China’s Moonshot AI pushes boundaries with cost-efficient model development amid US Chip restrictions

Beijing-based Moonshot AI is advancing its artificial intelligence capabilities using older Nvidia GPUs, showcasing China’s resilience and innovation in AI development despite US export curbs and the global race for computing power dominance.

Chinese artificial intelligence start-up Moonshot AI is making notable progress in developing advanced AI models despite limited access to high-end graphics processing units (GPUs) compared to its American counterparts. During a Reddit “ask me anything” session, company representatives revealed that the firm continues to build sophisticated models using fewer and older chips.

A Moonshot AI executive, posting under the handle “ppwwyyxx” – the same alias used by co-founder Wu Yuxin on X (formerly Twitter) – acknowledged that the company is “outnumbered” by US competitors when it comes to high-end GPUs. He confirmed that the firm’s latest reasoning model, Kimi K2 Thinking, was trained using Nvidia’s older H800 GPUs, which were banned from export to China in late 2023.

Innovation amid constraints and cost efficiency

Moonshot AI’s use of older chips reflects the adaptability of Chinese AI firms that continue to innovate under stringent US technology export restrictions. The company’s latest model, unveiled last week, has drawn significant attention within the AI community for its performance and cost efficiency.

Backed by major Chinese tech giants such as Alibaba Group Holding and Tencent Holdings, Moonshot AI — valued at around US$3.3 billion — has sparked discussion about a potential new “DeepSeek moment” in global AI progress. During the Reddit discussion, co-founder Yang Zhilin, under the handle “ComfortableAsk4494,” referenced OpenAI’s large-scale infrastructure plans, saying their next-generation model, the K3, would arrive “before Sam’s trillion-dollar data centre is built.”

Open-source excellence gains global attention

Co-founder Zhou Xinyu dismissed comparisons to OpenAI’s massive spending, emphasizing that Moonshot AI follows its own “way and pace.” While CNBC estimated the cost of training Kimi K2 Thinking at US$4.6 million, Yang clarified that this was not an official figure, noting that much of the expenditure went into research and experimentation.

Despite limited resources, Kimi K2 Thinking has earned recognition from the global AI community. Thomas Wolf, co-founder of Hugging Face, praised the model as another open-source achievement delivering “industry-leading performance.”

As Western firms like OpenAI and Meta continue to invest billions in GPU-intensive projects, Moonshot AI’s approach underscores how Chinese innovators are redefining AI advancement — achieving more with less amid global hardware constraints.