In January 2026, Zhang Peng limped into a three-hour interview on crutches — torn Achilles tendon from a business trip. English has an idiom for the occasion: break a leg.
Days later he rang the bell at the Hong Kong Stock Exchange: HK$116.20 offer, day-one close HK$131.50. The HK$57.9 billion market cap crossed HK$100 billion in three days.
The world's first large-model IPO belonged to a man an investor once asked: "Can you even con money out of people with it? How about you cut your valuation in half?"
The more telling line:
"Scaling Law is just a description of a phenomenon. It doesn't have a solid scientific basis."
A Lab That Shipped
Zhang entered Tsinghua's computer science department in 1998 — B.S., M.S., 2018 engineering doctorate. At the KEG lab, his job was turning research into products.
The lab's advisor was Zhang Bo, a founding figure of Chinese AI. His quadrant chart — a machine must know what it doesn't know — became Zhipu's north star.
In 2018 a policy window opened for researcher commercialization. "We were the first to eat the crab," Zhang said. After a year and a half of negotiation, Beijing Zhipu Huazhang was registered — Tsinghua CS's first complete tech-commercialization case.
The One-Fourteenth Bet
When GPT-3 shipped in May 2020, Zhang asked Zhang Bo for a verdict: "Milestone-level advance" — but it still "doesn't know what it doesn't know."
Zhipu's response: GLM, fusing BERT's understanding with GPT's generation. Then the bet: a 100-billion-class model on a startup budget.
GPT-3's training cost was estimated near $20 million. GLM-130B, finished July 2022: roughly 4 million yuan of compute, 10 million total. One-fourteenth.
By Zhang's account, Stanford's evaluation ranked it the only Chinese model in the roundup.
March 2023: ChatGLM-6B — 6.2 billion parameters, 6GB of VRAM. No. 1 on GitHub trending, 13 million Hugging Face downloads.
For Zhipu, efficiency was never a constraint. It was the product.
The Hard Choices
The second half of 2022 was brutal. "They didn't get it," Zhang said of investors. "What is this? How do you make money?"
Then ChatGPT launched on November 30. "ChatGPT did us a huge favor. People stopped questioning what this thing was."
Zhang picked enterprise over consumer. "ChatGPT in the US can charge $20 a month, and tons of people pay. In China, that's impossible."
He picked open source too, then had to defend it: "Open source and free can't be equated. Commercial customers need products, tools, and services — not just the model itself."
That is 2025 in a sentence: DeepSeek open-sourced, and customers asked why Zhipu charges.
What Restraint Bought
The prospectus is the doctrine, audited.
Revenue: 57.4 million yuan in 2022 to 312.4 million in 2024 — 130% CAGR, gross margin above 50%. R&D: 1.595 billion yuan in H1 2025, 839.5% of revenue. Losses: 2.466 billion yuan in 2024.
By September 2025: 12,000 enterprise customers, nine of China's top ten internet companies.
Zhang keeps the IPO small in his vocabulary. "Funding money is just travel money." How should Zhipu be remembered? "A pioneer in the history of AGI — a trailblazer."
Fair. But the market priced the journey, not the destination: can efficiency reach AGI alone, or does someone's brute force get there first?
Top comments (0)