Aaron Ginn on Nvidia GPU exports, the AI Belt and Road, and why the US must win the open-source race

Summary

Aaron Ginn, who runs sovereign AI infrastructure company Hydra and has been advising the Trump administration on GPU export policy, argues that the US has one viable strategy in the global AI race: aggressive, trusted proliferation of American hardware and open-source models before China fills the vacuum.

The Biden-era framework was the wrong map

The January 2025 Biden export rule, drafted under 1980s missile-proliferation law, created a three-tier country system that placed Portugal, Switzerland, Austria, Eastern Europe, Gulf states, and Mexico in a restricted licensing tier alongside genuine bad actors. Mexico — which manufactures most American servers — was flagged ostensibly because of cheap power costs, not any real diversion risk. Ginn has been pushing the Trump administration to scrap that framework in favor of what the Commerce Department announced last week: a trusted-proliferation model that lets allied and commercially important countries buy Nvidia hardware.

Saudi Arabia: 18,000 Blackwells, with more coming

The Saudi deal — 18,000 Nvidia Blackwell chips delivered, with hundreds of thousands more on the way — is the first major public expression of that policy shift. Ginn frames it as a wake-up call made necessary by the fact that Huawei had already deployed in the Gulf before the US moved. His analogy is fighter jets: you can't tell a country it simply cannot have a platform. If the US won't supply, China will, and the customer picks MiGs.

The dependency logic that makes the sale strategically durable, in Ginn's view, mirrors the F-35 supply chain. Buyers need ongoing resupply, software updates, and technical support. Every Nvidia deployment is, in his framing, American-owned real estate — infrastructure the world runs on that ties the operator to the US ecosystem over time. Ginn says Hydra is already involved in roughly a dozen sovereign AI projects internationally, and that Nvidia directs growing data center customers to them as a go-to sovereign deployment partner.

The Huawei threat is real and 12 months from a step-change

Huawei is currently manufacturing at roughly 7nm, possibly stretching to 5nm — versus TSMC's 2nm, heading toward 1.5nm. The process gap is real. What Ginn says should not be underestimated is China's near-complete supply chain sovereignty: it dominates rare earths and owns the full manufacturing stack to produce Ascend chips at scale. Within roughly 12 months, he expects Huawei to offer something comparable to an H100 — cheap, easy to install, fully integrated with DeepSeek and Manus models, and packaged as a turnkey solution. That bundle is already showing up at data centers Hydra works with across Europe and the Gulf.

DeepSeek and Manus are already widely deployed across Europe. Ginn's read is that China uses open source not out of ideological commitment but as a distribution weapon — get the models into every data center globally and make displacement harder. OpenAI's recent announcement of combined software-plus-infrastructure deployments is, in his view, a direct mimicry of that playbook.

The open-source gap is the real second-phase risk

Ginn argues the Trump administration's trusted-proliferation push solves phase one of the contest. Phase two — the one the US is currently losing — is open source. Closed model strategies are not a durable moat; most investors already accept that. The risk is that if Llama and Gemini don't meaningfully close the gap with Chinese open-source models, the infrastructure advantage won't matter. Europe and the Middle East will simply continue training and deploying on Chinese models regardless of where the chips come from.

On DeepSeek R2, Ginn says the consistent rumor is that it will run primarily on Huawei Ascend. He has not verified that, but expects the model to be strong — his framing is that AI research is math and physics, available to anyone who pursues it.

The structural bet: Nvidia is Boeing, not Google

Ginn's investment thesis is that the AI commercialization layer accrues to the platform — Jensen Huang and Nvidia — not to model companies sitting above it. He compares the dynamic to the airline industry: Boeing captures durable platform value, airlines create real businesses, but nobody achieves the parabolic power-law outcomes of a software-defined network like Google or Airbnb. Closed-model AI companies are, in his view, too commoditizable to generate that kind of outlier return. OpenAI's durable value is its brand as an accidental consumer product company — a framing he credits to Ben Thompson — not its model technology.

The fault line on the right

Ginn draws a sharp distinction between what he calls OG China hawks — who believe the US wins by engaging, trading, and proliferating American technology globally — and a newer protectionist strain that wants to restrict GPU sales and retreat from global markets. He argues the latter camp is incoherent on economics: the highest-tariff countries are the poorest, and trade has always been how American power and prosperity compound. The Gulf GPU deals, in his framing, are easy diplomatic currency — straightforward to offer in exchange for fentanyl commitments or reduced auto tariffs — and treating them as a national security concession misreads the actual leverage dynamic.