Zhipu AI hits HK$1 trillion market cap on coding model surge

Zhipu AI's pivot to coding models turned a money-losing B2B deployment business into China's first $1 billion ARR AI company, with its GLM 5.3 release on Aug. 14 tying Kimi K3 for open-source SOTA.

"It's like a Tsinghua engineering guy — very smart, very capable. Whatever you ask him to do, he'll do it brilliantly, but there's not much emotional value," Zhang Peng, Zhipu's CEO, said.

The numbers tell the story. Zhipu's ARR climbed from $500-600 million in May to $1 billion by July, nearly doubling in two months after GLM 5.2 achieved open-source SOTA on the Artificial Analysis composite leaderboard. The stock surged 32 percent on the first trading day after that release and reached a HK$1 trillion market cap within a week — roughly 2.5 times Meituan and 3.5 times JD.com. GLM 5.2's call volume growth was the fastest of any model in 2026, surpassing DeepSeek V4, according to cloud deployment platform Vercel.

The valuation carries heavy expectations. Zhipu trades at more than five times Anthropic's price-to-sales ratio, and to stabilize at HK$1 trillion, its ARR would need to reach approximately $6.2 billion — six times current levels. The company has raised its year-end ARR target to $2.5 billion, but the gap between current revenue and the implied valuation remains wide.

The May 2025 Pivot That Changed Everything

The turning point came in May 2025, when Zhipu convened an emergency strategy meeting with only core executives and a handful of shareholders. DeepSeek R1, released during the 2025 Spring Festival, had undercut the entire B2B market — matching OpenAI o1's performance at one-thirtieth the price. Zhipu lost nearly 30 percent of its customers to DeepSeek, according to company employees.

The decision was to abandon separate vertical models for text, multimodal, and coding capabilities, and instead bet on a single large-parameter model fusing Reasoning, Coding, and Agentic data. Shareholders opposed the move because of the cost — Zhipu's 2025 financial report showed a net loss of 4.718 billion yuan ($702.9 million), with R&D alone consuming 3.182 billion yuan ($474 million), four times total revenue.

The resulting GLM 4.5, launched in July 2025 after a three-month delay, was trained on 15 trillion tokens of general data plus 8 trillion tokens of Coding, Reasoning, and Agentic data — roughly 1.5 times the training data of comparable models. It became Zhipu's first model with a strong coding reputation, approaching Claude Sonnet 4's capability at one-seventh the price. The "Claude substitute" label began circulating on social media.

The Talent Moat Behind the "No Mistakes" System

Zhipu's advantage extends beyond model architecture. More than 70 percent of the several hundred researchers at its AI Institute are Tsinghua University computer science graduates, creating a web of mentor-student and peer relationships that competitors cannot replicate. Executive search headhunters say core algorithm researchers have rejected offers worth tens of millions of yuan, with stock options now valued at over a billion yuan each.

The company's disciplined focus shows in its resource allocation. The multimodal model training team has been reduced to around 10 people, while the text and coding team numbers in the hundreds. Zhipu built a meticulous A/B testing system for technical decisions and invested heavily in data strategy and post-training — areas where experience from training the trillion-parameter WuDao 2.0 model in 2021 provides an edge.

The competitive race is far from settled. Moonshot AI's K3, released July 17, displaced GLM 5.2 from the open-source SOTA position it held for a month; GLM 5.3 reclaimed the title just over 20 days later. Moonshot, DeepSeek, and Alibaba have all released new flagship models with parameter counts two to four times that of GLM 5.3, betting on pre-training scale while Zhipu focuses on post-training refinement.

For investors, the key question is whether Zhipu's MaaS revenue can close the gap to its valuation. The company's year-end ARR target of $2.5 billion would still leave it far from the $6.2 billion needed to justify the current market cap at Anthropic's price-to-sales multiple. Zhipu's stock, which has already priced in significant growth, faces a binary outcome: continued ARR acceleration or a valuation correction.

This article is for informational purposes only and does not constitute investment advice.