Palantir's chief technology officer says China built its best AI models by siphoning answers out of American ones. He's less worried about that copying than about Americans turning against the data centers built to stay ahead of it.
Shyam Sankar sat down with Bloomberg Television on Wednesday, July 15. He said the quiet part out loud. "These Chinese open source models are really the result of distillation attacks," Palantir's chief technology officer told the network, describing them as built on "stolen American IP" from frontier labs like OpenAI, Anthropic and Google. It's a blunt charge from a company whose business is selling AI-powered software to the Pentagon and its allies.
Distillation itself isn't exotic. It's how a developer trains a smaller, cheaper model on the outputs of a bigger one, letting it inherit much of the larger model's behavior without paying the full training cost. Anthropic made the sharper version of this argument two weeks earlier, telling the White House that operators tied to Alibaba's Qwen lab opened almost 25,000 fake accounts and ran 28.8 million exchanges through Claude between April and June. Anthropic called it "adversarial distillation" aimed at Claude's strongest skills: agentic reasoning and long, multi-step coding tasks. Alibaba hasn't addressed the specifics.
DeepSeek, Moonshot and MiniMax have faced similar allegations from Anthropic before. None of those claims has been independently proved in public. That's important, because once you move from a terms-of-service violation to a charge of theft, you need evidence a reader can see.
The Real Risk Isn't Copying, Sankar Says
Here's the turn in Sankar's argument, and it's the more interesting one. He said the bigger economic risk to the US isn't Chinese copying at all. It's Americans refusing to build the data centers that keep frontier labs ahead in the first place. He compared the current backlash to the fight over US nuclear power in the 1970s, when Three Mile Island and local activism helped turn a widely backed technology into a stalled industry for a generation.
The comparison has teeth. Data Center Watch tracked more than 75 US data center projects worth about $130 billion that were blocked or delayed in early 2026, and Gallup found 71% of Americans oppose a data center in their own area, more than the 53% who object to a nuclear plant nearby. This week gave Sankar another example. The Financial Times reported that New York became the first US state to suspend large new data center development, with Governor Kathy Hochul pausing projects above 50 megawatts while officials study power and water use, plus what it means for consumer costs.
Sankar's point is simple. Distillation is a copying problem. A stalled data center is a capacity problem. If you believe the next frontier models are still decided by chips and power - and land you can actually build on - capacity is the thing that decides who gets to move first.
A Convenient Argument for Palantir
He's not a neutral narrator here. Palantir's stock and its government contracts depend on the US staying ahead in AI and keeping the compute pipeline moving, so warning Washington against slowing data center growth doubles as an argument for Palantir's own future. That doesn't make him wrong. It does mean you should read the interview as advocacy, not just analysis.
Frankly, calling distillation "theft" is also more contested than Sankar lets on. Training on another model's outputs sits in a legal and ethical gray zone that predates this fight by years, closer in some cases to standard model-building practice than to industrial espionage. American labs have used model outputs in their own development work too. Anthropic's own complaint to the White House leaned on violations of its usage terms, not a courtroom finding that Alibaba stole Claude.
The harder fact for US labs is the price gap. Chinese developers have shipped models that users compare with GPT-4 and Claude-class systems at far lower cost. DeepSeek's R1 model did that in January 2025, when its claimed $5.6 million training cost rattled Nvidia's stock and forced investors to rethink how much compute frontier AI actually requires. Whether that efficiency came from real research gains, aggressive distillation - or some mix of both - is exactly the question Sankar and Anthropic are pressing Washington to answer.
Don't miss the commercial fight underneath the national security language. If cheaper Chinese models can get close enough to US performance, buyers won't wait around for an abstract debate about intellectual property. They'll use what works. You would too, unless regulators or your own risk team tell you the model is off limits.
This won't resolve this year. The direction is clear enough. Expect more public accusations from US labs and defense-tech companies, especially when local governments are also making it harder to build the power-hungry infrastructure those same companies say America needs. Sankar is right about one thing: policing borrowed outputs won't matter much if the US can't persuade its own towns to host the machines needed to stay ahead.
Also read: TSMC posts record revenue and hands Wall Street its verdict on the AI buildout; What Is an AI Browser and How Do Agentic Tools Like Comet Work; Bora Pharmaceuticals and Insilico Medicine Strike a $2.5 Billion AI Alliance
