The Gulf's Hyperscaler Bet: Who Holds the Leverage When the Weather Turns?

Sometime in the third quarter of this year, the first 200 megawatts of Stargate UAE will hum to life on the edge of Abu Dhabi. More than 100,000 cubic metres of concrete already sit in the ground — foundations for the first gigawatt of what is planned as a five-gigawatt UAE–US AI campus, the largest of its kind outside America. Down the road in Riyadh and Dammam, Humain's first data halls are lighting up with Nvidia GB300s. The cranes are real, the chips are arriving, and the press releases have never been glossier.

So here is the question I keep being asked by strategy teams in the region, and the one this piece tries to answer honestly: are the Gulf states betting their AI future on the hyperscalers — Microsoft, OpenAI, Nvidia and friends — or on sovereign alternatives like G42, Humain and the Falcon model family? And when geopolitics shifts, as it always does, which of those bets actually holds?

My answer, up front: it's both, it's deliberately messy, and the messiness is the strategy. But not all parts of the bet are equally durable — and the part everyone celebrates loudest is the part most exposed.

The shopping list, totalled

Start with what is actually signed, because the headline numbers blur together after a while.

On the UAE side, Microsoft's commitment now stands at $15.2 billion through 2029 — $7.3 billion already spent since 2023, including the $1.5 billion equity stake in G42 that made Redmond a shareholder in Abu Dhabi's sovereign AI champion, plus another $7.9 billion planned for 2026–2029. The latest tranche adds 200MW of data centre capacity through Khazna, G42's infrastructure arm, coming online before the end of this year. Stargate UAE layers OpenAI, Oracle, Nvidia, SoftBank and Cisco on top of G42's land and power, with G42 cleared to buy the equivalent of up to 35,000 Blackwell-class chips — a jump from the "couple of thousand" it could access two years ago.

On the Saudi side, Humain — the PIF-backed company launched with a mandate to make the Kingdom the world's third AI power — has broken ground on its first 100MW-class sites in Riyadh and Dammam, with 1.9 gigawatts targeted by 2030. It has regulatory approval for 18,000 Nvidia GB300s with "several hundred thousand" more in the pipeline, a $10 billion joint venture with AMD, a planned 500MW project with xAI, and a $100 billion investment vehicle behind it. Aramco Digital, meanwhile, signed with Cerebras in May to deploy CS-3 wafer-scale systems into its cloud business — a deliberate hedge on the non-Nvidia architecture.

And floating above all of it sits MGX, Abu Dhabi's AI investment vehicle, which has quietly assembled the most interesting portfolio in the industry: a founding equity position in the Stargate venture itself, repeated follow-ons into OpenAI, a co-lead of Anthropic's $30 billion Series G and participation in its $65 billion Series H at a $965 billion valuation, plus a position in xAI. By early 2026, total committed Gulf capital to US technology — across Saudi pledges, UAE deals and Qatari ventures — was estimated at roughly $2.5 trillion.

That is not a country picking a vendor. That is a region buying optionality at sovereign scale.

The sovereign counter-narrative is real — and smaller than it looks

The Gulf's sovereign AI story is not vapourware. Abu Dhabi's Technology Innovation Institute shipped Falcon-H1 Arabic in January — a hybrid Mamba-Transformer model in 3B, 7B and 34B sizes that tops the Open Arabic LLM Leaderboard. Inception, G42's model house, maintains the JAIS family up to 70B parameters, purpose-built for Arabic NLP. These models matter: they encode Gulf dialects, regional legal context and cultural ground truth that frontier Western models still handle clumsily, and they can be deployed inside UAE jurisdiction so sensitive government data never leaves.

But hold the two stories side by side and notice the asymmetry. Falcon-H1's largest Arabic variant is 34 billion parameters. Anthropic and OpenAI are training systems on gigawatt-scale clusters that the Gulf is paying to build — for them. The sovereign models are genuinely sovereign and genuinely small; the frontier compute is genuinely enormous and genuinely not sovereign at all. Calling both "the national AI strategy" papers over the fact that they live in different weight classes and different jurisdictions of control.

The paradox nobody puts in the brochure

Here is the compute-sovereignty paradox in one sentence: every "sovereign" AI facility in the Gulf runs on American silicon, shipped under American export licences, governed by American security conditions, revocable by the stroke of an American pen.

The receipts are public. The Biden-era AI diffusion rule put the UAE and Saudi Arabia in Tier 2 — capped, case-by-case, second-class. The Trump administration rescinded that framework, and in November 2025, after the Saudi Crown Prince's Washington visit, Commerce approved sales of up to 35,000 advanced chips each to G42 and Humain. But read the conditions. The UAE got there by ripping out its Chinese-developed AI stack, letting go of Chinese personnel, and divesting Chinese tech holdings. G42 operates under what it calls a "regulated technology environment" — a gold-standard compliance regime designed in dialogue with the US Bureau of Industry and Security, which monitors compliance on an ongoing basis. Larger allocations require government-to-government assurances. And as of March, Washington is drafting new rules that would condition future chip exports on foreign firms investing in the United States — formalising the Gulf playbook into the price of admission.

