Sovereignty Is No Longer a Compliance Problem. It’s a Control Problem.

Second in a four-part series unpacking the pillars of CASA26.

A year ago, the story of AI in the Middle East was a story of pure ascent. The Gulf had decided it would not merely consume artificial intelligence; it would build the ground it runs on. Abu Dhabi broke ground on a gigawatt-scale compute cluster. Saudi Arabia’s sovereign wealth stood up a national AI champion and raised its ambitions into the trillions. Washington cleared the export of tens of thousands of advanced GPUs that had been frozen for months. With more than three trillion dollars of sovereign capital behind them and regional data-centre capacity set to triple by the end of the decade, the Gulf states could build while everyone else waited for financing.

Then, in late February, the other face of the bet arrived. As the conflict with Iran escalated, drone and missile strikes hit data-centre infrastructure across the Gulf — and AWS confirmed its UAE region had been knocked offline, with recovery measured in months and customers told to move their workloads elsewhere. The region that had marketed itself as a safe harbour for the world’s data watched a flagship cloud region burn. In the space of twelve months, sovereignty went from a slide about data residency to a question with missiles attached: who physically controls the ground intelligence runs on — and what happens when that ground is contested?

That is the real subject of CASA26’s second pillar — AI, sovereign infrastructure and agentic systems. It’s also why e& enterprise’s Ahmed Omer is returning to CASA26 to talk through sovereignty from a Middle East vantage point that looks nothing like it did a year ago.

The hard part is no longer building the model

For most of the AI cycle, the centre of gravity sat in a handful of hyperscaler data centres where the largest models were trained. That era is closing. Models are commoditising — capable, open-weight, increasingly interchangeable. The defensible value is moving to a different question.

The hard part is no longer building the model. The hard part is deciding where it runs.

Training is a one-time, centralised act. Inference — the actual work of intelligence, performed millions of times a day — is not. It wants to be close to the data, close to the user, close to the decision. That pulls it out of the hyperscaler core and toward the edge: into networks, devices, enterprise premises and regional infrastructure. The economics point the same way. The binding constraint on AI is no longer chips alone; it is power, latency and the cost of moving data around. Inference at the edge is often cheaper, faster and easier to govern than inference shipped to a distant cloud and back.

This is the opening the telecom industry has been waiting for, whether or not it realises it.

Why this is the operators’ second chance

Hyperscalers own scale. What they cannot easily own is proximity, regulatory standing and trust — the three things that decide where regulated intelligence is allowed to run. Operators have all three almost by default: physical infrastructure close to the user, deep compliance relationships with national regulators, and decades of being trusted with sensitive data.

Sovereign AI is usually framed as a burden — GDPR, the EU AI Act, data-residency rules, the long list of reasons a regulated enterprise cannot simply pipe its data into a centralised model. Framed that way, it’s a cost. Framed correctly, it’s a market. The very rules that make centralised LLMs unusable for a bank, a hospital or a government create demand for intelligence that runs on controlled, in-country, compliant infrastructure. Someone has to host that. The operator with edge sites, a power footprint and a regulator’s trust is better placed to do it than almost anyone.

We saw the early shape of this at CASA25. Deutsche Telekom set out a blueprint for a sovereign CPaaS built on European terms. KPN made the case that reinvention is as much about AI culture and capability inside the operator as about the technology itself. Intel joined the uncomfortable conversation about why telcos keep failing at innovation — which is, at heart, this exact gap between holding the right assets and actually moving on them. And BT’s work on UC Edge points at the same instinct: push capability out to where the customer and the data already are.

The assets are real. The question CASA26 has to answer is whether the industry can convert them into a commercial position before the window closes.

Agents make the question urgent

There is a third force tightening all of this: agentic AI. Once intelligence is distributed across networks, devices and enterprises, the next step is agents that don’t just answer but act — booking, paying, negotiating, resolving, on a customer’s or a company’s behalf, with a degree of autonomy.

That changes the stakes of “where it runs” entirely. An agent acting autonomously inside a regulated business is not a chatbot; it’s an actor with access, authority and consequences. The control plane — the layer that governs which agents may operate, on what data, under whose rules, with what audit trail — becomes the most valuable real estate in the stack. It is, in effect, the new orchestration layer, and it sits naturally next to the infrastructure that hosts the inference.

CASA has to be honest here, because the industry often isn’t. Much of what is sold today as “agentic AI” is still a demo. At CASA25, the agentic sessions — including the “beyond the buzz” discussion Kyle Nel led, with Intel among those at the table — kept returning to the unglamorous questions: consent, identity, trust, governance, and whether any of it survives an enterprise procurement process. The capability is racing ahead. The governance, the economics and the trust model are not. That gap is the opportunity, not the disappointment.

What CASA26 will do with this pillar

The AI and sovereignty track at CASA26 is built around the questions that decide who captures this — not the ones that demo well:

• Where will inference actually run in five years, and who pays for the power, the edge sites and the compliance?
• Is sovereign AI a genuine commercial line for operators and regional infrastructure players, or a slogan they’ll cede to the hyperscalers anyway?
• What does the control plane for agentic AI look like in practice, and who is positioned to own it?
• And from the Middle East to Europe, how does a region turn geopolitical exposure into a reason to keep intelligence on home soil?

These are analyst-led discussions, in a room of around 150 senior leaders who build, regulate, finance and deploy this infrastructure — operators, chipmakers, investors and the regional voices, including e& enterprise, who are living the sovereignty question in real time.

This is also one of the CASA26 tracks open for a partner to help shape and lead. If your company’s future depends on where intelligence runs — and on being seen as one of the organisations defining the answer — this is the conversation to be inside, while the track is still open.

The cloud was built for humans. The next infrastructure is being built for intelligence — and the question of whose soil it stands on is no longer academic.

Amsterdam. September 21–23. The next chapter starts now.

Next in the series: Network APIs & Telecom Transformation — from API counts to programmable networks, identity, and the business model operators keep missing.