The Distillation Dilemma: Chinese AI Firms and the Future of AI Competition

In the rapidly evolving landscape of artificial intelligence, a new challenge has emerged: the distillation of capabilities from large language models.

TL;DR
• Chinese AI firms accused of distilling capabilities from Anthropic's Claude model.
• National security risks and AI export control concerns raised.
• Chinese AI companies pursuing performance gains under compute constraints.

Summary
The AI industry is witnessing a significant shift in competition, with a focus on access control, API governance, and data sovereignty. Anthropic has accused Chinese firms DeepSeek, Moonshot AI, and MiniMax of systematically extracting capabilities from its Claude large language model through large-scale distillation, aiming to accelerate training of their own systems.

Key points include:
• Anthropic reported that the three firms used proxy services and fabricated accounts to flood Claude with engineered prompts in bulk, harvesting high-quality responses to strengthen their own models.
• The alleged methods varied by actor, with DeepSeek conducting over 150,000 conversations, Moonshot AI conducting over 3.4 million conversations, and MiniMax conducting 13 million conversations.
• Anthropic responded by activating blocking and detection mechanisms, tightening API traffic monitoring, and behavioral analysis.

The dispute signals a broader shift in large language model competition, with implications reaching beyond a terms-of-service dispute. Industry observers note that distillation is a common practice, but the question remains whether this activity crossed legal or contractual lines.

The developments illustrate how Chinese AI companies are pursuing performance gains under compute constraints. Moonshot AI, for instance, raised over $1.2 billion in funding and reportedly hit revenue in under 20 days that exceeded its full-year 2025 total. Other Chinese developers, such as Zhipu, are pressing ahead with technical upgrades, launching models with improvements to agentic, reasoning, and coding capabilities.

The fallout extends well beyond model training, potentially reshaping demand for advanced AI chips and affecting procurement cycles for high-end GPUs and accelerators. Cloud providers face pressure to implement stronger API traffic monitoring and anomaly detection to prevent platform misuse. Taiwanese suppliers are at the crossroads of the US-China tech rivalry, securing orders for AI servers, accelerator cards, and data center infrastructure now comes with tighter customer due diligence and end-use compliance demands.

As the AI competition deepens, the line between regulatory compliance and global market ambition will define the structure and direction of the AI supply chain.