U.S. Resumes Export-Compliant AI Chip Sales to China Amid New Tariff Regime
Following a July 2025 policy reversal, the U.S. government has cleared Nvidia and AMD to resume shipments of export-compliant AI accelerators to China, subject to a new 15 percent levy on sales.
After a brief ban on compliant chips in April 2025, the Trump administration reversed course in July 2025, allowing Nvidia and AMD to resume shipments of export-controlled AI accelerators to China. According to Built In, this decision followed significant lobbying efforts by Nvidia CEO Jensen Huang and White House AI and crypto czar David Sacks. The administration also authorized the sale of Nvidia’s H200, the company’s second-most-powerful chip, alongside the H20. While this restores access to vital compute capacity, the government introduced an unprecedented 15 percent fee on these specific chip sales, which is payable directly to the U.S. Treasury.
The H20, which Nvidia engineered to meet U.S. export-control limits, features 96 GB of HBM3 memory and 4.0 TB per second of memory bandwidth. According to Built In, Nvidia has shipped over a million of these units to China since late 2024. While these chips allow Chinese firms to maintain existing AI research pipelines, they remain distinct from the flagship Blackwell-based GPUs, such as the RTX PRO series launched in March 2025, which remain restricted. The current regulatory framework effectively creates a green zone for lower-powered models while keeping top-tier processing power reserved for the U.S. and its allies.
The reopening of these sales channels is a commercial priority for U.S. vendors, as China accounts for roughly 13 percent of Nvidia’s revenue and hosts approximately half of the world’s AI development companies. Built In reports that Chinese firms like DeepSeek, Baidu, and Alibaba have been forced to innovate by optimizing model architectures to extract performance from midrange chips. For instance, DeepSeek achieved near-frontier performance by utilizing millions of H800 chips—which fall below the export-control threshold—through efficient software techniques. This reliance on U.S. hardware persists, with an estimated 75 percent of AI training in Chinese data centers still running on Nvidia’s CUDA platform.
Despite the renewed access, Chinese firms are continuing to diversify their hardware portfolios to mitigate future supply chain risks. Major players such as Baidu and Tencent are investing in domestic chip alternatives, aiming to fortify self-reliance against potential future restrictions. According to Built In, while Chinese labs have successfully stretched existing hardware capabilities, they remain deeply dependent on American GPUs to sustain frontier-level progress. The current policy represents a calibration between maintaining U.S. competitiveness and allowing American firms to capture revenue from the Chinese market.
The interplay between export controls and tariff brinkmanship defines the current landscape. By levying a 15 percent fee on compliant chip sales, the U.S. government has turned a regulatory tool into a revenue stream. As Nvidia ramps up H20 shipments to meet pent-up demand, the race for AI supremacy continues. While domestic self-sufficiency remains a long-term goal for Beijing, the immediate availability of green-zone GPUs is expected to accelerate AI development in China, allowing companies to move beyond the stopgap workarounds necessitated by the previous export restrictions.