“You either have a national security problem or you don’t,” said Deborah Elms, director of trade policy for the Hinrich Foundation. If you have a 15% payment, it doesn’t somehow eliminate the national security issue. Her candid analysis gets to the essence of a deal that has both surprised the semiconductor sector and policy circles: Nvidia and AMD will pay the U.S. government 15% of their China revenues on two cutting-edge AI chips for export licenses.
The deal, which was confirmed by a Trump administration official, covers Nvidia’s H20 and AMD’s MI308 chips designed for artificial intelligence tasks and, until recently, caught in the tightening web of U.S. export restrictions. It marks the first instance in which American firms have consented to share revenue with Washington as a requirement for access to the market, a development that Stephen Innes of SPI Asset Management characterized as “part diplomacy, part protection racket… a political tariff in everything but name.”
The H20, which was made following the 2023 restrictions of the Biden administration, was specifically made to meet previous performance limits for the Chinese market. It is based on Taiwan Semiconductor Manufacturing Company’s cutting-edge process nodes, and it provides high throughput for AI inference while remaining just below interconnect and compute limits enforced by U.S. regulators. MI308, AMD’s data center-class graphics processor, provides comparable AI acceleration features, with tensor-optimized cores and high-bandwidth memory interfaces being essential for training and use of massive language models. Both processors are pivotal to AI infrastructure, where hardware represents more over 80% of the bill of materials in data center installations.
So far, these items were essentially excluded from China under April’s expanded restrictions, as part of a larger U.S. strategy to restrict Beijing’s access to high-end AI compute. That policy framework, which began in 2022, aimed not just at chips but also at the lithography equipment and design software to make them. The controls have held back China’s indigenous chip manufacturing barred Semiconductor Manufacturing International Corporation from purchasing extreme ultraviolet lithography tools but have not dissuaded Chinese companies from developing rival AI models or from smuggling banned processors into the country via shell firms and intermediaries.
For Nvidia, the risks are astronomical. Being excluded from China already cost the company an estimated $4.5 billion in one quarter. Analysts anticipate pent-up demand to fuel a spurt in H20 orders now that licenses are issued, with the company already committing hundreds of thousands of extra wafers from TSMC. AMD’s possible upside is just as substantial Melius Research estimates up to $3 billion of recurring quarterly revenue from MI308 sales to Chinese buyers once production ramps.
The geopolitical context for the deal is as important as the commercial terms. President Donald Trump has threatened to impose a 100% tariff on foreign-made semiconductors unless firms produce in the United States, using Section 232 national security authorities to remake supply chains. Whereas the White House has made exemptions for companies such as Apple and Samsung that have committed large U.S. investments, the revenue-sharing paradigm of Nvidia and AMD is a new hybrid of export and tariff thinking permitting sales but taxing them at their origin.
Security officials are still agitated. In a letter to Commerce Secretary Howard Lutnick, 20 former officials cautioned that such chips as the H20, while sold to civilian companies, might “enable autonomous weapons systems, intelligence surveillance platforms and rapid advances in battlefield decision-making.” Beijing has brushed aside such fears as “unilateral bullying” even as it picks up the pace of domestic GPU research through state-sponsored companies such as Biren Technology.
Technically, H20 and MI308 are not the strongest members of their respective families, but their designs continue to provide significant amounts of AI acceleration nonetheless. The H20, for instance, supports high-speed NVLink interconnects though with lesser bandwidth than unlimited models and accommodates large-scale inference clusters. The MI308 supports advanced matrix compute units and high-bandwidth memory stacks, both critical to training efficiency as well as inference throughput. These abilities make them desirable not just for commercial AI applications but also for military-related purposes based on timely data processing.
The economic mechanics of the 15% tax are still unclear; the government has not indicated how it will use the anticipated billions in proceeds. It is certain that the deal reflects the exorbitant price of market access at a time when semiconductor policy comes locked with national security strategy. As Forrester’s Charlie Dai summarized it, the agreement “creates substantial financial pressure and strategic uncertainty for tech vendors” crossing the fault line of the world’s two biggest economies.
For the moment, the deal provides Nvidia and AMD with a thin wedge back into a market they can ill afford to forsake, while providing Washington a new wedge in the continuing battle for AI hardware dominance. Whether it also provides Beijing greater access to the compute power U.S. policy sought to withhold is one open and significant question.
