• The Trump administration has reversed earlier export bans, allowing Nvidia (NVDA) to ship its H200 AI accelerators to approved Chinese customers in exchange for a 25% fee on sales revenue, collected during U.S. security reviews before re-export.
  • This move follows a July decision permitting the lower-spec H20 chips with a 15% cut, but Chinese regulators and buyers largely rejected H20 as underpowered, making it a side issue in the broader shift from security controls to transactional trade deals.
  • The H200, while below Nvidia's newest Blackwell architecture, offers about six times the performance of H20, potentially boosting China's AI compute capabilities if Beijing permits large-scale purchases amid ongoing regulatory and political scrutiny.

In a significant policy pivot, the U.S. government has approved exports of Nvidia's H200 AI chips to select Chinese buyers, marking a departure from the Biden-era "small yard, high fence" security approach toward a more transactional model. According to people familiar with the matter, the deal includes a 25% revenue cut for the U.S. government, collected when parts enter the U.S. for security checks before being re-exported. This development comes after earlier efforts to restructure export rules hit a snag with the H20 chip, which faced lukewarm reception in China due to its inferior performance compared to domestic alternatives.

Nvidia's shares briefly lifted on the news, reflecting investor optimism that even constrained access to the Chinese market could support continued data-center revenue growth. However, the company remains barred from selling its top-end chips like the Blackwell series, and China's regulators are reportedly considering tight approval procedures, including potentially barring public-sector buyers and requiring liability waivers for cybersecurity risks. These measures have already chilled demand, with some Chinese tech firms hesitant to rely heavily on U.S. hardware amid fears of backdoors and sanctions.

Behind the scenes, U.S. Commerce officials are still finalizing the formal export rules, while a bipartisan group in Congress has introduced the "SAFE CHIPS Act" seeking to block exports of advanced chips like the H200 for 30 months. This legislative push directly challenges the Trump administration's move, with several senators condemning the approval as a "colossal economic and national security failure." In China, regulators are weighing restrictive procurement rules that could limit H200 uptake, such as allowing foreign accelerators only when domestic chips are inadequate and making purchasers personally liable for cybersecurity consequences.

The shift underscores a broader trend where advanced GPUs are treated as dual-use technologies, critical to both commercial AI and military applications like autonomous weapons systems. Analysts note that this transactional approach, which monetizes export permissions rather than enforcing pure security controls, complicates coordination with U.S. allies such as Australia, which have aligned their export rules with earlier security concerns. Efforts to reach Nvidia for comment on the H200 deal were unsuccessful, but industry sources suggest the company views it as a standalone regulatory matter within a larger geopolitical contest.

Looking ahead, the future of H200 exports hinges on final U.S. rules, potential congressional action, and Beijing's approval regime. In the short term, Nvidia may gain additional China revenue, but volumes are uncertain due to these regulatory hurdles. Long-term, experts expect China to accelerate its domestic GPU ecosystem, using any temporary H200 access as a bridge rather than a dependency. Meanwhile, related developments like "Operation Gatekeeper," targeting smuggling networks that diverted Nvidia parts into China, highlight enforcement challenges and pressure to create formal export channels. This decision, while narrow, symbolizes a pivotal moment in the ongoing AI hardware race, balancing economic interests against security risks in an unpredictable policy landscape.