American companies will spend roughly $700 billion this year building out the physical infrastructure of artificial intelligence: data centers, chips, and raw computing power. Every dollar, as former deputy national security advisor Matt Pottinger and former National Security Council China director Liza Tobin put it, 'is a bet on American capitalism' — that the advances those facilities produce will earn enough to fund the next breakthrough.
That bet is being undermined. Not by fair competition, but by a technique the authors call distillation.
Here is how it works. A state-of-the-art American model is asked millions of questions. Its answers are recorded. A new model is then trained on those answers, inheriting much of the original's capability without requiring anything close to the research investment that produced it. Done with permission, distillation is an accepted research practice — Apple pays Google to distill its Gemini model to improve Siri, and university labs use the technique legitimately. Done without permission, it is something else entirely.
Chinese AI models — DeepSeek, Moonshot's Kimi K3, Alibaba Cloud's Qwen — barely registered in American boardrooms before last year. They now run inside Airbnb, Coinbase, DoorDash, and many others. Their open-weight architecture, meaning the code is freely downloadable and customizable, makes adoption frictionless. American firms, chasing cost and convenience, are integrating products that Pottinger and Tobin argue owe their capability in significant part to unauthorized distillation of proprietary U.S. models.
The authors are direct about the pattern: 'Beijing has run its playbook of brute-force economics over and over in recent decades.' Steal intellectual property. Subsidize domestic producers. Block foreign competitors from the home market. Then flood global markets with the resulting product. Solar panels. Semiconductors. Electric vehicles. Now, artificial intelligence.
The open-weight framing deserves scrutiny. Because Chinese models are freely downloadable, the cost of switching to them approaches zero for any American developer. That is a powerful market incentive — and it is precisely the kind of incentive that Beijing's model is designed to exploit. The subsidy is invisible; the disruption is real.
---
Say it plainly: the United States is financing, through its own private capital, a research base that a strategic adversary is systematically harvesting. The $700 billion in annual AI infrastructure spending assumes that the returns on innovation will accrue to the innovators. Distillation-at-scale, conducted without authorization, breaks that assumption at the foundation.
The free market works when property rights are enforced. When a competitor can replicate years of R&D investment by querying a model millions of times and recording the answers, the incentive to invest in the next breakthrough erodes. This is not an argument against open-weight models in principle — it is an argument that the rules governing what can be distilled, by whom, and under what conditions, are a matter of national economic consequence, not merely a licensing footnote. The record is public. The question is whether Washington will read it before the next round of the playbook completes.
