The Great Unbundling of the Hyperscalers
There is a quiet reordering underway in how large organisations think about where their artificial intelligence runs, and it has moved well beyond the abstract. Two years ago, the conversation between infrastructure vendors and enterprise buyers turned almost entirely on efficiency and cost optimisation, the familiar grammar of a cloud-first decade. That grammar has changed. Rabii Al Khonaizi, Regional Sales Director for KSA and North MEA at Submer, described a market in which the questions have moved up the stack.
“Today, enterprises are asking how to run dense AI and GPU workloads on-premises, with sovereignty and control becoming central priorities, especially in Saudi Arabia and across MEA where data residency expectations are rising,” he said. He attributed the shift to the rapid growth in AI compute demand and a deliberate move away from unpredictable cloud operating expenditure toward more stable investment models, adding that the conversations now begin earlier, “often around defining the infrastructure strategy rather than selecting a solution.”
That observation is not confined to one vendor’s order book. It describes a structural movement that independent research has now quantified with some force.
The retreat from cloud-first is no longer anecdotal, it is measured
The scale of the repositioning has become difficult to dismiss as sentiment. A 2026 enterprise survey of 203 IT decision-makers commissioned by Cloudian found that 93% of organisations have already repatriated some AI workloads from the public cloud, are in the process of doing so, or are actively evaluating such a move, with 79% having already shifted workloads and 73% planning to move further toward on-premises or hybrid infrastructure over the following two years. The same study identified three converging forces behind the trend: data sovereignty concerns, cloud cost unpredictability, and real-time performance requirements, a triad that maps almost exactly onto what Al Khonaizi reported hearing in his own regional conversations.
The financial dimension of that retreat deserves particular attention because it is the point at which buyers most often discover the gap between what they were promised and what they paid. Cloudian’s research found that 40% of enterprises reported actual cloud AI spending exceeding their initial projections, while nearly half cited cloud-specific cost unpredictability as a barrier to expanding AI adoption at all. The Barclays CIO Survey, conducted at the end of 2024, recorded that 86% of chief information officers planned to move at least some public cloud workloads back to private cloud or on-premises environments, the highest figure ever registered in that survey. Al Khonaizi located the sharpest version of this pushback in the region.
“The biggest pushback in MEA is around cost predictability in hyperscaler-based AI workloads, particularly training, where actual usage often exceeds forecasted spend,” he explained. That pattern, he said, has driven customers to reassess how they balance cost, performance, and sovereignty, with the discussions often beginning “with building a clearer financial and strategic case for infrastructure independence before moving into any technical design.”
Sovereignty has turned a financial argument into a national one
What distinguishes the Gulf version of this story from its European or American counterparts is the degree to which the commercial case has fused with a national one. Al Khonaizi was direct on this point. “In Saudi Arabia, this is amplified by national priorities around digital sovereignty under Vision 2030, leading to a stronger preference for owned and controlled infrastructure,” he said. The Kingdom has matched that rhetoric with capital at a scale few jurisdictions can rival.
The launch of HUMAIN under the Public Investment Fund in May 2025, chaired by Crown Prince Mohammed bin Salman, established a vehicle mandated to build the entire AI stack from data centres and cloud infrastructure through to sovereign Arabic models, and the country’s data centre sector has expanded sixfold since Vision 2030 began, with investment exceeding $4.26 Billion and more than 60 operational facilities developed by over 20 companies according to the Saudi Ministry of Communications and Information Technology.
This is the context in which enterprise repatriation in the region acquires a momentum it lacks elsewhere. When the national infrastructure strategy targets gigawatts of sovereign compute capacity and the Personal Data Protection Law stimulates local hosting, an individual enterprise’s decision to bring workloads home stops looking like a contrarian bet and starts looking like alignment with the direction of national policy. Al Khonaizi described a parallel shift in how technology leaders themselves now value infrastructure.
“CIOs are increasingly reframing the conversation from simply running workloads to asking how infrastructure can deliver strategic and financial value over time,” he said, noting that this includes performance, scalability, and, in some cases, monetisation potential, with the discussion moving toward “defining the right architecture and economic model first, ensuring infrastructure decisions are aligned with broader business and national AI ambitions.”
The roadmap is being rewritten around a question buyers cannot yet answer
If the demand signal is clear, the shape of the response is more revealing still. Al Khonaizi said the strongest message from customers carried a clear instruction about what they wanted from a vendor. “The strongest signal from customers is that they don’t want a catalogue of solutions; they want help clearly defining what they are actually trying to solve,” he explained, describing a signal that is shaping Submer’s 2026 roadmap toward more structured, advisory-led engagement in the earliest stages of infrastructure planning. The intent, he added, was to align architecture decisions with each customer’s business model constraints and growth trajectory, because “given the speed of AI evolution, infrastructure now needs to be designed with flexibility and rapid iteration in mind rather than fixed long-term assumptions.”
The reason that the advisory layer matters becomes apparent only when one examines what these buyers are not asking about, and here the analysis arrives at its central tension. The entire repatriation thesis rests on the assumption that enterprises can run dense AI workloads on their own infrastructure as effectively as a hyperscaler can. Whether that assumption holds depends almost entirely on a part of the stack that most buyers, by Al Khonaizi’s account, still treat as an afterthought.
The blind spot Al Khonaizi identified is specific and consequential. “Many customers are still not fully considering the role of power and cooling infrastructure in AI performance, especially at high GPU density,” he said, explaining that the focus tends to fall on compute costs while the broader facility efficiency layer is underestimated despite its significant impact on total cost of ownership. As GPU scale increases, he added, these elements become critical to both cost and performance, which is why the company had taken on a particular responsibility in the early conversation. “Our role is to bring this into the conversation early, helping customers evaluate the full stack, not just compute, but the systems that enable it,” he said.
The external data emphatically validates that warning. The International Energy Agency’s 2025 Energy and AI report found that global data centre electricity consumption reached 415 terawatt-hours in 2024, roughly 1.5% of total global electricity use, and projected that figure to more than double to around 945 terawatt-hours by 2030, slightly more than Japan’s entire annual electricity consumption today, with AI as the most important driver.
The IEA’s subsequent analysis recorded that electricity demand from data centres rose 17% in 2025 alone, with AI-focused facilities climbing faster still, well outpacing the 3% growth in global electricity demand overall. Crucially for any enterprise contemplating its own build, the IEA noted that the share of cooling systems in total data centre consumption ranges from about 7% in efficient hyperscale facilities to over 30% in less efficient enterprise data centres, a spread that quantifies precisely the penalty an unprepared operator pays for underestimating the physical layer.
That figure is the quiet heart of the whole movement. An enterprise repatriating workloads to reclaim cost predictability and sovereignty can surrender both if its facility runs cooling at four times the overhead of the hyperscaler it left. The desert climate of the Gulf sharpens that risk rather than softening it, which is why market analysis of the Saudi data centre sector has flagged water-frugal and hybrid cooling design as a genuine competitive differentiator rather than a technical footnote.
The unbundling of the hyperscalers, in other words, does not free enterprises from the hyperscalers’ hardest problem. It hands that problem to them. The organisations that recognise this early, and treat power and cooling as a first-order architectural decision rather than a procurement detail settled after the GPUs are chosen, are the ones for whom the economics of independence will actually hold. Al Khonaizi’s wager and Submer’s roadmap is that the rest will learn it the expensive way.
