Zhipu AI

Zhipu AI was founded in 2019 as a spin-off from Tsinghua University's Knowledge Engineering Group — one of China's most prestigious AI research centres. The founding team had spent years researching language models and knowledge graphs at Tsinghua, building expertise in natural language understanding that the commercial AI industry was only beginning to recognise as transformative. The company's name, Zhipu (智谱), translates roughly to "knowledge spectrum," reflecting its academic roots and ambitions. The founders — Tang Jie, Wang Tao, and Liu Xin — believed that large language models would become foundational infrastructure for enterprise software.

Zhipu's flagship model series, GLM (General Language Model), was developed as an alternative architecture to the Transformer models underlying GPT. GLM's design made it particularly efficient for bidirectional text understanding — strengths in both generating text and comprehending existing text simultaneously — which the team argued gave advantages in enterprise applications where accuracy and contextual understanding mattered more than raw generation speed. GLM models were published in research papers that attracted international academic attention, establishing Zhipu's credibility within the global AI research community even as geopolitical tensions were rising.

When ChatGPT launched in late 2022 and triggered a global AI race, Zhipu's research advantage translated into commercial momentum. ChatGLM — Zhipu's conversational AI product — launched in early 2023 and quickly became one of the most widely used Chinese-language AI assistants. Zhipu attracted investment from Alibaba, Tencent, Meituan, and Xiaomi — a coalition of China's most powerful technology companies that effectively endorsed Zhipu as a national AI champion. The valuation exceeded $3 billion, positioning Zhipu alongside DeepSeek and Baidu's Ernie as China's leading generalist AI model companies.

Zhipu scaled its enterprise customer base to over 10,000 Chinese companies, banks, and government agencies using its AI infrastructure for applications ranging from financial analysis to government document processing. The American export controls on advanced Nvidia chips — which restricted Zhipu's access to the most powerful training hardware — forced the company to develop techniques for training capable models with less compute, including expertise in model distillation and efficient training architectures. Researchers noted that Zhipu's GLM models consistently performed competitively on benchmarks despite operating under hardware constraints that American companies did not face.

Zhipu's valuation exceeded $3 billion by 2025, with the company providing AI infrastructure to institutions that could not or would not rely on American AI providers. The company's model-agnostic approach — offering both API access to its own models and tools for customers to fine-tune models on proprietary data — reflected the enterprise AI market's demand for data sovereignty and compliance with Chinese regulations. Zhipu had become what Zhipu's founders had envisioned from Tsinghua: the AI infrastructure layer for the Chinese enterprise economy, built by researchers who had studied the problem for a decade before most of the world understood what a large language model was.