Shortlisting a model, we expected capability to be the hard filter. It wasn't. The hard filter was proving a model was allowed to touch the data at all. Residency did the work that benchmarks usually get credit for.

The UAE's Federal Authority for Artificial Intelligence and Data, established 14 June 2026, is meant to settle this nationally. Much of the detail was still being written when we needed it. So we built against the parts that existed and made defensible calls on the rest.

Residency comes before capability, not after

The mandate's technology track exists separately from operations, services and training, and its job is governance, deployment and security standards for the AI layer itself. One recent industry analysis of the UAE's agentic AI rollout put it plainly: data residency is now shaping architecture and supplier choices, not trailing behind them as a compliance afterthought.

That matched what we saw. Residency was the first filter, applied before capability entered the conversation, and it removed contenders that would otherwise have led the field.

Who we actually tested

We ran a formal RFI, not informal conversations. The same nine-category, weighted scorecard went to three vendors: Core42 and Du, two competing sovereign AI cloud platforms, and Emircom, a systems integrator route rather than a platform.

The nine categories, in weight order:

• Sovereignty and data residency (heaviest weight, by design)
• Which models are exposed as a managed service, and where they're hosted
• Document AI / OCR trained on UAE-specific documents
• Pay-as-you-go pricing
• Dedicated GPU pricing, for a self-hosted option
• Provisioning and connectivity into the entity's network
• Agent framework support
• Observability and security certifications
• Commercial terms and federal references

We're not naming which vendor we built on, consistent with how this series treats the entity's technology choices. What's worth sharing is the shape of the test itself: a platform vendor and a systems integrator answer a different subset of those nine categories well. Knowing that upfront meant we asked each one only the questions it could actually answer.

Storage and inference are not the same question

This is the distinction that cost us the most time, and the one we see other teams miss. Where data sits at rest and where a model actually reasons over it are two different questions. A vendor can store your data inside the UAE and still send the live inference call somewhere else entirely, which still gets marketed as a “compliant, UAE-hosted service,” and still fails the mandate the moment a regulator asks where the thinking happened.

Getting a straight answer on inference location took longer than any benchmark we ran. Vendors were specific about storage and vague about inference. The vagueness was itself the answer.

What we ruled out

Three patterns, in order of how fast they eliminated a candidate:

• No in-country inference option at all — the strongest benchmark scores in our shortlist, gone on day one.
• Stored locally, reasoned elsewhere — passed the marketing pitch, failed the direct question.
• Right infrastructure, no audit trail on request — mattered to the regulator as much as server location did.

What survived was a shorter list than expected. Once every candidate met the residency and audit bar, the remaining call looked like an ordinary technology decision: Arabic/English handling, task fit, and whether the vendor's roadmap kept pace with a calendar that doesn't grant extensions.

Why we didn't buy GPUs on day one

“Agentic AI is expensive” gets repeated often enough that it shapes budgets before anyone checks the number. It's not wrong, it's incomplete. Dedicated infrastructure sized for volume you don't have yet is expensive. Paying per token for what you actually use is not.

Sovereign cloud vendors in the UAE sell two ways:

• Model-as-a-service — pay per token, in-country, nothing to provision.
• GPU-as-a-service — rent a dedicated chip, in-country, pay for it whether it's busy or idle.

Core42's Compass platform publishes its pay-as-you-go rates openly. Its frontier reasoning model runs a little over a dollar per million input tokens and around ten dollars per million output tokens. A lighter model for routine extraction costs a fraction of that. Document AI processes text at roughly a dollar per thousand pages. A dedicated H100, by comparison, runs about $2.50 an hour on demand — call it $22,000 a year running around the clock, doing work or not.

Run the two curves against each other and the crossover lands far out: on Core42's published rates, a dedicated GPU only earns back its cost north of half a million transactions a year on a typical agentic workload. A ninety-day pilot doesn't come close. Neither does a busy service running flat out for a full year.

So we ran pay-as-you-go through the pilot and into year one, with the model kept behind a thin abstraction so it can be swapped without reworking the agents around it. We also route: simple, high-confidence cases go to the cheaper model, only the hard cases escalate to the expensive one. That alone cut model cost by more than half against sending everything through the frontier model by default. Dedicated capacity earns its place once volume nears the break-even line, or earlier if policy demands an air-gapped environment, a sovereignty call, not a cost one.

At pilot volumes the model is one of the smallest line items in the build. The real cost is the engineering around it: runtime, guardrails, audit logging, integration. That doesn't change whether you pay per token or rent a GPU. The infrastructure choice doesn't move the real cost. It only moves the risk, from underpaying for a spike you didn't plan for, to overpaying for capacity you're not using yet.

What's coming next

Two posts left: the governance layer (classification, consent, the audit trail), and a plain account of what changes for a citizen and an officer once a service like this goes live.

Frequently asked questions

If you're inside a federal entity working through your own shortlist, we'd like to compare notes. Follow along for the next post, on the governance layer that makes a service like this safe to put in front of a regulator.

---

Syed M. Umair is Head of Product and Digital ventures at Microvision, where he is currently working with a federal entity on its path through the UAE's assistive AI mandate. This is the third in a six part series on what that work actually involves.

Connect with Syed on LinkedIn