The US strategy to restrict AI and chip exports is unraveling. What was meant to protect national security is instead fueling global tech sovereignty movements in China and Europe - exactly the outcome Washington hoped to prevent. Instead of slowing rivals, the policy is pushing allies and adversaries alike to build independent stacks.
On July 17, the Rabbit Hole Recap highlighted how Chinese open-weight models like Kimi K3 are already competitive and 90% cheaper than US counterparts. Marty Bent warned that domestic lobbying for strict rules on open-source models plays into Beijing’s hands, creating an opening for Chinese AI to dominate global developer ecosystems. This economic gravity, he argued, could shift hardware optimization toward Huawei and away from American platforms.
"If the US restricts its own developers from building and sharing open models, global users will flock to Chinese alternatives."
- Marty Bent, Rabbit Hole Recap
The next day, Nathaniel Whittemore reported from the G7 summit where the absence of Anthropic’s Fable 5 and Mythos models loomed large. French President Emmanuel Macron voiced concern that the US holds an AI 'kill switch' - a sentiment echoed across EU policy circles after the US denied UK Prime Minister Keir Starmer an export carve-out. European leaders now see tech sovereignty not as theory but necessity.
Italy’s Bruno Benafee confirmed plans for five European AI 'gigafactories' - a 20-billion-euro effort dwarfed by US hyperscaler spending but symbolically significant. Meanwhile, Microsoft is already adapting: reports suggest it’s fine-tuning DeepSeek V4, a Chinese model, for parts of Copilot. The irony is stark - US policy blocks frontier models from China while US companies embed Chinese code into core productivity tools.
"The mood among EU policy circles turned sour after the US denied UK Prime Minister Keir Starmer a carve-out from export restrictions."
- Nathaniel Whittemore, The AI Daily Brief
A week of mounting evidence shows the strategy is backfiring. Enterprises are abandoning single-model reliance for 'model councils' mixing cheap open-source workers with expensive advisors. Harvey’s experiment pairing GLM 5.1 with Opus 4.7 outperformed Opus alone at a fraction of the cost. The shift isn’t just tactical - it’s structural. As inference optimization becomes a competitive edge, dependency on any single US model becomes a liability.
The US may have won the initial AI race, but it’s losing the geopolitical follow-through. Allies are no longer waiting for concessions. They’re building alternatives - fast.


