American AI coding agents are impressive. But so are China’s
Speakers
Tuhin Srivastava
— Co-founder and CEO, Baseten
Deirdre Bosa
— Anchor/Reporter, CNBC Tech Check
Summary
CNBC's Deirdre Bosa discusses the rapid rise of Chinese AI coding agents with Baseten CEO Tuhin Srivastava. The conversation focuses on Zhipu AI's GLM model and Alibaba's Qwen, and whether Chinese open-source models are closing the gap with US frontier labs. Srivastava explains how reasoning and agentic AI are driving a 100x increase in inference demand and massive compute buildout, and he remains positive on Nvidia while noting hardware diversification.
- Zhipu AI's coding agent saw viral demand, including from US developers, and the company went public in Hong Kong.
- Alibaba released Qwen 3 Max, claiming benchmark competitiveness with US frontier models.
- Tuhin Srivastava says Chinese open-source models are not a blip and are closing the gap.
- He highlights cost efficiency, tool calling, and on-prem/VPC deployment as enterprise advantages.
- Inference demand has grown 100x, driven by adoption, reasoning models, and agentic applications.
- US AI infrastructure spending is framed as justified by global demand and long-term value.
- Nvidia is seen as well-positioned with CUDA and inference-specialized architectures, though diversification is coming.
- Huawei chips, TPUs, and AMD are mentioned as alternative hardware paths.
Ideas
Inference growth drives massive compute demand
Inference demand is experiencing explosive growth—about 100x over the past year—driven by broader AI adoption, reasoning models that consume more compute, and agentic applications that require lots of back-and-forth with models. More inference inevitably leads to more compute, and the massive US data center buildout is justified by serving a global market and the potential for enormous long-term value creation.
Chinese open-source AI models are competitive
Chinese open-source AI models are not a blip. Multiple Chinese labs, including Alibaba’s Qwen, Zhipu’s GLM, DeepSeek, and Kimi, are producing models that are increasingly competitive with US closed-source frontier models, with a gap that is shrinking monthly. These models are very cost-efficient, strong at tool calling, customizable, and can be deployed inside enterprise VPCs for control and data privacy, driving real production adoption.
Zhipu AI's GLM is cost-efficient and competitive
Zhipu AI’s GLM models are very cost-efficient, excellent at tool calling, open source, and perform well on benchmarks. This makes them attractive to developers and enterprises, and the newer GLM 4.7 coding assistant is gaining global traction.
Nvidia remains key AI compute beneficiary
Nvidia is a key AI compute beneficiary with a deep CUDA moat and a long history of chip generations. Its recent Groq acquisition signals a move toward LPU-type architectures and inference-specialized chips, positioning it well for the next phase of AI hardware even as TPUs and AMD provide diversification.
