
Moonshot AI released its latest multimodal model with a 1-million-token context window and advanced agentic capabilities, closing the gap with top U.S. offerings on key benchmarks. The development added to pressure on Nasdaq 100 futures, which fell as much as 2.3%, and semiconductor names already reeling from valuation debates. Z.ai, formerly Zhipu AI, also advanced with its GLM-5.2 open-weight model that rivals or exceeds certain Western frontier systems on engineering tasks while running efficiently on domestic hardware.
DeepSeek continued its rapid gains with ongoing chip development efforts aimed at reducing reliance on Nvidia Corp. and Huawei Technologies Co. processors for inference workloads. Chinese firms surveyed by Bloomberg Intelligence plan to allocate 46% of AI accelerator budgets to domestic suppliers over the next 12 months, up from 30% currently, as Beijing pushes a nationwide data-center buildout valued at roughly 2 trillion yuan, or $295 billion, over five years. At least 80% of hardware and software in the project will come from local vendors.
U.S. tech shares faced broader headwinds as Microsoft Corp. Chief Executive Officer Satya Nadella publicly noted that leading models remain too expensive, with the company exploring lower-cost Chinese alternatives such as DeepSeek variants. Elon Musk predicted Chinese models could match certain Anthropic capabilities by early 2027, prompting swift rebuttals from Chinese lab founders claiming even faster timelines. The Philadelphia Semiconductor Index extended losses, with memory and foundry names posting sharp daily declines amid the rotation.
Despite headlines emphasizing a U.S. tech rout, Chinese AI advancements have driven measurable adoption by global developers and enterprises seeking cost efficiency. Six of the top 10 most-used AI models on certain developer platforms are now Chinese, with token usage from U.S. companies on platforms such as OpenRouter exceeding 30% for Chinese models in recent weeks and peaking near 46%. Domestic training runs, such as Meituan’s trillion-parameter LongCat-2.0 executed entirely on local accelerators, demonstrate progress in closing the full-stack gap without U.S. chips.
Moonshot AI’s flagship model supports up to 128,000 output tokens and is scheduled for open weights release by late July, details often omitted in broad selloff coverage. Z.ai’s GLM-5.2 operates under an MIT license with no restrictions, enabling rapid global fine-tuning and commercialization. Chinese AI labs have quietly achieved breakthroughs in embodied and video-action models, with Ant Group releasing an industry-first dual-stream mixture-of-transformers architecture for causal video prediction and policy execution. Beijing’s discussions with Alibaba Group Holding Ltd., ByteDance Ltd. and Z.ai on potential restrictions for overseas access to certain models remain at an exploratory stage.









