The U.S. Commerce Department's decision to let Nvidia resume sales of its H20 artificial intelligence chip to China came with a price tag attached: a 15 percent cut of revenue from those sales goes to the federal government. AMD agreed to the same arrangement for its MI308. The deals, struck in mid-2025, reopened a market that had been closed by export controls and immediately drew objections from lawmakers and former officials who said national-security decisions were being traded for revenue.
The arrangement is unusual. Export-control licenses in the U.S. have historically been granted or denied on security grounds, not priced. Section 4815(c) of Title 50 prohibits the Bureau of Industry and Security from collecting fees for considering or issuing export licenses. Whether the 15 percent payment is framed as a license fee, a tariff, or something else, it puts the question of authority squarely before Congress.
The revenue deal and its critics
Senators Mark Warner, Jack Reed, Elizabeth Warren, Christopher Coons, and Charles Schumer wrote to Commerce Secretary Howard Lutnick in July 2025 asking how the payments could be reconciled with the statute. House Select Committee on the CCP Chairman John Moolenaar sent a similar letter. Former officials raised a sharper concern: that national-security decisions were being negotiated in exchange for trade concessions or government revenue, contrary to past U.S. practice of refusing to bargain over such terms.
Jensen Huang, Nvidia's chief executive, told CNBC in May 2025 that U.S. chip restrictions had cut Nvidia's China market share nearly in half. Commerce Secretary Lutnick, speaking the same month, described the H20 as Nvidia's '4th best' AI chip, a framing meant to argue that the licensed product is not the company's most advanced. Critics were not persuaded. The Wall Street Journal reported in August 2025 that Nvidia had agreed to the 15 percent payment with billions at risk.
The framing matters because it defines what the license actually buys. If the H20 is genuinely a fourth-tier product, the revenue share is a modest toll on a low-value export. If it is still useful for training large models at scale, the toll is being collected on something that strengthens a competitor's AI capacity. Lutnick's own department has not published a technical assessment that settles the question.
Beijing's countermove
China's response came in two parts. In July 2025, China's cybersecurity regulator summoned Nvidia over a chip-security issue, and Bloomberg reported that Beijing urged domestic firms to avoid the H20. Then on September 15, 2025, China's State Administration for Market Regulation announced that Nvidia had violated the Anti-Monopoly Law and that it would conduct a further investigation.
The SAMR probe turns the commercial opening into a liability for Nvidia. A company that pays 15 percent of its China revenue to Washington while facing an antitrust investigation in Beijing is caught between two governments that both want leverage over the same supply chain. It is not clear this will hold.
The timing of the SAMR announcement, coming weeks after the licensing deal, looks less like a coincidence than a signal. Beijing has used regulatory tools against foreign firms before, and the anti-monopoly statute gives SAMR broad discretion. Nvidia's Mellanox acquisition, completed in 2020, had already been reviewed once; reopening it keeps a standing threat on the table without requiring new legislation.
What the controls actually cover
The Congressional Research Service report R48642, published September 19, 2025, lays out the scope. Advanced logic chips, including FPGAs and ASICs, advanced memory such as high-bandwidth memory and DRAM, and GPUs fall under the restrictions. So do parts of the supply chain: design IP, materials, specialized manufacturing processes, semiconductor manufacturing equipment, electronic design automation tools, and advanced packaging and testing.
Parts of the chain remain open to China. Mature-node technology, third-party computing, some advanced chips through licensing, R&D and open-source technology, materials and intermediates, and training for Chinese staff are still accessible. The controls are not a wall. They are a filter, and the filter has holes that both sides keep finding.
The CRS report also notes that before 2018, U.S. controls and licensing policies allowed firms across the supply chain to sell a range of products and services in China, providing revenue for U.S. firms while allowing China to develop competitive capabilities. The current regime is an attempt to reverse that, but the open channels show how hard a full reversal is.
Supply chain math in 2026
The Semiconductor Industry Association reports that announced U.S. projects will create and support over 525,000 American jobs, including 71,500 facility jobs. SEMI projects memory equipment spending to grow 32 percent in 2026, 15 percent in 2027, 12 percent in 2028, and 21 percent in 2029, driven by AI-related memory demand. The build-out is real, but it runs on a timeline measured in years, not quarters.
Meanwhile, export controls on advanced fab tools to China have hit memory suppliers with large DRAM fabs in China, including Samsung and SK hynix. Memory prices could move as a result. The AI upswing and the structural bottlenecks are happening at the same time, and neither waits for the other.
The domestic build-out also depends on the same equipment that the controls restrict. ASML's extreme ultraviolet lithography systems, Tokyo Electron's deposition tools, and Applied Materials' etch equipment are all subject to licensing. A U.S. fab cannot be built without them, and the queue for delivery is long. The policy that restricts China's access also constrains America's own expansion.
What happens next
Congress is assessing how to advance U.S. competitiveness in semiconductors and AI, and whether the revenue-sharing arrangement survives legal scrutiny. The constitutional question over Article I, Section 9, Clause 5, which prohibits federal export taxes and duties, sits alongside the statutory question over Section 4815(c). Either could unwind the deal.
For Nvidia, the stakes are market share in the world's second-largest economy. For AMD, the same. For Beijing, the question is whether to let domestic firms buy the licensed chips or push harder toward self-reliance. China's 2014 industrial policy set a goal of world-leading semiconductor capabilities across the supply chain by 2030, and U.S. controls have only sharpened that ambition.
The filter will keep shifting. The money, the probes, and the fabs will not wait for it to settle. A useful next data point is whether SAMR expands its Nvidia investigation beyond the anti-monopoly finding into export-control retaliation, and whether Congress moves to codify or block the 15 percent revenue share before the 2026 midterm cycle.
According to Reuters, the framework agreed between the U.S. and China in June 2025 set the stage for these licensing decisions, and the terms are still being negotiated as of late 2025.
See also desk coverage for related stories.