DeepSeek, a Hangzhou-based AI firm, is gearing up to launch its V4 model in late April—a Mixture-of-Experts design with ~1 trillion total parameters but only 32-37 billion active per task. The release, delayed twice since February, could signal a geopolitical shift if trained entirely on domestic Huawei silicon, according to Carthage Capital's Stephen Wu. Meanwhile, the company posted rare on-site job listings for server maintenance in Ulanqab, Inner Mongolia.

Silicon Valley's OpenAI, Anthropic, and Google announced a united front on April 6 through the Frontier Model Forum to block Chinese model replication. Anthropic identified 16 million queries from 24,000 suspected fake accounts tied to three Chinese firms, alleging adversarial distillation tactics. OpenAI previously accused DeepSeek of model copying in a February 12 congressional memo.