US Reinstates Nvidia H20 and AMD MI308 China Sales — But 15% of Revenue Now Goes to Washington

Overview: Export Licenses Return, With a Price Tag

The US government has lifted a previous ban on sales of Nvidia's H20 and AMD's MI308 AI accelerators to China — but only under a new and surprising condition: exporters must remit 15% of the revenue from those shipments to the US Treasury. The move reverses an outright prohibition introduced amid rising tech tensions and places a novel financial tariff on high-performance AI chips.

Background: Policy U-Turn and the Timeline

After the Biden administration imposed strict limits on advanced AI processors destined for China over national security concerns, the Trump administration expanded those measures in early 2025 into a de facto ban. That ban was partially reversed in July 2025, when Washington reopened export licenses for the H20 and MI308 — paired with this unprecedented 15% revenue carve-out.

What chips are affected?

The policy specifically targets Nvidia's H20 and AMD's MI308 — data-center AI processors optimized for inference and large-scale model deployment. Both chips were originally engineered to comply with prior US export control parameters, but they still represent significant compute for training and inferencing workloads.

Product Features and Technical Profile

Both the Nvidia H20 and AMD MI308 are designed for AI workloads: high-throughput tensor operations, low-latency inference, and dense matrix multiplications that accelerate transformer models and other neural networks. Key features include specialized accelerators for AI inference, memory architectures tuned for model weights and activations, and software stacks that integrate into common AI frameworks.

Comparisons and advantages

• Nvidia H20: Known for mature software ecosystems and broad framework support, strong developer tools, and established deployment footprints in cloud and enterprise environments.
• AMD MI308: Competes on performance-per-dollar and open software interoperability in targeted workloads, increasingly popular with vendors seeking alternatives to Nvidia.

Both chips are optimized for inference rather than only raw training, making them attractive across consumer, enterprise, and research use cases.

National Security Debate

Security experts have long warned that advanced AI accelerators can be repurposed for military applications. A coalition of around 20 security specialists urged Commerce Secretary Howard Lutnick to block sales, arguing that chips 'optimized for AI inference' could enable autonomous weapons, surveillance platforms, and sharpened battlefield decision-making.

Critics say the 15% revenue requirement does not address the core national security risk. Deborah Elms, head of trade policy at the Hinrich Foundation, summed it up: 'You either have a national security problem or you don't. A fee doesn't make that problem disappear.'

Market Impact, Vendor Pressure and Pricing Questions

Vendors face a strategic dilemma: accept the fee and preserve market access to the world's largest AI hardware buyer, or forgo Chinese sales to adhere strictly to security concerns. Nvidia told the BBC it follows US export rules for global market participation and hopes export controls will allow American firms to compete in China and beyond. Whether Nvidia and AMD will absorb the 15% levy or pass it to Chinese buyers remains uncertain.

Charlie Dai, VP and principal analyst at Forrester, called the arrangement 'unprecedented' and warned that it highlights the steep cost of market access amid escalating tech trade tensions. The new policy injects substantial commercial pressure and strategic uncertainty into supply-chain decisions.

Use Cases and Risks

Potential civilian uses include cloud inference workloads, recommendation engines, and industrial automation. But the dual-use nature of AI chips makes diversion to military projects a persistent risk, exacerbated by reported import loopholes and third-party routing that have previously allowed powerful processors to reach restricted end users.

What’s Next for AI Chips and Export Controls?

The 15% revenue mechanism is likely to provoke legal, commercial, and diplomatic responses. Expect greater scrutiny of end-user declarations, tighter compliance regimes, and renewed debate over whether financial offsets can substitute for hard export controls. For tech vendors and cloud providers, the choice will be a balancing act between commercial opportunity and geopolitical risk — with AI inference hardware squarely at the center of the contest.