklist covers how.

What I'd do this quarter

1. Inventory AI use cases and classify each by data sensitivity and the regime that applies.
2. Stand up a gateway with routing, redaction and audit before adding more use cases.
3. Pick an in-country cloud region for personal-data workloads and confirm the AI services you need are available there.
4. Pilot one private open-weight deployment for your most sensitive use case, and measure it against a hosted baseline.
5. Record the placement decisions as architecture decision records, so they can be revisited as the rules and the compute evolve.

How I can help

I help UAE and GCC teams design AI platforms that put each workload in the right place, with the governance to prove it. See fractional CTO and AI advisory in the UAE, or analyse your challenge and bring the result to a call.