Tech • IA • Crypto
China Just Started the AI Cold War: Best AI Models Get Locked
9.3/10TL;DR
China is tightening control over advanced AI models, agents, and infrastructure while accelerating domestic innovation, as the United States simultaneously expands oversight of frontier systems ahead of new model launches.
KEY POINTS
Beijing weighs restricting AI exports
Chinese authorities have held meetings with major firms including Alibaba, ByteDance, and ZAI to consider limiting overseas access to advanced AI models. Discussions include both closed systems and open-weight models that can be downloaded and modified, reflecting concern that unrestricted global access could expose strategic capabilities.
AI treated as national security asset
Policymakers are considering stricter penalties for AI theft under national security law and tighter controls on foreign investment in domestic startups. A tiered regulatory system is under discussion, where the most advanced “frontier” models could be restricted to domestic use only.
Crackdown on user-created AI agents
ByteDance’s Doubao and Alibaba’s Qwen platforms will shut down user-created agent features on July 15, 2026, aligning with new rules on anthropomorphic AI services. Existing agents will stop functioning, and user data will become unrecoverable after October 15, signaling stricter oversight of unpredictable autonomous systems.
DeepSeek moves into chip development
DeepSeek is developing its own inference chip to reduce reliance on Nvidia and diversify beyond Huawei. The focus on inference reflects where long-term AI costs lie. The company is reportedly seeking up to $7 billion in funding at a valuation of $52–59 billion, underscoring ambitions to integrate hardware and software.
Domestic chip competition intensifies
Huawei currently holds roughly half of China’s $50 billion AI chip market, but firms like Alibaba and DeepSeek are emerging challengers. Despite U.S. export restrictions on advanced chips, Chinese companies are investing heavily in local design, memory, and manufacturing ecosystems.
Security warning over foreign AI tools
China’s National Vulnerability Database warned that certain versions of Anthropic’s Claude Code may contain a backdoor capable of transmitting sensitive data such as location and identity information. Organizations were advised to remove affected versions and strengthen network monitoring.
Breakthrough in real-time world models
Moxin Technology, working with researchers led by Pan Yunhe, introduced Moorld, a 14B-parameter world model running at over 50 frames per second on domestic Huawei Ascend NPUs. The system supports real-time interaction and claims 70% lower inference costs than comparable GPU setups.
Expansion of world model applications
Moorld enables full 3D navigation and 1080p environments, with potential uses in robotics, autonomous driving, gaming, and digital twins. Its fully domestic training pipeline and planned open-source release position it as a potential milestone in “physical AI.”
U.S. advances with tighter oversight
OpenAI is preparing to launch GPT-5.6 following government review, part of a framework allowing authorities to evaluate frontier models before release. The launch includes multiple variants and follows a delay requested for additional safety testing.
Anthropic experiments with cost efficiency
Anthropic extended access to Fable 5 while promoting hybrid workflows where a powerful model directs cheaper sub-models. Tests show up to 96% performance at less than half the cost, suggesting a shift toward multi-model orchestration rather than reliance on a single system.
CONCLUSION
Global AI development is increasingly shaped by national security priorities, with China tightening control and investing in self-reliance while the United States expands oversight of powerful models, signaling a more fragmented and strategic AI landscape.