American AI coding agents are impressive. But so are China’s

Key Concepts

• Jipu AI (GLM 4.7): A Chinese AI startup gaining traction with its coding assistant, trained on Huawei chips.
• Inference: The process of running a trained AI model to generate outputs from user inputs. Crucial for real-world AI application usage.
• Agentic Applications: AI systems capable of autonomously performing tasks, requiring significant back-and-forth interaction with models (and thus, high inference).
• Open Source vs. Closed Source Models: The debate surrounding accessibility, customization, and control in AI model development. Chinese models are largely open source.
• Compute: The computational power required to train and run AI models, a growing cost and bottleneck.
• VPC (Virtual Private Cloud): A logically isolated section of a public cloud where enterprises can deploy and run applications securely.
• TPU (Tensor Processing Unit): AI accelerator developed by Google.
• LPU (Language Processing Unit): A newer type of AI accelerator, exemplified by Nvidia’s Groq acquisition, designed for efficient language model inference.

The Rise of Chinese AI and the Coding Battleground

The discussion centers on the increasing competitiveness of Chinese AI models, particularly in the critical area of coding. Jipu AI, with its GLM 4.7 coding assistant, has experienced rapid adoption, notably in both China and the United States. This is significant because coding is identified as the “key battleground in AI” – the space where models transition from experimentation to practical, daily use. The viral adoption of Jipu, following DeepSeek’s earlier emergence, suggests that DeepSeek wasn’t an isolated incident but a harbinger of more sophisticated AI capabilities coming from Chinese labs. Jipu’s GLM 4.7 was notably trained entirely on Huawei chips, highlighting China’s growing self-sufficiency in AI hardware.

Inference Growth and the Cost of AI

Tuin Sui Vastiva of B10 highlights a 100x growth in inference over the past year, driven by increased user adoption and the development of more capable models. This surge in inference directly translates to increased demand for compute power. The conversation explores the question of who is footing the bill for this escalating compute cost. While individual users pay subscription fees (e.g., Anthropic Claude Max), it’s suggested that venture capital and large technology companies are likely subsidizing a significant portion of the costs, mirroring the early days of the gig economy. However, the value delivered by AI, particularly in enterprise applications like healthcare (citing OpenEvidence and Bridge), is argued to be exceeding the incremental cost of compute.

Open Source Advantage and Chinese Innovation

A key theme is the rise of open-source AI models originating from China. Five Chinese companies – Alibaba (Qwen 3 Max), Jipu (GLM), DeepSeek, Kimmy, and others – are now producing models competitive with those from US giants like OpenAI, Anthropic, and Google. While a raw ability gap still exists, it’s closing rapidly, with Chinese models potentially being only a few months behind their American counterparts. The open-source nature of these models offers advantages in terms of customization, control, and deployment within a company’s own infrastructure (VPCs). This is particularly appealing to enterprises concerned about data security and vendor lock-in. Tuin notes that the ability to fine-tune these models further narrows the performance gap.

Hardware Landscape and US vs. China

The discussion touches on the hardware landscape, acknowledging Nvidia’s dominance but also highlighting the emergence of alternatives like TPUs (Google) and AMD. China’s development of its own chips, as demonstrated by Jipu’s training of GLM 4.7 on Huawei chips, is seen as a significant step towards self-reliance. The conversation raises the question of why AI infrastructure buildout costs are so much higher in the US compared to China. Tuin suggests that the US serves a global market, requiring a larger and more extensive infrastructure. He also points to the excitement and potential surrounding AI as driving factors in the substantial investment. Nvidia’s recent acquisition of Groq, a company specializing in LPUs (Language Processing Units), signals a move towards more specialized hardware optimized for inference.

The Shift from Chatbots to AI-Powered Creation

Debbosa and Jasmine’s personal experience with AI tools underscores a broader trend: a shift from simple chatbot interactions to using AI for building applications. They successfully used Jipu to create an app tracking Hong Kong’s top companies by market cap, mirroring similar results achieved with American tools like Replet and Claude. This ease of application development suggests that AI is becoming increasingly commoditized, empowering even non-technical users to create functional tools.

Notable Quotes:

• Debbosa: “Coding is the key battleground in AI. It's the first place models turn into daily tools and where agents do real work.”
• Tuin Sui Vastiva: “There’s two separate exponents that are kind of multiplying against each other… one is just greater adoption of AI in general and the other one is you know these models of thinking and reasoning more.”
• Tuin Sui Vastiva: “I don’t think anyone can look at the current state of affairs and not see… project a year forward and be like oh I I can see where this is going.” (referring to Chinese AI advancements)

Technical Terms:

• Inference: The process of using a trained AI model to make predictions or generate outputs.
• VPC (Virtual Private Cloud): A secure, isolated environment within a public cloud.
• TPU (Tensor Processing Unit): A specialized AI accelerator developed by Google.
• LPU (Language Processing Unit): A specialized AI accelerator designed for language models.
• Fine-tuning: The process of further training a pre-trained AI model on a specific dataset to improve its performance on a particular task.
• Agentic Applications: AI systems capable of autonomous action and task completion.

Logical Connections:

The conversation flows logically from observing the rise of Jipu AI to exploring the broader implications for the AI landscape. It connects the increasing demand for compute to the growth of inference, the emergence of open-source models, and the hardware race between the US and China. The discussion consistently returns to the idea that coding is a crucial benchmark for AI progress and that Chinese models are rapidly closing the gap with their American counterparts.

Data and Statistics:

• 100x growth in inference over the past year (B10 data).
• $1.4 trillion estimated investment in US AI infrastructure (OpenAI alone).

Conclusion:

The conversation paints a picture of a rapidly evolving AI landscape where Chinese models are emerging as serious competitors, particularly in the critical area of coding. The open-source nature of these models, coupled with their increasing efficiency and cost-effectiveness, presents a compelling alternative to closed-source American models. While the US continues to invest heavily in AI infrastructure, the Chinese approach, characterized by innovation and a focus on open-source development, is proving to be a potent force. The shift from chatbots to AI-powered creation is accelerating, and the battle for dominance in AI is far from over.