China just mass released 10+ frontier AI models in 2 weeks and Western markets barely noticed

China's tech giants and AI startups dropped a coordinated wave of frontier models during Chinese New Year. Alibaba, Baidu, ByteDance, Tencent, plus startups like Moonshot AI and MiniMax all released major updates within days of each other.

The scale is wild. Baidu's ERNIE 5.0 hits 2.4 trillion parameters. Moonshot's Kimi K2.5 is 1.04 trillion and fully open sourced. Over 700 generative AI services are now commercially deployed in China. Baidu alone has 200M+ monthly active users on their AI platform.

But the real story is pricing. These models are charging $0.05 to $0.15 per million tokens for API access. That's roughly 1/20th of comparable Western pricing. Alibaba's Qwen 3.5 activates only 4% of its parameters per inference through sparse architecture, which is how they're hitting those price points without sacrificing quality.

US chip export restrictions didn't slow them down. It forced algorithmic innovation instead. When Nvidia GPUs aren't available, teams optimize architecture. iFlytek's Spark X2 was trained entirely on domestic Chinese chips, proving the full stack works without Western hardware.

ByteDance's Seedance 2.0 video generator got deployed on China's Spring Festival Gala broadcast. Hollywood is already filing IP complaints. This isn't demo day stuff, it's production scale.

The infrastructure play looks interesting here. Every one of these models needs domestic compute. Cambricon and other AI chip names in holdings like CNQQ are positioned as the picks and shovels for this buildout. DeepSeek V4 is expected in early March with rumored trillion parameter multimodal capabilities, which should put more pressure on domestic chip supply.