Zhipu AI

Zhipu AI (智谱AI, also operating under the brand Z.ai) was founded in Beijing in 2019 as a commercial spin-out of Tsinghua University's Knowledge Engineering Group (KEG), a research lab long focused on knowledge graphs and, later, large language models. Its founders, Tsinghua professors Tang Jie and Li Juanzi, retained close ties to the university's research pipeline even as the company commercialized; Zhang Peng serves as CEO and Liu Debing as chairman. That academic pedigree distinguishes Zhipu from commercially-founded rivals such as Moonshot AI or the hedge-fund-spun-out DeepSeek — Zhipu's GLM (General Language Model) architecture traces its lineage directly to KEG's pre-ChatGPT-era language modeling research, giving the company a multi-year head start on Chinese-language pretraining data and evaluation infrastructure relative to newer entrants. Zhipu is frequently grouped with Moonshot, MiniMax, and Baichuan as one of China's so-called "AI tiger" startups, the cohort of well-funded, model-focused labs that emerged in the wake of ChatGPT's launch, though Zhipu's academic roots and subsequent public listing set it apart from that cohort's more purely commercial members. In January 2025, the US Bureau of Industry and Security (BIS) added Zhipu to the Entity List, citing its role in advancing China's military modernization. The listing cut off Zhipu's access to Nvidia H100 and H200 GPUs and any other US-origin technology absent an individual license, forcing an abrupt pivot in its hardware roadmap. Rather than attempt to route around the restriction, Zhipu committed to training its flagship models entirely on domestic silicon. The clearest demonstration came with GLM-5, released February 11, 2026: a 745-billion-parameter model trained from scratch on roughly 100,000 Huawei Ascend 910B accelerators using Huawei's MindSpore framework, with no Nvidia hardware involved at any stage of training — one of the first public proof points that a frontier-class Chinese model could be built entirely outside the CUDA/Nvidia ecosystem. Zhipu's GLM family spans the GLM-4 and GLM-4.6 generations released in 2024-2025, GLM-5 in February 2026, and GLM-5.2, released in mid-June 2026, alongside CodeGeeX, its dedicated code-generation model. GLM-5.2 narrowed the benchmark gap with the leading overseas closed models to roughly 1-4% on several evaluation suites, a result JPMorgan cited when it raised its 2026-2030 revenue forecast for Zhipu by 7-16%. The GLM line is released with permissive open-weight licenses for most versions, which has made it one of the most widely deployed open Chinese model families among both domestic enterprises and international developers building on Hugging Face and similar platforms. Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, under the legal entity name Knowledge Atlas Technology (ticker 2513.HK), becoming the first Chinese frontier AI foundation-model company to go public. The IPO raised approximately $558 million at an offer price of HK$116.20 per share, and the stock closed its debut day up roughly 13%. Zhipu has since conducted a further placement of new shares raising roughly $4 billion. The stock's subsequent run has been extraordinary. By early June 2026, shares were up nearly 1,000% year-to-date, minting chairman Liu Debing a fortune of roughly $22.4 billion and making him one of the richest people in China created directly by the AI boom. Following the GLM-5.2 release, Zhipu's market capitalization surged further, briefly spiking as much as 42% in a single session before closing up about 15% and crossing HK$1 trillion (roughly $128 billion) on June 22, 2026 — a year-to-date gain that some reports place above 2,400%. Zhipu counts Alibaba, Tencent, Meituan, and Xiaomi among its investors and customers, and reportedly serves more than 100,000 enterprise clients. Zhipu's Entity Listing, which looked like a severe constraint in January 2025, has instead become the foundation of its market narrative: a Chinese AI lab that built a genuinely competitive frontier model on domestic Huawei Ascend compute, insulated from any future tightening of Nvidia export rules in a way that H100/H800-dependent rivals are not. That insulation is not free — Ascend's per-chip performance and interconnect bandwidth still lag Nvidia's Blackwell-generation parts, and MindSpore's software ecosystem remains thinner than CUDA's — but Zhipu's results suggest the gap is closing faster than US policymakers may have anticipated when they wrote the Entity List rules. As the company works to extend a stock rally now among the largest wealth-creation events of the current China AI cycle, its central bet is that a compute roadmap free of American input controls, however performance-constrained today, is the only durable long-term strategy for a Chinese frontier lab.