Treasury Secretary Scott Bessent didn’t mince words. Chinese open-source AI models that steal American intellectual property face sanctions. The warning came amid fresh alarm over Moonshot AI’s Kimi K3. It just claimed the top spot on a front-end coding leaderboard.
But here’s the rub. Once those model weights drop publicly on July 27, stopping their spread gets tough. Nearly impossible, some analysts say. Enterprises already download them. They run the software on private servers. Data stays inside the firewall. Costs plummet.
Bessent spoke plainly on Fox Business. “There’s a very technical AI word for it called distillation, but you and you and I would call it theft,” he said, per Business Insider. “This administration supports open source models, but what we do not support is IP theft.” He pointed to detected watermarks from U.S. large language models embedded in many Chinese counterparts. “We are finding watermarks of our US large language models on many Chinese models, and that’s unacceptable.”
The Open-Weights Challenge
The Trump administration has circled this issue before. Last year officials at Commerce weighed adding Chinese AI labs to the Entity List. NSA and White House cyber teams drafted advisories urging companies to avoid the models. Even an executive order floated requirements that U.S. firms guarantee security and accept liability if a hosted Chinese model suffered a breach. Those ideas stalled. Pro-innovation voices inside the White House pushed back hard.
Now momentum builds again. Kimi K3 changed the conversation. The 2.8-trillion-parameter model from Beijing-based Moonshot AI doesn’t just compete. It outperforms in specific tasks. It matches or exceeds Anthropic’s latest Fable 5 and OpenAI’s GPT-5.6-Sol on several benchmarks. And it activates only 16 experts out of 896 per token. Efficiency by design. Full open weights arrive in days. Tom’s Hardware reported the revival of ban discussions, citing cybersecurity fears.
David Sacks sees danger in the policy response. The former White House AI czar and current co-chair of the President’s Council of Advisors on Science and Technology fired off a post on X. “We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition.” He wrote it Sunday, right after Kimi K3’s leaderboard surge. Sacks argues America ties itself in knots with data-center restrictions and pre-approval rules while Chinese models advance unchecked.
And the economics pull hard. DeepSeek-V4-Pro delivers output for 87 cents per million tokens. Claude runs $50 for the same. Coinbase slashed its AI spending nearly in half by switching. Enterprises gain privacy too. No data leaves the building. Self-hosting shifts GPU, power and maintenance costs to the user. For heavy, sustained workloads it pays off fast.
Yet national-security voices warn of backdoors. Hidden triggers. Data exfiltration risks. The open nature amplifies the threat. Once weights hit Hugging Face or BitTorrent mirrors, recall becomes fantasy. Fine-tunes and quantized versions multiply. Enforcement across borders collapses.
Bessent insists on parity. “These models have to be held to the same standard as US models are, but more importantly, are they using stolen materials?” He named Meta, Nvidia and Reflection AI as American players advancing open-source efforts. The contrast stings. U.S. labs pour resources into transparency while Beijing’s offerings surge in capability and availability.
Recent reporting shows the White House weighs an executive order targeting open-source AI. Not a blanket prohibition. Targeted measures. Procurement bans for federal contractors. Guidance that creates doubt among regulated firms. Fear, uncertainty and doubt. Dean Ball, a former Trump adviser now at OpenAI, suggested as much and drew swift criticism. Bill Gurley penned an op-ed framing open source as the antidote to an AI duopoly.
Anthropic itself has complained about distillation attacks. Yet the company built its own models on vast stores of copyrighted material without explicit permission. Skeptics note the hypocrisy. The debate grows messy. Technical. Geopolitical. Commercial.
So what happens next? Moonshot plans its release for July 27. Markets reacted. Nasdaq dipped about 1% on Friday after the initial Kimi K3 news, according to TechEchelon. U.S. AI leaders watch closely. Some push for outright blocks. Others warn such moves hand China a propaganda win and slow domestic innovation.
The Treasury chief left little ambiguity. Sanctions remain on the table if evidence of theft holds. Watermarks provide one forensic clue. Distillation techniques let smaller models mimic larger ones by training on their outputs. The practice spreads fast in open environments.
But enforcement faces reality. Chinese labs operate beyond direct U.S. jurisdiction. Model files propagate globally within hours. Developers in Europe, Asia and Latin America grab copies. Enterprise adoption accelerates precisely because the models deliver performance without vendor lock-in or eye-watering API bills.
Earlier this month Sacks highlighted the leaderboard jump. Kimi K3 didn’t just place. It led. The result rattled assumptions about American dominance. It also handed ammunition to those who favor lighter regulation at home.
Administration insiders tell Axios the focus has shifted from last year’s aborted efforts. Personnel changes removed some of the strongest innovation defenders. Cybersecurity concerns gained fresh traction. No final executive order has circulated yet. Discussions continue. Options range from Entity List additions for specific labs to softer public advisories that highlight risks without formal prohibition.
One thing looks clear. The era when U.S. policy could ignore capable Chinese open models has ended. Kimi K3 marks a threshold. Performance meets accessibility. Benchmarks meet real-world cost savings. Policymakers scramble to catch up.
Bessent’s message carries weight. Support for open source stops where theft begins. The question now becomes detection, proof and consequences. Watermarks offer signals. But in a world of distilled derivatives and community fine-tunes, tracing origins turns complicated fast.
Industry watchers expect more statements in coming days. The July 27 weight release will test the administration’s resolve. Companies already testing Kimi variants will face fresh scrutiny. Boards will ask general counsels hard questions about compliance and risk.
And through it all the models keep improving. Chinese teams train on domestic silicon. They bypass export controls. They release code and weights that anyone can inspect, modify and deploy. The transparency that open source promises now collides with security fears that span oceans.
Short term, expect guidance. Long term, the contest will test whether regulation can contain software that copies itself across the internet at the speed of download. Bessent drew his line. The coming weeks will show how firmly Washington holds it.
WebProNews is an iEntry Publication