Palantir CTO calls Chinese open models a threat to America

Signals Inbox·July 18, 2026·AI Trust

Palantir CTO Shyam Sankar says Chinese open models are not simply catching up through better engineering. He argues they are using American frontier models as teachers, then releasing the copied capabilities cheaply and openly. The tension is that America may be paying the huge training bill while China captures the distribution advantage.

Q1What exactly did Sankar say?

In the recorded interview clip, Palantir CTO Shyam Sankar says many Chinese open models are the result of distillation attacks on American frontier labs. His claim is that Chinese developers are extracting valuable behavior from closed models, rebuilding it more cheaply, and releasing the result as open source.

Q2What is a distillation attack?

Distillation normally means using a powerful model as a teacher for a smaller model. The student sends many questions, studies the answers, and learns to produce similar results. The technique itself is common. It becomes an attack when developers allegedly use fake accounts, automated requests, or forbidden access to copy a model at scale.

Q3Is there evidence behind the accusation?

There are serious allegations, but Sankar's full claim is not independently proven. Anthropic has accused DeepSeek, Moonshot, and MiniMax of using around 24,000 fraudulent accounts and more than 16 million Claude exchanges. Anthropic says the campaigns targeted reasoning, coding, agents, and tool use. The Chinese companies have not publicly accepted those accusations.

Q4Why is this becoming urgent now?

Because Chinese open models are getting much closer to the frontier. The UK AI Security Institute reportedly found that China's cyber capability gap fell from roughly six to ten months in 2025 to about four months. Moonshot's Kimi K3 is also competing with leading American models in coding while being released openly and at a lower price.

Q5Why could open source threaten American prosperity?

American labs may spend billions on chips, researchers, electricity, and training runs. If a rival can learn from the finished model for a fraction of that cost, it avoids much of the expensive trial and error. It can then release the copied capability openly, attract developers, lower prices, and weaken the American lab's ability to recover its investment.

Q6What happens next?

Frontier labs will probably tighten account checks, rate limits, model fingerprinting, and access to detailed reasoning outputs. Washington could also treat model access more like chip access and expand export controls beyond hardware. The hard part is balance: making American AI harder to copy without making it harder for legitimate developers to use.

Q7So is China's progress mostly copied?

That would be too simple. Chinese labs have produced real engineering advances in efficiency, model architecture, and open distribution. Distillation may help them close specific capability gaps faster, but it does not explain everything. The real signal is that frontier-model outputs have become strategic assets, and American labs may have protected their chips better than their intelligence.