Data residency used to settle the sovereignty question. If the database sat in-country, the data was considered under local control. AI workloads have pulled that answer apart. A prompt can be stored in one jurisdiction and executed on GPUs in another. It can then be processed by a model owned by a company headquartered somewhere else entirely, with networks and downstream services carrying it between them. Keeping the database home is still necessary. It no longer decides what happens to the data once a model touches it. That depends on who holds the key to the model.

Nicholas Borg is Chief AI Officer at 3CX, the business communications platform serving more than 350,000 customers worldwide. Based in Dubai, he leads the company's AI product work, including its AI Voice Agents, call screening and real time transcription features. He holds NVIDIA certifications in AI infrastructure and operations and in generative AI with large language models. Borg comes to AI from years in telephony, and he writes regularly about AI infrastructure and the deals reshaping it. His view comes from watching the Gulf work through AI access in real time, and from fielding the same question from his own customers in Europe.

"We can't just talk about the data and the inference and not talk about who the owner is, who has the key to the model, and where that model is going to reside," Borg says.

Residency as a negotiated middle ground

The UAE has long taken a firm line on telecom and data. WhatsApp messaging works in the country, but WhatsApp voice and video calling remain restricted under the Telecommunications and Digital Government Regulatory Authority's (TDRA) VoIP licensing rules, which limit internet calling to licensed providers. With AI, that kind of control ran into a hard limit. Building frontier models at home takes GPUs and talent at a scale the country is still assembling, and sitting AI out was not an option. "So the answer is a middle ground," Borg explains. "The data stays here. The inference can run here. But the model still belongs to someone else."

In August 2026, OpenAI introduced inference residency in the UAE, letting eligible API, ChatGPT Enterprise, and Edu customers keep GPU execution of prompts and responses in-country. Not every model or feature is covered by that configuration.

Borg sees that as real progress, with a limit. "In-region execution is a genuine step forward," he says. "But residency is not ownership. Running a model in your country is not the same as holding the key to it."

Every handoff adds an owner

Connectivity raises the same question, and the handoff is easy to miss. Each time data moves onto another provider's network or service, another party enters the chain. Sovereignty claims rarely map every one of those hops.

Borg expects public sovereignty claims to be technically accurate and practically incomplete. "Sovereignty is going to be sold like a contract with fine print," he says. "The headline will be true. The full answer is in the architecture, so buyers need to have that conversation with the engineers, not just the sales deck."

Scarce data becomes the asset worth protecting

The stakes are rising because model builders face growing pressure on the stock of high-quality public text. Epoch AI has projected that developers could exhaust the supply of high-quality, human-generated public text sometime between 2026 and 2032, a constraint flagged in Stanford's AI Index and widely covered, including by Forbes. "The best public text has largely been used," Borg says. "If you want a model that's actually different, you need data nobody else has."

He points to Spirit Airlines' bankruptcy, where Google placed the $10 million winning bid for the airline's business data. The sale remains subject to the court process. The data reportedly covers emails, chats, operational records, and pricing models, with passenger and personal information excluded. Decades of operating records are now being valued as a potential AI asset.

That market makes provider assurances harder to take on faith. "Disclaimers are not audits," Borg says. "When a lab tells you your workload won't be used for training, the question is how you verify that. Most customers have no way to check independently."

The interaction itself has value too. "How people talk to an AI, what they ask, what they do with the answers. That's valuable data in its own right," he adds. "Customers deserve verification, not just a promise."

Bringing the model inside the boundary

Borg sees enterprises responding less through renegotiated terms than through a change in posture toward their providers. "The shift isn't coming through signatures and contracts. It's coming through attitude," he says. "'I'm paying, you're getting my data, and I can't see what happens to it.'"

The answer he sees taking shape is to take an open-weight model, post-train it on proprietary data, and run it in the same environment as the database. Some large enterprises are already moving that way, partly for control and partly to rein in inference bills that grow with every employee seat. "Then you bring the model in, you own the whole circuit, and you can honestly say nothing is leaving," Borg says.

That approach keeps the database as the durable asset while moving the key back to the enterprise. "The real answer right now is to own the model," Borg concludes. "Find a good open-weight model, and there aren't many, then post-train it on your own data. It's a big lift. It's also the only way to hold the key yourself."