National security advisor Divyansh Kaushik on AI diffusion to the Gulf, Huawei's chip gap, and why America needs an open-source flood strategy

Summary

The Trump administration's Gulf AI deals represent a strategic inversion of Biden-era policy. Where the January 2025 Biden diffusion rules prioritized preventing chip smuggling to China, blocking Gulf offshoring, and restricting cloud-based remote access by Chinese entities, the current approach treats advanced AI technology as economic leverage — chips in exchange for reciprocal one-to-one investment in the United States and security concessions. Key details remain unresolved, including whether security conditions attach to specific companies or to the UAE and Saudi governments as sovereign entities, and whether Gulf partners will be permitted to co-locate Huawei infrastructure alongside US chips in the same data centers.

Divyansh Kaushik, national security advisor at Beacon Global Strategies, points to the Microsoft-G42 deal from April 2024 as the template that worked. That transaction was accompanied by an intergovernmental assurance agreement requiring G42 to phase out Chinese hardware within a defined period and restrict military and intelligence access. The current Gulf agreements have not yet reached that level of specificity.

Huawei's Chip Gap Is Real but Shouldn't Invite Complacency

TSMC shipped approximately 2.9 million dies to Huawei in violation of US export controls — sufficient to produce roughly 900,000 to 1 million Ascend 910B chips. That is a significant stockpile, but domestic Chinese AI demand is likely to absorb most of it. A recent Malaysian announcement of an Ascend data center, which the minister subsequently walked back, involved only 3,000 Ascend chips — a figure Kaushik describes as inconsequential.

On a quality basis, the latest Huawei server uses technology Nvidia was deploying seven years ago, compensating by packing more GPUs onto a single chassis to approximate comparable memory and interconnect bandwidth. Yields on Ascend die production sit at roughly 20%, making unit economics non-viable without heavy state subsidy. Export controls on ASML's extreme ultraviolet lithography machines have kept China locked at 7-nanometer production using multi-patterning techniques, a constraint Kaushik estimates set Chinese semiconductor development back seven to ten years.

China's own export control framework is nascent by comparison. Its export control law passed in 2020, and the detailed dual-use item rules did not arrive until December 2024, leaving the system operationally immature relative to the US Bureau of Industry and Security.

The Open-Source Flood Strategy

For countries that Kaushik characterizes as geopolitical swing states — nations currently operating within China's orbit that could be drawn toward American technology infrastructure — he argues the US should pursue an aggressive open-source deployment strategy. The commercial pricing argument is straightforward: no one in Kenya will pay $200 per month for access to frontier AI models. The US Development Finance Corporation and Export-Import Bank should be financing deployment of American open-source models in these markets before Chinese alternatives entrench.

The security case for crowding out Chinese models is more pointed. DeepSeek, which has documented ties to the PLA that Kaushik says have been scrubbed from the internet, is a state-influenced system. He references an Anthropic paper on sleeper agents — models with latent behaviors that activate under specific conditions — as a credible threat vector. A state actor with that capability and access to critical infrastructure or consumer devices could trigger harmful behavior months or years after deployment. The US currently focuses export controls on outbound technology flows but has not paired that with equivalent import restrictions on Chinese AI software entering American systems or being deployed abroad with US backing.

Direction of Travel

Asked whether the US is better positioned heading into the second half of 2025 than it was a year ago, Kaushik's answer is cautiously affirmative. The strategic goal, in his framing, is not to freeze Chinese AI development entirely — that has never succeeded in any technology race — but to arrive first and maintain a compounding lead. On that metric, he believes the US is moving in the right direction, provided it pairs export restrictions with a genuine offensive strategy to extend its advantage rather than relying on denial alone.