Anthropic wants tougher US export controls on AI chips

Smugglers are hiding high-end AI chips inside prosthetic silicone baby bumps. Others are packing GPUs into seafood containers right next to live lobsters.

These aren't plot points from an international thriller. They are documented smuggling methods cited directly by Anthropic in its latest policy submission to the U.S. Department of Commerce.

Ahead of the May 15 deadline, Anthropic is pushing Washington to aggressively tighten the "Diffusion Rule"—the interim framework rolled out in January that restricts the global flow of advanced chips and model weights.

Closing the 1,700-chip blind spot

The Commerce Department's framework splits the world into three tiers based on national security risks. Tier 1 covers close allies with minimal red tape, Tier 3 locks out adversaries, and Tier 2 includes most everyone else under moderate limits.

Anthropic points out a massive loophole in Tier 2. Right now, buyers in those countries can import up to the equivalent of 1,700 Nvidia H100 GPUs—about $40 million worth of hardware—without needing a government license.

That threshold makes it absurdly easy for shell companies to place repeated orders just under the limit and funnel silicon into restricted territories.

Anthropic is asking Commerce to lower this exemption threshold significantly. It also wants increased funding for the Bureau of Industry and Security to crack down on evasion, while offering a carrot: Tier 2 nations that prove strict data center physical security can unlock more capacity via direct bilateral treaties.

Crucially, Anthropic warns against pushing back the May 15 implementation date, noting that rival firms are currently on an aggressive buying spree to stockpile whatever they can.

The DeepSeek counter-narrative

While many observers viewed the rise of DeepSeek as proof that hardware export controls are ineffective, Anthropic argues the exact opposite.

Anthropic says Chinese labs like DeepSeek openly face compute bottlenecks that force them to consume two to four times more electrical power just to match U.S. frontier performance. DeepSeek also relied heavily on advanced processors stockpiled before restrictions kicked in.

Because top-tier computing power still roughly doubles every two years, Anthropic projects that by 2027, training frontier models on older hardware will be ten times more expensive than doing so on cutting-edge American silicon.

Anthropic also draws a parallel to green tech. The U.S. share of global chip manufacturing crashed from 40% in 1990 to just 12% today, with 90% of advanced manufacturing located overseas. Anthropic insists that without strict export frameworks and compute caps abroad, advanced AI data center buildouts will flee American shores just like solar panels and batteries did.

Why it matters

For developers and enterprise builders outside Tier 1 allied nations, the hardware landscape is about to get much more complicated. Buying large-scale GPU capacity off the shelf will soon involve intrusive compliance checks, background reviews, and significantly higher pricing.

For the AI industry at large, Anthropic's filing signals that the battle for dominance is no longer fought solely on algorithmic efficiency. Top U.S. labs are openly leveraging federal trade policy as a defensive moat against foreign open-weight alternatives.

We think Anthropic's maneuver is politically brilliant but strategically precarious. Tying corporate strategy to national security wins applause on Capitol Hill and keeps competitors off balance.

However, setting the bar too high risks driving emerging markets into alternative chip architectures. When leading labs feel forced to lobby for tighter borders to preserve their lead, it looks like pure technical moats are getting much harder to defend on code alone.