Inference residency: How AI vendors keep model computation inside one country

Inference residency is a promise about where an AI model does its computing. When a company turns it on, the processors that read a prompt and write the answer sit inside a chosen country or region. It goes a step beyond data residency, which only says where stored data lives.

Storing a file at home is not the same as reading it at home

Picture a bank that keeps its paper records in a vault in Abu Dhabi. Under a plain storage promise, that is all you are guaranteed. A clerk in another country could still be handed the pages to read and answer questions about them.

Now picture an engineer at an energy company who pastes a draft contract into an AI assistant and asks for a summary. The saved chat might sit in a local data center. But the moment the model reads the contract, the text has to be processed by a chip somewhere. Inference residency is the promise about that chip.

Three terms are enough to follow the rest:

• Data residency covers content stored at rest, such as saved chats and uploaded files.
• Inference is the moment a trained model processes a prompt and generates a reply.
• Inference residency says that step happens on hardware in the chosen region too.

What actually happens when a prompt is processed

Running a large language model takes specialized chips called GPUs (graphics processing units). A prompt travels to a data center, a GPU runs the model on it, and the answer travels back. The whole trip can take a few seconds.

Without a residency promise, a vendor may send that work to whichever data center has spare capacity. That can be in another country, and the customer may never see where. This is a normal way to run a global service efficiently, but it is exactly what some buyers cannot accept.

With inference residency, the customer’s workspace is pinned to inference engines in one region. Prompts and responses are handled by GPUs there. OpenAI describes it this way for the UAE: eligible customers can restrict GPU inference to UAE-based engines, extending its UAE data residency offer for data at rest.

Why regulated organizations ask for it

Some organizations are barred or discouraged from sending sensitive material abroad. Energy companies, government agencies and universities often fall into this group. For them, storing data locally is not enough if prompts containing that data are processed elsewhere, even briefly.

The reasoning is simple. A prompt often contains the most sensitive material of all, because that is what people ask the model to work on: contracts, permit files, investment notes, student records. If the computation leaves the country, so does the content, at least for the moment it is being read.

Our report on ADNOC, G42 and OpenAI shows the idea in practice. According to the Abu Dhabi Media Office, ADNOC’s deployment runs on OpenAI’s GPT models using in-country inference infrastructure in the UAE, and a government pilot uses OpenAI’s UAE inference residency. The announcements did not publish deployment timelines or financial terms.

How it is offered today

OpenAI is the example with public documentation. It offers inference residency to eligible API, ChatGPT Enterprise and ChatGPT Edu customers. The UAE is one of three locations with the option, alongside the United States and Europe, according to OpenAI’s UAE announcement. The company has said it plans to add regions over time.

There is a hard prerequisite: data residency must already be enabled in the same region. In other words, inference residency is an add-on layer, not a replacement. A customer first fixes where data is stored, then fixes where it is computed on.

Press coverage of the UAE launch says data residency is included at no extra cost on Enterprise and Education plans. Pricing for inference residency was not confirmed in the sources reviewed. Other cloud vendors offer similar regional controls under their own names, so the term is not unique to one company.

The fine print: What can still leave the country

The guarantee is narrower than it sounds. OpenAI’s help documentation, as summarized by news coverage, says it covers GPU execution of in-scope customer content for supported workloads. It does not confine every step. Non-GPU work such as authentication and analytics can still happen outside the region. External web search, if used, is not kept in-region either.

Model choice is another limit. One report on the UAE launch says only GPT-5.2 was available for UAE inference residency, that the GPT-5.6 family was not supported, and that features such as image generation, internal search and improved memory were unavailable. Outlets differ on details, so buyers should check the vendor’s current documentation.

This matters because a regional deployment can trail the main service. A team that pins its workspace to one country may get a narrower product than a team that does not. That is a trade-off, and for regulated buyers it is often one they accept on purpose.

What to watch

• Model parity: whether regional deployments get new models at the same time as the main service, or lag behind.
• Feature gaps: which tools stay unavailable inside a pinned region.
• More regions: OpenAI says expansion is planned, and rivals face the same demand from regulated buyers.
• Audit evidence: how customers can verify where processing took place, beyond the vendor’s own statement.

For buyers, the practical step is to ask which parts of a request stay local. Storage, GPU compute, logging, search and support access are separate questions with separate answers. A vendor that can answer each one in writing offers a much firmer guarantee than one that says only that data stays in the country.