The Gulf sovereignty paradox
Gulf AI strategies are an effort to manage different types of dependencies. At the hardware layer, access to state-of-the-art computing power remains shaped by US export controls and licensing decisions, making American approval a continuing condition of many Gulf projects. At the cloud and data layer, local data centres can provide residency, but not necessarily insulation from the jurisdiction of the foreign companies that operate them. At the model layer, Chinese open-source systems may offer lower-cost and more adaptable alternatives to proprietary American tools, but they can create switching costs once embedded in public services or trained on local data. These costs, including the expense of retraining models on alternative architectures and the political cost of unwinding ties with a strategic partner, are not unique to Chinese systems. But the consequences of dependency on Beijing versus, for instance, Washington carry a different strategic and political weight for Gulf states. European providers, including France’s Mistral, meanwhile, offer another route for hedging, though not yet a complete alternative to the American stack.
The Gulf states are already mixing these layers: the United Arab Emirates (UAE) has accepted US conditions on the technology relationships of its principal AI company, G42, while also investing in Mistral and experimenting with Chinese Alibaba-derived model architectures; Saudi Arabia has preserved Chinese technology links while partnering with American firms; and Qatar has emphasised Arabic-language AI. These strategies increase optionality, but they do not amount to technological independence.
The scale of capital is extraordinary: sovereign investors globally deployed US$66 billion in AI and digital infrastructure in 2025, with Gulf funds the largest contributors – including US$12.9bn from the UAE’s Mubadala Investment Company, US$6bn from the Kuwait Investment Authority and US$4bn from the Qatar Investment Authority. The seven largest Gulf funds together accounted for 43% of all sovereign capital deployed globally that year across all sectors, a historic high; this includes completed transactions, capital commitments and announced partnerships. The UAE has committed to what may become the largest AI infrastructure deployment outside the US. The US–UAE AI Acceleration Partnership, established during Trump’s visit to Abu Dhabi in mid-2025, provides the framework for a 5 GW AI campus. G42 partnered with OpenAI, Oracle, Nvidia, Cisco and SoftBank to build Stargate UAE, a 1 GW facility within the campus. The first 200-megawatt cluster is on track for delivery in 2026.
G42’s trajectory illustrates the Gulf’s sovereignty paradox precisely. Microsoft’s US$1.5bn investment in the company in 2024 for a minority stake and board seat came with conditions that included G42’s divestment of partnerships with Huawei and other Chinese technology firms. The deal was brokered with assurances to both the US and Emirati governments regarding security compliance. G42 secured access to advanced Nvidia chips and OpenAI’s models, but only after accepting conditions that made American permission a precondition for building ostensibly sovereign infrastructure.
Yet hedging is already visible. Abu Dhabi’s MGX invested in both the American stack (through Stargate UAE) and the European alternative (through Mistral), with the UAE committing US$35.4–59bn to French data-centre expansion as part of a bilateral UAE–France Framework for Cooperation in Artificial Intelligence. Meanwhile, the Mohamed bin Zayed University of Artificial Intelligence launched K2 Think, a reasoning model built on Alibaba’s open-source Qwen architecture (a Chinese model) in partnership with G42 and America’s Cerebras (non-Chinese hardware), with no evident contractual barrier.
Saudi Arabia’s approach differs structurally. HUMAIN, backed by the Public Investment Fund (PIF), partners with Qualcomm, Google Cloud and Nvidia on a deal-by-deal basis. The PIF allocated approximately US$40bn to AI investments across geographic lines. Critically, Saudi Arabia has not replicated G42’s clean break with Chinese technology: Huawei’s role in the country’s 5G infrastructure through STC remains intact, preserving a channel for Chinese AI integration that Abu Dhabi foreclosed. Washington has pressed Riyadh on Huawei’s network role since 2019, and US approval of recent Saudi AI-chip exports has come with security and reporting requirements. Saudi Arabia, however, has so far avoided the kind of disentanglement condition Microsoft applied to G42. This divergence raises questions about whether Washington is willing to apply equal leverage across the Gulf, or whether the size of the Saudi relationship makes such conditions politically harder to impose.
