US Permits Nvidia H200 Chip Exports to China with 25% Surcharge
The United States has authorized exports of Nvidia’s advanced H200 AI processors to select customers in China, imposing a 25% import tax to fund security reviews. President Donald Trump announced the policy shift on Truth Social, stating it protects national security while reopening a multibillion-dollar market for American semiconductor firms. The move balances economic gains against risks of bolstering China’s AI capabilities, drawing sharp rebukes from security experts.
Trump disclosed the decision after informing Chinese President Xi Jinping, who responded positively. The policy extends to comparable chips from AMD and Intel, with exports limited to approved entities. U.S. customers of Nvidia shift to next-generation Blackwell and Rubin architectures, excluded from this arrangement to preserve technological edges.
The H200, a high-bandwidth memory variant of Nvidia’s Hopper GPU architecture, delivers up to 141GB of HBM3e memory at 4.8 terabytes per second bandwidth. It supports large-scale AI training models, outperforming prior H100 models by 1.4 times in inference tasks. China’s domestic alternatives, like Huawei’s Ascend 910B, lag significantly, with benchmarks showing H200’s 30% faster token generation in Llama 3.1 workloads.
Implementation involves routing chips through Taiwan to the U.S. for mandatory security vetting, higher than the 15% fee floated in August 2025. This process adds latency but ensures compliance with export controls under the Bureau of Industry and Security. Nvidia’s supply chain adapts, with TSMC fabricating H200s on 4nm nodes for enhanced efficiency.
Market reactions split along global lines. Nvidia’s stock climbed 2% in after-hours trading post-announcement, building on a 3% daytime gain from earlier Semafor reporting. China’s SSE STAR Chip Index and CSI Semiconductor Industry dipped over 1% at open but rebounded, reflecting investor bets on eased tensions. U.S. semiconductor ETF PHLX rose 1.2%, signaling broader sector relief.
Critics, including Democratic Senators and Republican Representative John Moolenaar, decry the policy as a “colossal economic and national security failure.” Moolenaar warned China could repurpose H200s for military enhancements and reverse-engineer designs. Eric Hirschhorn, ex-Obama Commerce official, labeled it a “terrible mistake” undermining bipartisan efforts to curb China’s military modernization.
The Institute for Progress analysis highlights H200’s superiority, noting it exceeds China’s legal chip limits by enabling 2.5 times larger models without distillation. Security hawks argue the surcharge fails to offset proliferation risks, potentially accelerating Beijing’s surveillance and autonomous weapons programs. Proponents counter that export bans since October 2024 cost Nvidia $8 billion in lost China revenue, per company filings.
Trump framed the deal as job-creating, projecting 5,000 U.S. manufacturing roles from heightened domestic production. It aligns with his tariff-heavy trade agenda, using fees to subsidize American fabs under the CHIPS Act. Commerce Secretary Gina Raimondo’s team vetted the framework, incorporating end-user certifications to block military diversions.
Broader implications ripple through AI geopolitics. China’s $47 billion national chip fund, announced in May 2025, targets self-reliance, but H200 access could shortcut progress by two years, per RAND Corporation estimates. U.S. allies like Japan and the Netherlands, co-signatories to 2023 export pacts, monitor for harmonization.
Nvidia maintains H200 shipments comply with Wassenaar Arrangement guidelines, capping performance below supercomputer thresholds. The company reports 70% of H200 capacity allocated to non-China markets, with Blackwell ramping in Q1 2026. AMD’s MI300X and Intel’s Gaudi 3 face similar scrutiny, potentially delaying their China re-entry.
Enterprise adopters in China, including Baidu and Tencent, eye H200 for cloud inference, where it reduces energy draw by 20% versus H100. U.S. firms like Meta and OpenAI, reliant on Nvidia for training, benefit indirectly from stabilized supply chains. Yet, the policy tests Biden-era alliances, with EU officials probing reciprocal measures.
As implementation unfolds, quarterly audits by the Commerce Department will track diversions. Trump’s post cited “ironclad safeguards,” but skeptics demand congressional oversight. The surcharge revenue, estimated at $2.5 billion annually, funnels to DARPA’s AI resilience programs.
This pivot underscores tensions in U.S.-China tech decoupling, where economic interdependence clashes with strategic rivalry. While reviving Nvidia’s 15% China revenue share, it exposes fault lines in export regimes, potentially spurring Beijing’s $100 billion retaliatory investments in indigenous silicon. For American startups, it signals opportunities in secure AI hardware, with 25% of venture funding in Q4 2025 targeting post-quantum chips.
