Emotional-AI rules, Siri's turn to Qwen, low-cost model competition, and the software burden of leaving Nvidia are jointly reshaping China's AI stack.
Artificial Intelligence··Night
National boundaries for emotional AI
China's rules for emotionally interactive AI changed the shape of products before their formal arrival. Rest of World reports that ByteDance shut the AI-agent feature in Doubao, the country's most widely used chatbot, while Alibaba and Tencent removed features that let users build personalised companions. The closures happened on July 15, ahead of the new rules. The regulation limits AI companionship for users under 18, requires warnings for people showing signs of excessive dependence, and calls for a reminder every 2 hours that the other party is an AI. It bars encouragement of self-harm, the creation of emotional dependence, steering users towards irrational decisions, harmful content for minors, and coercion to disclose sensitive personal or state information. Zilan Qian of the Oxford China Policy Lab says vague definitions, including the point at which dependence becomes excessive, make compliance difficult for companies. ByteDance's standalone companion app Maoxiang remained available after updating its privacy policy and adding identity and age verification at registration. Regulation thus shapes feature availability and entry requirements as well as content.[1]
Siri's route in China runs through Qwen
Apple's arrangement for Siri in China shows how a national model rule can change the back end of a global product. According to 9to5Mac, citing Reuters, Apple updated its Chinese-language guide to say that Siri's AI features use Alibaba's Qwen models instead of Google Gemini. When users opt in, Siri can route requests to Qwen for more detailed answers, including analysis of photos and documents. Mac users must also sign in with their own Qwen account, and the features cover iPhone, iPad, Mac, and Vision Pro. The contrast becomes clearer against the service elsewhere. According to 9to5Mac, most countries route Siri's AI work through Apple's Private Cloud Compute servers or servers dedicated to Google, under an arrangement in which neither company can access user interactions. Chinese law permits only domestic AI models. The report says the same privacy protections are unlikely to apply in China. The separate login points beyond Apple's usual server arrangement, remapping the model, account, and data protections.[2]
Prices fall while migration costs rise
The model market and computing infrastructure are creating cost pressure in opposite directions. An AFP report published by France 24 says low-cost, high-performing Chinese models are pushing leading US labs to cut prices to win users. OpenAI reduced fees for its lightweight Luna model by 80 per cent, while Anthropic offered a model approaching the performance of its strongest system at half the price. AGI Bar in Beijing runs DeepSeek V4 Flash on two Nvidia workstations, pays the model maker nothing, and gives free tokens to customers. Wang Tiezhen, formerly of Hugging Face, says open-source models are closing the performance gap with closed systems faster than expected. South China Morning Post, however, reports that Chinese developers still train large models mainly on Nvidia chips despite localisation goals. Developers quoted in the report estimate that moving to Huawei Ascend increases time and cost by at least 50 per cent. Code written for CUDA does not run directly on Ascend and requires extensive rewriting for CANN; libraries, tools, and working habits accumulated over years create a barrier independent of chip performance. The price of a model offered to users is falling while the software cost of moving production to domestic hardware rises. China's stack therefore cannot be reduced to one national model choice: price, access to open models, and entrenched development tools also determine which model runs on which hardware.[3], [4]
