Summary: In Azure UAE North, the Azure OpenAI GPT models run only under provisioned throughput, so a small sovereign AI assistant costs a fixed monthly reservation, not a token bill. This is what the region tables allow, why the reservation came to 69% of our running cost, and how to quote it.
The spreadsheet that was wrong by a factor of a hundred
The first cost model we built for a 100-user AI assistant, hosted in the UAE, looked tidy. Fifty concurrent users, a few thousand questions a month, tokens priced at the public pay-as-you-go rate. The model line came out so small that it hardly deserved a row.
It was wrong by two orders of magnitude. Nobody caught it until the subcontractor’s architecture note said, in one flat sentence, what we had missed, and we checked the Microsoft region tables ourselves on 23 September.
What Azure OpenAI in UAE North actually serves
Microsoft publishes a table of which models run in which region, and under which deployment type. Read the uaenorth row slowly. We wrote about why the deployment type, not the region name, decides data residency earlier. Pay-as-you-go deployments there serve embeddings and speech-to-text. The GPT models (GPT-4.1, GPT-4o, GPT-5-mini, GPT-5.1) run only under Regional Provisioned Throughput.
Provisioned throughput means you reserve capacity in units and pay for the reservation whether anyone asks a question or not. The smallest deployment is 25 units for the small model and 50 for the larger one. There is no 3 a.m. discount for an empty office.
The global and data-zone deployment types do offer pay-as-you-go GPT, but they may process the request outside the country. If the tender says in-country processing, those are out. The table is public. Most quotes we see were written without opening it.
Why the reservation swallows everything else
Once we rebuilt the model, the reservation came to 69% of the monthly running cost. The application tier, the database, the API gateway and the logs, added together, were a small slice of what was left. The tokens themselves were a rounding error on top of a reservation that had already been paid for.
That changes what you are selling. A buyer who hears “AI assistant” assumes a usage bill: more questions, higher cost, and a cheap pilot that grows. In UAE North a small assistant and a busy one cost nearly the same, because the floor sets the price and usage barely touches it. A 100-user assistant is an infrastructure line, closer to renting a rack than to paying a telephone bill.
I find this mildly annoying, to be fair. Everyone wants the usage story. It is simply not what the region tables allow.
How to quote it
Quote a fixed monthly fee, sized on the reservation. Hold it for twelve months. Put one sentence in the contract that ties the fee to the Azure list price on a stated date, and say the fee will be restated before signature if that price moves. Reservation prices for the region were not confirmable from the public list when we bid, so the clause is not decoration.
Two more details we wrote into our own pricing:
- Buy the annual reservation at signature. Hourly provisioned pricing is several times the monthly rate and would sink any fixed fee.
- Carry the larger model as a priced option. If the evaluation shows the small model cannot handle Arabic well enough, the minimum doubles. Better that the customer sees that number on day one than that we absorb it in month three.
A review clause that starts in month thirteen protects year two only. The risk sits in year one.
What to check before you buy or bid
Ask whoever quotes you for the deployment type behind every GPT line, not the model name. “GPT-4o, in-country” can mean three different things on the invoice.
Ask for the reservation size in units and the term, and whether it is hourly, monthly or annual.
Ask what happens to the fee when the list price moves. If the answer is “we absorb it”, ask who signed that.
Check the alternatives, because they have moved. Since August, OpenAI’s own API offers UAE processing for several of its newer models, priced per token with a ten per cent uplift and no reserved block. It comes with its own gates: an abuse-monitoring approval, a retention amendment and a separate UAE approval. And several features, such as files, batches and fine-tuning, are storage-only in the country, so processing may still leave it. For some buyers that is the better route. For others the sovereignty clause rules it out. Either way the answer comes from reading the tables, not from the brochure.
If the buyer refuses any reserved block at all, the in-region options shrink to open-weight models on GPU machines, which are billed by the hour and fixed all the same. There is no pay-as-you-go sovereign route for GPT-class models in UAE North today.
Frequently asked questions
Can I run GPT models pay-as-you-go in Azure UAE North?
Not in-region. Standard deployments in uaenorth serve embeddings and speech-to-text. The GPT models run only under Regional Provisioned Throughput. Global and data-zone deployments are pay-as-you-go, but they may process requests outside the UAE.
What is the smallest provisioned throughput reservation?
25 units for the smaller GPT model and 50 for the larger one. You pay for the reservation whether anyone uses it or not.
Is there a pay-as-you-go route for sovereign AI in the UAE?
OpenAI’s own API offers UAE processing for several newer models, priced per token with a ten per cent uplift, subject to approvals. Some features are storage-only in the country, so processing may still leave it. Open-weight models on in-region GPU machines are billed by the hour.
How should a sovereign AI assistant be priced?
As a fixed monthly fee sized on the reservation, held for twelve months, with one contract sentence that ties the fee to the Azure list price on a stated date.
Next step
If you have an in-country assistant on the roadmap and a quote that looks suspiciously cheap, send us the line items. We will tell you which row the reservation is hiding in. See how we build self-hosted enterprise AI, or Get in touch.
Photo by İsmail Enes Ayhan on Unsplash
