Conditional chip access
Nvidia has received White House approval to export its H200 artificial intelligence graphics processing units (GPUs) to “approved customers” in China, according to President Donald Trump. The decision adds a new layer to the export-control regime that has shaped the global AI supply chain since Washington began restricting the highest-end accelerators. It also signals that the administration is moving toward conditional market access rather than blanket prohibitions for every class of advanced chip.
Trump said the approval comes with national-security conditions and that the US government will take a 25 percent cut of the GPU sales tied to the arrangement. That revenue-share element is unusual in export policy, because it frames licensing not only as a security mechanism but also as a direct fiscal instrument. The announcement positions the deal as supportive of US jobs and manufacturing, while preserving a performance gap at the frontier by excluding Nvidia’s newest flagship Blackwell GPUs and the next-generation Rubin line from the permitted exports.
The H200 sits above Nvidia’s previously allowed China-facing offerings, including constrained models designed to comply with restrictions, but remains below the company’s newest top-tier products. That tiering matters: policy can slow a rival’s access to the most capable training hardware while still permitting meaningful commercial deployment for inference and mid-range training. In AI infrastructure, that is often enough to keep large ecosystems running, even if it limits the pace of frontier model development.
The decision arrives after months of political debate over whether easing chip limits helps US industry compete or undermines US leverage. A core tension is that restrictions can reduce revenue for US chipmakers and accelerate domestic substitution efforts abroad, while looser rules can speed AI progress in competing markets. As a result, administrations tend to oscillate between tight controls and negotiated carve-outs, and each move creates planning risk for suppliers, cloud operators, and enterprise buyers.
Market impact will depend on implementation details and actual demand. Trump’s statement emphasized “approved customers,” which implies a controlled end-user list and compliance oversight that could narrow who can buy and deploy the chips. Even if approvals exist on paper, sales volumes can be limited by uncertainty about future rule changes, the administrative burden of licensing, and the willingness of buyers to commit capital to hardware that could face new constraints later.
There is also a strategic reality on the China side. China has been pushing a stronger domestic semiconductor pipeline and has incentives to reduce reliance on US-origin compute hardware. If Chinese buyers treat supply as potentially unstable, they may prefer domestic accelerators for long-lived deployments while using imported chips tactically where they offer immediate performance gains. That behavior can cap the long-term commercial upside for exporters even when licensing windows open.
For global supply chains, this policy shift reinforces a trend: the chip trade is increasingly governed by deal structures that blend thresholds, carve-outs, and negotiated conditions. That makes compliance a competitive advantage. Companies with stronger legal, logistics, and customer-vetting capabilities can move faster, while smaller intermediaries face higher risk and cost. Over time, those pressures can concentrate distribution and channel power in fewer hands.
Two dynamics will be watched next. One is whether the “25 percent cut” becomes an enforceable mechanism that survives scrutiny and can be extended to other vendors or categories. The other is whether the US draws a stable line between “allowed performance tiers” and “frontier-protected tiers,” because stability is what data center buyers need to plan multi-year capacity and financing. Without that clarity, every approval can also act as a reminder that the rulebook is still in flux.
