Anthropic CEO Dario Amodei has issued a warning that has resonated in Washington: if the United States does not completely halt the export of advanced chips to China, the American technological edge in artificial intelligence could shrink to “just a few months.” The statement, reported by the South China Morning Post from Shenzhen, is no mere technical opinion—it reveals a battle tearing through Silicon Valley. While Amodei calls for a total blockade on hardware, giants like OpenAI are negotiating to avoid restrictions on open-weight models, and Chinese labs are launching models such as Kimi K3 that are closing the gap at breakneck speed.
The Offensive Against Distillation and Smuggling
Amodei stated that Anthropic has never advocated for a ban on open-weight models. That nuance is key. The Anthropic CEO does not want to shut down open-source; he wants to strangle hardware. In his own words, he described the chip ban as the most efficient and direct way to prevent Beijing from developing frontier AI models for military use and national surveillance. The concrete proposal: the United States should not sell powerful chips or chip manufacturing equipment to China and must crack down on rampant smuggling and workarounds that evade current sanctions.
The most unsettling point was the direct accusation that Chinese developers are using large-scale distillation of American models to bypass hardware limitations. Distillation—a technique that allows training smaller, more efficient models from large ones—has become the Trojan horse of Chinese AI. According to Amodei, this practice could narrow the gap between U.S. and Chinese capabilities to “just a few months,” a horizon that in the industry amounts to strategic irrelevance.
The Silicon Valley Fracture: Open Weights vs. Total Blockade
Amodei’s plea does not come out of nowhere. It arrives amid intense negotiations in Washington, where there is speculation about introducing restrictions on open-weight models to curb Chinese advances. But here the rift widens. While Anthropic pushes for iron-fisted hardware control, OpenAI, as reported by Chinese media 36Kr, has held meetings with the Treasury Secretary, the Commerce Secretary, and members of Congress to discuss AI policies. Sam Altman’s agenda is well known: to prevent software restrictions from damaging the U.S. innovation ecosystem, especially as its own models compete with open-source ones.
The paradox is clear: Anthropic, which builds closed, secure models, calls for shutting down the hardware spigot; OpenAI, which also operates with proprietary models, negotiates to keep software channels open. But the real elephant in the room is that both know that controlling hardware is, at this point, a losing battle. Chinese labs have demonstrated an astonishing capacity for adaptation: the launch of Moonshot AI’s Kimi K3 model is just the latest example of how Chinese innovation advances even with lower-capability chips.
The Shift Toward Commercial Licenses: The End of Open Source
As Amodei spoke in Washington, Goldman Sachs released a report that could change the game. According to the investment bank, Chinese AI developers may be shifting toward a “paid weights” commercial licensing model—a change that would transform the nature of open source in the industry. The idea is simple: instead of releasing model weights for free, Chinese companies would begin charging for commercial use, following a model already being tested by some U.S. companies.
This shift would have profound consequences. If Chinese labs widely adopt paid licenses, the open-source ecosystem that has driven global AI innovation could fragment. On one hand, U.S. companies would lose free access to models that, though trained with limited hardware, are demonstrating surprising efficiency. On the other hand, Beijing would gain a new revenue stream that could be reinvested in R&D, further accelerating the race.
The Geopolitics of Silicon: Who Really Controls the Value Chain?
Amodei’s offensive reveals an uncomfortable truth: control of the AI value chain no longer depends solely on hardware. For years, the U.S. strategy was based on the premise that without access to NVIDIA’s and TSMC’s most advanced chips, China could not compete. But reality has proven otherwise. Distillation, algorithm optimization, and software engineering have allowed Chinese labs to achieve remarkable results with previous-generation hardware.
The launch of Moonshot AI’s Kimi K3 model, which competes directly with U.S. frontier models on complex reasoning tasks, is proof that the gap is closing. And it is not alone: GLM-5.2, developed by Zhipu AI, has demonstrated capabilities that just two years ago seemed reserved for Silicon Valley labs. The question is no longer whether China will catch up to the United States, but when.
A Reflection on the Future: The Blockade That Is No Longer Possible
Dario Amodei’s plea to Washington has a hint of desperation. It is the plea of someone who knows that control tools are becoming obsolete. Banning the sale of advanced chips to China is necessary, but insufficient. Smuggling, distillation, and software innovation have created an ecosystem that evades sanctions with an ease that should worry U.S. strategists.
The future of artificial intelligence will not be decided in Silicon Valley labs or Taiwanese factories. It will be decided by the ability of governments to understand that hardware control is a short-term battle, while the real war is fought on the ground of software, data, and, above all, the capacity to innovate under constraints. China has shown that chip scarcity is not an insurmountable obstacle, but an incentive for creativity.
While Washington debates whether to shut down open-weight models or not, Beijing has already moved to the next phase: commercializing its AI. The shift toward the “paid weights” licenses anticipated by Goldman Sachs is not just a business strategy—it is a declaration of technological independence. And on that new chessboard, Amodei’s chip bans seem less like a solution and more like the final gesture of an empire watching its advantage slip away.