AI's Next Moat Is Made of Friction

A robot stops at the border. A model doesn't. In November the FCC banned imports of advanced Chinese humanoid robots on national-security grounds ^[1]. Washington is still trying to figure out whether it can legally ban a Chinese AI model at all — the First Amendment gets in the way, and an open-weight model can be pulled from a private server no customs agent will ever see ^[2]. That asymmetry is the whole story of where AI is going. The reason it matters now is that the model layer has stopped being a place to build a moat. OpenAI cut prices on its Luna model by 80 percent to fend off Chinese rivals, and Anthropic, xAI, and Meta all followed ^[3]. Open-source Chinese models like Qwen and DeepSeek have closed the performance gap with the closed US systems. And the pure software players are now trying to build walls out of restricted access and premium pricing, on the frank admission that frontier models "are easily replicated" ^[4]. When the thing you sell can be copied, you stop selling the thing and start selling the gate. Every player with manufacturing muscle has drawn the same conclusion. Tesla is scrapping the Model S and Model X to convert its Fremont factory into a hub for its Optimus humanoid, which Musk calls the company's most important product ^[5]. NVIDIA is building the entire stack, from sensor-to-processor integration with Infineon and NXP to a standardized safety system called Halos, already deployed at Amazon and Toyota factories ^[6][7]. Google DeepMind launched Gemini Robotics 2 with the stated goal of "physical AGI" — the intelligence layer for every robot ^[8]. Fei-Fei Li, who built ImageNet, and Yann LeCun, the deep-learning pioneer who has spent years criticizing language models, have both left text behind for "world models" that learn the physics of space and time rather than the statistics of words ^[9]. And China is already at industrial scale: more than 40,000 humanoid robots shipped in the first half of this year, 97 percent of the global total ^[10]. But the physical path does not glide past regulators the way software did. It runs straight into them. Hyundai's 40,000-member union has vowed that not a single robot using new technology enters a workplace without a labor agreement ^[11]. The UAW's Shawn Fain has framed the fight in stark terms.

We are in a fight for humanity. — Shawn Fain

At CES this year, humanoids that danced on stage couldn't walk on carpet ^[12]. And the deepest bottleneck is the hand itself: China's own standard-setting body identified tactile sensing as the unsolved problem, with 80 percent of the tasks where humans still beat machines being touch-related ^[13]. Here is the reversal. Every one of those frictions is also a wall. A Chinese robot can be stopped at customs — the FCC already did it. A Chinese model cannot be stopped anywhere, because it is a file. The industry spent two years outrunning regulators in software, where the product was weightless and the border was meaningless. Now it is walking voluntarily into the most regulated domain on earth — factories, unions, safety codes, import bans — because that is where the moat is. Consider what that means for who wins. The advantage is moving from the best coders to the best manufacturers. And the thing that still defeats 80 percent of human tasks — the sense of touch, the thousands of receptors in a hand — is also the one thing no competitor can copy by downloading a file. You can replicate a model overnight. You cannot replicate a hand.