The US-China AI chip situation has developed a layer of absurdity that would be funny if the stakes weren't so high. Earlier this year, the Trump administration announced it would allow Nvidia to export its H200 AI chips to China — provided the US received a 25% cut of the revenue. China's response was to instruct its customs authorities to block imports of H200 chips and warn Chinese tech companies against purchasing them. So Nvidia has US permission to sell, and Chinese buyers who won't buy. CNBC reported in late February that Nvidia has essentially sold none of its approved China chips. The export permission is real; the market for it is not.
Beijing's logic is straightforward. Accepting US-approved AI chips on US terms — with a revenue cut flowing back to Washington — would be a visible subordination of China's technology sector to American commercial leverage. From China's perspective, better to develop domestic alternatives and accept near-term capability gaps than to depend on a supplier that can revoke access at will, attach revenue conditions, or use the commercial relationship as a future bargaining chip. The speed with which Chinese labs adapted after prior export controls — building capable systems on less powerful hardware, open-sourcing models to reduce dependence on any single vendor — suggests this calculation isn't entirely wrong.
The US policy debate has become a tangle of competing concerns. Congressional hawks are pushing back on any relaxation of chip exports, arguing that even H200s provide meaningful AI capability uplift to Chinese military and surveillance programs. Commerce Department officials counter that tighter restrictions simply accelerate China's domestic chip development while depriving US companies of revenue. TechCrunch reported in early March that the administration is now considering sweeping new controls that would require US government approval to ship AI chips almost anywhere outside the US — a significant tightening that would affect not just China but allied countries as well.
Meanwhile, the smuggling pipeline has not been shut down. The Centre for a New American Security estimated that between 10,000 and several hundred thousand AI chips were smuggled into China in 2024 alone — a figure with a wide range that reflects how hard this kind of activity is to measure. The channels run through Singapore, Malaysia, and other third-party jurisdictions. Export controls slow the flow but don't stop it, and the chips that do reach China arrive outside any licensing framework that might otherwise be tracked or conditioned.
Anthropic has entered this debate with a formal position paper on export controls, arguing that US compute advantage is a genuine national security asset worth protecting and proposing specific frameworks for how controls should be structured. It's an unusual move for an AI lab — taking a formal public position on trade policy — but consistent with Anthropic's broader posture of treating AI governance as something it has a stake in shaping rather than merely responding to. Given the company's simultaneous fight with the Pentagon, the timing is also a reminder that Anthropic's disputes with the US government are selective rather than comprehensive: it opposes specific uses of its technology, not US technological leadership as such.