I don't say this to sneer. The Gulf states made a clear-eyed trade: alignment with the American stack in exchange for access to the only compute that matters this decade. Given that the alternative — Chinese silicon — is two-plus generations behind on AI training workloads and getting further sanctioned by the month, it was the right trade. But let's name what was traded. The Gulf did not buy compute sovereignty. It bought a seat at America's table, on a lease that renews at Washington's discretion, with rent denominated in capital commitments and strategic compliance. A licence that was granted in November 2025 because of a successful state visit can be narrowed in 2027 because of an unsuccessful one.

The equity hedge is the cleverest part of the bet

Which is why I think the most underrated actor in this story is not G42 or Humain but MGX — because equity behaves differently from licences.

Chips can be embargoed. Data centres can be idled by a licensing pause. But a shareholding in OpenAI, Anthropic and xAI simultaneously cannot be un-owned by a policy change, and it pays out regardless of which lab wins. While the infrastructure deals lock the Gulf into the American stack, the equity programme gives Abu Dhabi a claim on the upside of the entire frontier — including labs that compete with the very hyperscalers building on Emirati soil. Co-leading Anthropic's Series G while funding OpenAI's Stargate is not confusion. It is the same logic a sovereign wealth fund applies to oil markets: own the spread, not the position.

Saudi Arabia is running a version of the same play with different weights — more domestic build (Humain's 1.9GW), more architectural diversification (AMD, Cerebras, Qualcomm alongside Nvidia), an xAI joint venture rather than broad lab equity. Riyadh is betting harder on becoming a place where AI happens; Abu Dhabi is betting harder on owning pieces of whoever makes it happen. Both are rational. They are not the same bet, and it will matter which one ages better.

Where leverage actually lives when the weather turns

So: when geopolitics shifts — a new US administration, a Taiwan crisis, a regional war that makes Washington nervous about where its GPUs sleep at night — what does the Gulf actually hold? This is where I'll give you my opinion rather than the consensus.

The GPUs themselves are the weakest form of leverage in the entire stack. Hardware depreciates, licences are conditional, and the supplier holds the renewal pen. Anyone in the region whose AI strategy reduces to "we have the most Blackwells per capita" is holding an asset whose strategic value Washington can mark down unilaterally.

Capital is stronger, but only while it is still being deployed. The $2.5 trillion in commitments buys enormous influence today precisely because America's AI buildout is capital-hungry and its grid is strained. The moment the cheques clear, the leverage transfers to the recipient. Pledged money is power; spent money is sunk cost.

Energy is stronger still, and chronically underrated in this conversation. The Gulf can deliver firm gigawatts — gas, solar, nuclear — at speeds and prices that American utilities currently cannot. Stargate UAE exists in Abu Dhabi rather than Ohio partly because of this. Unlike chips, energy advantage is native, unsanctionable and compounding. If I ran strategy for a Gulf entity, I would treat the energy-to-compute pipeline as the crown jewel and staff it accordingly.

And the durable layer — the one that survives any licensing freeze — is the boring sovereign stack: Arabic-first models and the data flywheels behind them, the regulatory regimes (the UAE's data-residency rules, Saudi's PDPL) that make local deployment mandatory, the talent pipelines at MBZUAI and KAUST, and equity positions in the model layer. None of it photographs as well as a gigawatt campus. All of it is still yours the morning after a policy reversal.

My honest read: the Gulf is running a barbell, whether or not it is articulated that way internally. One end is deep, deliberate dependence on the American stack at the infrastructure layer — accepted, priced, and exchanged for access and influence. The other end is genuinely sovereign but modest: local models, local data, local rules, plus equity claims on everyone's upside. The middle — the comforting idea of "sovereign hyperscale compute" — is mostly branding. American GPUs under American licence conditions are not sovereign no matter whose flag flies over the data hall, and the smartest people in Abu Dhabi and Riyadh know it. The strategy works precisely because they know it.

What I'd watch over the next twelve months

Four tells will reveal which way this is breaking. First, the fate of Washington's proposed export framework: if chip access is formally tied to US investment quotas, the lease terms just got codified — watch whether the Gulf negotiates carve-outs or simply pays. Second, whether Stargate UAE's Q3 phase-one delivery actually lands on schedule, because the region's credibility as a build partner is priced into every future deal. Third, whether Humain's architectural diversification (AMD, Cerebras) produces production workloads or remains an insurance premium. Fourth — and most telling — whether the next generation of Falcon and JAIS models trains on Gulf soil at meaningful scale, or whether "sovereign AI" continues to mean small models at home and frontier ambitions outsourced to partners.

The Gulf's bet is not hyperscaler versus sovereign. It is a wager that by the time the geopolitical weather turns, the region will have converted a window of American need — for capital, for energy, for aligned partners — into assets that no licence can revoke. The deals signed so far buy the window. They do not, yet, buy the conversion. That part is still to play for.