The US-Gulf AI Alliance and the Future of AI Power

On 14 February 1945, Franklin D. Roosevelt met Saudi King Abdul Aziz aboard the USS Quincy and put together an important deal: Washington guaranteed security and engineering expertise (via what became Aramco), and in return, Riyadh guaranteed reliable petroleum priced in U.S. dollars. The agreement birthed both OPEC-era petrodollars and a successful seventy-five-year geostrategic alliance.


Fast-forward to 2025. President Donald Trump recently visited Saudi Arabia and other Gulf countries, striking a new bargain. Gulf sovereign funds will bankroll and host vast AI campuses, stocked with top American chips, power infrastructure, and cloud software. American technology companies will provide engineering know-how and advanced semiconductors. US companies will receive Gulf funds data-center build-outs in the US, keep majority ownership of the Gulf clusters and maintain strict control of the semiconductors. This new partnership for the AI age will align Gulf countries to the American technology ecosystem and its standards.

The Global Race for AI Supercomputers

The US is making sure its global AI leadership is solid by keeping the foundation strong at home. Some of the major AI infrastructure projects underway include:

China — and to some degree, Europe, India, and ASEAN states — are also building “compute sovereignty” programs and offering financial support for building data centers. They believe that whichever bloc trains the largest frontier models will enjoy military and economic advantages.

Although gains from scaling up are decreasing, being first place in a winner-takes-most market can still yield outsized leverage.

A Decentralized Alternative?

In some obvious ways, AI is not like oil. A tanker can’t quietly switch from Gulf crude to Russian Urals mid-voyage, but a data center can easily switch a US model for a European or Chinese one. The political allegiance of compute is therefore more fragile.

This creates space for decentralized alternatives. Spare GPUs worldwide could be used for AI training and inference. Entrepreneurs are already active in this space. For example, OpenDiLoCo (short for Open Distributed Low-Communication) reproduces Google’s distributed learning system with open-source code, letting thousands of hobbyist nodes co-train a 10 billion parameter model.

A couple of experiments that push the envelope include:

· Swarm learning, which lets data stay on-site while gradients are shared

· INTELLECT-2 (32 billion parameters) shows reinforcement-learning agents trained across a heterogeneous “compute commons,” with no single data center required

These projects are tiny compared with Stargate. Though they demonstrate that “AI factories” don’t need to be steel monoliths, they have also not yet unveiled a viable alternative model at scale.

Expect a hybrid architecture

The AI race is well underway. While the US leads in AI model innovation, and its government strikes wide-reaching deals that mirror FDR’s oil-handshake, China is a fast-follower in AI model innovation and leads the world in related areas such as battery technology and the adoption of industrial robots.

However, the next decade will look less like the Exxon-Aramco duopoly in AI and more like a patchwork. This will include state-aligned megaprojects to guarantee minimum national compute baselines (like Stargate and Colossus in the US), allied regional hubs (in the Gulf and likely in Southeast Asia), and decentralized alternatives that treat GPUs as fungible atoms on the open internet. While the intangibility of AI will create chances for decentralized competitors, a scalable decentralized model to train world-leading frontier AI systems will likely require both new technical advances and significant user awareness and adoption.