Qatar, meanwhile, is pursuing a linguistically focused strategy. The Fanar Project, developed by the Qatar Computing Research Institute with government support, builds Arabic-language AI models across multiple specialisations. The rationale extends beyond language: AI models trained on American or Chinese data embed cultural assumptions into their outputs, generating skewed representations of social institutions when deployed across societies with distinct cultural frameworks. For Gulf states integrating AI into education and public services, linguistic sovereignty is inseparable from cultural sovereignty. The Qatar Investment Authority has formed Qai, a national AI company, alongside a US$20bn joint venture with Brookfield for data-centre infrastructure.
The competitive dynamic between the Gulf states compounds the paradox. Rather than pooling computational capacity or coordinating procurement to strengthen bargaining power – as they have periodically done through the Organization of the Petroleum Exporting Countries – the UAE and Saudi Arabia are building parallel infrastructure in competition, fragmenting regional leverage against technology providers.
The Iran war added a physical dimension to the sovereignty question. The Gulf Cooperation Council data-centre market, projected to reach US$9.5bn by 2030, is concentrated in a region where energy infrastructure has been targeted repeatedly – from the strikes on Aramco’s Abqaiq oil facilities in 2019, which knocked out 5.7 million barrels per day of Saudi crude, to the severing of four undersea cables in the Red Sea in 2024, which disrupted a quarter of Asia–Europe internet traffic. The Gulf AI stack depends on power grids, fibre-optic corridors and cooling water, and exposure runs upstream of the Gulf’s infrastructure as well. More than 70% of the world’s advanced semiconductor chips are fabricated in Taiwan, which draws over 40% of its electricity from gas – much of its liquefied natural gas (LNG) is imported from Qatar. Meanwhile, South Korea, the source of roughly two-thirds of global memory chips, imports some 70% of its crude oil and a fifth of its LNG from the Middle East. Chip fabrication also rests on feedstocks concentrated in the same region: Qatar supplies over a third of the world’s helium, which is critical to the process, and Israel and Jordan together account for roughly two-thirds of global bromine, which is used in etching circuit patterns into wafers. Current and possible future Hormuz closures therefore strike the stack in two places: directly, at Gulf data-centre capacity, and upstream, at the energy sources and feedstocks on which chip fabrication depends.
Experts had therefore highlighted the need for data centres processing sovereign government data to be integrated into national-defence architectures. In the face of retaliatory strikes by Iran that began in February 2026, Gulf air-defence systems performed effectively – the UAE intercepted thousands of attacks from ballistic missiles or drones. The Aramco precedent cuts both ways, however: the September 2019 strikes on the Abqaiq processing plant briefly halted nearly half of Saudi production, yet full output was restored within weeks through spare capacity, redundant processing trains and rapid component replacement. Thus, while concentrated infrastructure can be struck devastatingly, redundancy and operational depth can reverse the damage.
The inference question
The global AI sovereignty debate has thus far largely focused on the physical location of data. A less examined dimension concerns inference data – the queries submitted to AI systems and the outputs they generate. When a national oil company runs trading optimisation through a foreign-operated AI cloud, the queries reveal production forecasts, hedging strategies and counterparty analysis. When a government ministry uses AI-assisted scenario planning, the questions disclose which contingencies the state is preparing for. The risk is twofold. Under the CLOUD Act, inference logs held by US-headquartered operators remain accessible to the American legal process regardless of server location. Independently, the queries themselves constitute a continuous intelligence stream – what a state is preparing for, hedging against and modelling – generated in close to real time. Training-data residency does not address either exposure.
France’s military framework agreement with Mistral, for instance, addresses this implicitly by requiring all AI operations to run on nationally controlled infrastructure. The Iran conflict made the risk concrete: the US military used Anthropic’s Claude model for intelligence assessments during the strikes, and Iran stated that it targeted cloud-computing facilities allegedly for ‘their role in supporting military operations’. Commercial cloud infrastructure in the Gulf was thus simultaneously serving civilian, corporate and military inference – and an adversary targeted it for the third of those reasons. No Gulf state has publicly identified inference-data exposure as a sovereignty risk, though a recent Aramco–Microsoft memorandum of understanding commits to ‘sovereign-ready digital infrastructure’ – suggesting a growing awareness that operational AI generates strategically sensitive information beyond training data.
|
