The US-China AI rivalry just escalated sharply. On July 22, 2026, Michael Kratsios, director of the White House Office of Science and Technology Policy (OSTP), accused Chinese AI startup Moonshot AI of covertly distilling Anthropic’s flagship Fable 5 model to develop Kimi K3, its recently released open-source model that has taken the AI world by storm.

The Treasury Department has since threatened sanctions if the allegations are confirmed.

What Is Distillation, and Why Does It Matter?

Distillation is a technique where you feed the outputs of a more powerful AI model to a weaker one during training, essentially teaching the weaker model to imitate the stronger model’s behavior. Done legitimately, it is a standard part of AI development. Done covertly using a competitor’s proprietary model, it is widely regarded as IP theft.

According to Kratsios, Moonshot built a sophisticated internal platform capable of conducting distillation attacks across multiple US AI models while “rapidly switching between access methods to avoid detection.” The administration claims Moonshot used this infrastructure to harvest outputs from Anthropic’s Fable 5 while the model was briefly available for API access.

If the allegations hold up, Kimi K3 would essentially be a frontier-quality model built partly on the back of billions of dollars in R&D that Anthropic spent developing Fable 5.

Why Kimi K3 Raised Eyebrows

When Moonshot released Kimi K3 on July 16, 2026, the model immediately attracted attention because it performed at or above the level of Anthropic’s and OpenAI’s best models on major benchmarks. For a Chinese AI startup with a fraction of the compute budget, that result was striking. Experts at the time noted that the leap forward was faster than the investment or timeline would normally explain.

Some researchers have since questioned the timeline. Anthropic’s Fable 5 was only re-released on July 1 after being briefly taken offline due to US export controls, leaving a narrow 15-day window for a distillation campaign to produce a working frontier model. Skeptics argue that large-scale distillation significant enough to produce Kimi K3’s results would take longer than that. The White House has not publicly disclosed how it believes the timeline was compressed.

Moonshot AI has denied the allegations.

What the Government Says It Knows

Kratsios framed the accusation as more than suspicion. He stated the administration had specific intelligence indicating Moonshot AI used Fable while developing K3, and described the platform Moonshot allegedly built as capable of concealing the distillation activity from API providers.

The Treasury Department’s threat of sanctions signals that the administration is treating this as a national security issue, not just an intellectual property dispute. The AI sector has already seen restrictions on chip exports to China. Sanctions targeting a specific AI company for model distillation would be a significant escalation.

Whether sanctions materialise depends partly on what evidence the administration is willing to make public and what diplomatic calculus the White House is running.

What This Means for Business

For enterprises using AI, this story carries several practical signals.

AI IP theft is a real and growing risk. If even Anthropic, one of the most security-conscious AI labs in the world, is dealing with allegations of distillation attacks, every company that makes its models accessible via API should be thinking about how its outputs could be used. This is not a distant geopolitical issue. It is a business risk that becomes more relevant as more companies build proprietary models.

The frontier model gap may be narrowing faster than expected. Whether or not the distillation allegations prove true, Kimi K3’s performance is real. Businesses that assume frontier AI capabilities are only available from OpenAI and Anthropic should update that assumption. Open-weight models that match closed frontier systems change the cost equation for self-hosting and custom development significantly.

US-China AI competition will drive regulatory pressure. Sanctions threats, export controls, and IP disputes are going to produce regulation that affects how enterprises procure and deploy AI. Companies that are currently building on models or infrastructure with ambiguous geopolitical exposure should factor that into their vendor and architecture decisions now, not after the rules solidify.

Trust is becoming a differentiator. As allegations of AI IP theft escalate, enterprises will increasingly need to answer the question: where did this model’s intelligence come from, and can we verify it? Vendors who can demonstrate provenance and clean development practices will command a premium.

At Enterprise DNA, we work with business leaders navigating exactly this kind of complexity. The AI landscape is moving fast enough that a decision you make today about which models and platforms to build on can look very different in 12 months. Getting that foundation right matters.

The Treasury sanctions threat and the Moonshot case are still developing. Expect more detail to emerge over the coming weeks as the administration decides how hard to push.