ByteDance Bans AI Model Distillation From US Systems as Anthropic Moves Toward Custom Claude Chip
China’s artificial intelligence industry is becoming increasingly cautious about one of the most controversial practices in AI development: model distillation. After recent accusations involving Chinese AI lab Moonshot and its Kimi K3 model, ByteDance has reportedly taken a firm internal stance by banning employees from distilling US-developed AI models.
The move highlights the growing tension between Chinese AI companies and Washington, especially as the race to build more powerful large language models becomes tied to national security, semiconductor access, and the future of major platforms like TikTok.
Model distillation has become a major concern in the AI race
Model distillation is a technique where one AI model is trained to imitate the outputs or behavior of another, often larger or more advanced model. In legitimate settings, it can help developers create smaller, faster, and cheaper AI systems. But in the current US-China AI rivalry, it has become a politically sensitive issue.
US officials and leading AI companies have raised concerns that Chinese labs may be using distillation to quickly close the performance gap with top American frontier models. The concern is that even if China faces restrictions on advanced chips, developers could still use indirect methods to capture the capabilities of leading US models and transfer them into their own systems.
Moonshot’s Kimi K3 model recently became a flashpoint in this debate after allegations emerged that it had been distilled from Anthropic’s Fable model. Some officials also claimed that Moonshot had access to advanced NVIDIA GB300 servers and may have obtained additional GPU resources through Thailand.
These accusations have intensified fears in Washington that AI capabilities can be transferred across borders in ways that are difficult to monitor or prevent.
ByteDance takes a cautious approach to avoid US backlash
ByteDance, the owner of TikTok, is now reportedly telling employees not to use distillation from US AI models as a shortcut for advancing its own systems. The decision appears to be driven not only by technical ethics and compliance concerns, but also by the company’s broader geopolitical exposure.
TikTok remains one of ByteDance’s most valuable assets, and the platform has faced repeated scrutiny in the United States. Any perception that ByteDance is using US AI models improperly could create fresh political pressure and potentially threaten the company’s business interests.
ByteDance founder Zhang Yiming reportedly told staff that the company should be prepared to give up short-term advantages in order to protect its long-term goals.
That message reflects a broader shift among Chinese technology firms. Rather than risk retaliation from the US government, major AI players appear to be putting stricter boundaries around how their teams interact with American AI systems.
Alibaba has also moved to reduce distillation concerns
ByteDance is not alone in taking a more defensive stance. Alibaba has reportedly restricted employee access to Anthropic’s Claude in an effort to reduce concerns that its Qwen AI model could be accused of benefiting from distillation.
This suggests that China’s largest technology companies are becoming more aware of the risks surrounding AI training methods. While companies want to remain competitive, they also need to avoid actions that could trigger sanctions, investigations, or restrictions on international operations.
For Chinese AI labs, the challenge is clear: they must develop advanced models fast enough to compete globally, but without relying on methods that could be interpreted as copying or extracting capabilities from US frontier systems.
Anthropic is building custom silicon for Claude
While Chinese firms are dealing with distillation concerns, US AI companies are focusing on another major battleground: custom chips.
Anthropic has confirmed that it is working to build its own custom chip for Claude, its family of AI models. The company is assembling an in-house silicon design team and is pursuing what it describes as a multi-chip approach.
This move is significant because AI inference costs are becoming one of the biggest challenges for companies operating large-scale chatbots and enterprise AI services. Inference refers to the process of running AI models after they have been trained, such as generating answers, analyzing documents, writing code, or processing user prompts.
As AI usage grows, inference costs can become enormous. Custom chips, also known as ASICs, can help companies improve efficiency, reduce reliance on general-purpose GPUs, and gain more control over performance and supply chains.
Samsung and Broadcom are expected to play key roles
Anthropic’s custom chip effort is reportedly tied to its broader partnerships with Samsung and Broadcom. The company had previously made remarks about “logic chips,” sparking speculation that it was exploring dedicated AI silicon.
Now, Anthropic’s push into chip development looks more concrete. By designing hardware specifically suited for Claude workloads, the company could improve cost efficiency and scalability over time.
This strategy mirrors a wider trend across the AI industry. As demand for AI services grows, major companies are no longer relying solely on off-the-shelf GPUs. Instead, they are investing in tailor-made silicon that can handle their own models more efficiently.
OpenAI has also entered the custom chip race
Anthropic is not the only AI company moving in this direction. OpenAI has also unveiled plans for its own custom AI chip, reportedly developed with Broadcom. The chip, known as Jalapeño, is expected to see large-scale deployment by the end of 2026.
The push by OpenAI and Anthropic shows how important hardware has become in the AI race. The companies that control both advanced models and optimized infrastructure may gain a major long-term advantage.
Custom chips could allow AI firms to reduce operating costs, improve response speeds, and support more users without depending entirely on NVIDIA’s high-end GPUs, which remain in extremely high demand.
AI competition is moving beyond models
The latest developments show that the AI race is no longer just about who can build the smartest chatbot. It is also about who can train models legally and responsibly, who can access advanced chips, and who can scale AI services at the lowest cost.
ByteDance’s ban on distilling US models reflects growing anxiety among Chinese AI companies about political and regulatory consequences. At the same time, Anthropic’s custom chip plans show how US AI leaders are trying to build deeper control over their infrastructure.
The result is a rapidly changing AI landscape where software, chips, export controls, and geopolitics are becoming inseparable.
For ByteDance, avoiding distillation-related controversy may be essential to protecting TikTok and its global business. For Anthropic, building dedicated hardware for Claude could be key to competing at scale in the next phase of artificial intelligence.
