China's AI startups close gap with US as Z.AI hits $1B ARR

Chinese AI startups have closed the capability gap with US frontier labs to months, with Z.AI's GLM-5.3 scoring 84.5 percent on CyberGym versus Anthropic's Mythos 5 at 83.8 percent.

"Won't take that long," Tang Jie, co-founder of Z.AI, said on X in response to Elon Musk's prediction that China wouldn't match Anthropic's top model until the first quarter of 2027.

Z.AI, which went public on the Hong Kong Stock Exchange in January, reached $1 billion in annual recurring revenue by July, about twice that of DeepSeek. The company's GLM-5.3 model, with 743 billion parameters, trailed Mythos 5 on exploit development — scoring 54.4 percent on ExploitBench versus 78.0 percent — but Z.AI says its weights will be published by the end of August.

The rise of Chinese AI labs threatens the market positions of OpenAI and Anthropic, which command billions in revenue — Anthropic reached $65 billion in annual revenue in July — while Chinese firms operate on less than one-fifth of the capital and chip access of their US counterparts.

The Tsinghua Pipeline

Tang, a professor at Tsinghua University, spun Z.AI out of his lab in 2019, funding it in part with a data-analytics platform that counted Google and IBM among its clients. His former student Yang Zhilin, who earned a PhD at Carnegie Mellon, leads Moonshot AI, valued at tens of billions of dollars. Both companies trace their roots to Tsinghua's elite "Yao class," established in 2005 by Turing Award winner Andrew Yao to cultivate top-tier computer science talent.

The Tsinghua network extends further. Two more former Tsinghua researchers started AI model companies in Shanghai, and DeepSeek's founder Liang Wenfeng — who consulted with Tang before spinning out from his hedge fund in 2023 — studied in Hangzhou. Nearly half of total equity-capital investment in China flowed into AI in the first half of this year, most of it from government-backed funds, according to the Investment Association of Central State-Owned Enterprises.

More With Less

DeepSeek's January 2025 release of a powerful open-source model — the "DeepSeek shock" that briefly rattled US stock markets — demonstrated the efficiency gains Chinese labs have achieved. The company's multihead latent attention technique slashes memory usage by creating a shorthand version of conversational context, while its mixture-of-experts design routes problems to specialized sub-models, reducing chip demands. Moonshot incorporated both techniques into its K2 and K3 models, and DeepSeek in turn adopted a Moonshot-optimized training method.

Chinese labs operate under severe constraints. Researchers at top labs including Alibaba and Z.AI said they are often allotted one-fifth of the high-end chips available to peers at OpenAI and Google. Jefferies calculated that Chinese tech companies invested less than one-fifth of what US counterparts spent on AI during these years.

The efficiency gap shows in model economics. Z.AI reports GLM-5.3 reaching 31.4 percent on its internal coding benchmark at roughly 50,000 output tokens per task, against Anthropic's Opus 4.8 at 29.5 percent using 120,000 tokens — slightly better work for less than half the tokens.

Z.AI's GLM-5.3 release generated headlines for its cybersecurity capabilities, but the underlying data tells a more nuanced story. The CyberGym result — a single run across 1,507 tasks with no variance figures — represents a seven-tenths-of-a-point margin that Z.AI itself acknowledges narrows on harder tests. On ExploitBench, which requires converting discovered flaws into working attacks, GLM-5.3 scored 54.4 percent versus 78.0 percent for Mythos 5 and 76.5 percent for OpenAI's GPT-5.6 Sol. In timed ExploitGym tests, GLM-5.3 completed 105 attack-development tasks in two hours and 130 in six, compared with 181 and 247 for Mythos 5.

Z.AI also reported working with Chinese security teams to identify 2,436 vulnerabilities across 269 open-source projects, with 53 publicly disclosed and 2,383 under embargo. The company did not disclose how many findings were previously unknown or independently reproduced.

The model's distribution strategy adds another layer. Z.AI says GLM-5.3's weights will be published once safety evaluation and hardening are completed, with the release expected by the end of August. Its most sensitive cybersecurity functions will only be available to verified users. Snowflake has announced plans to add GLM-5.3 and DeepSeek V4 Flash to its Cortex AI catalog, alongside models from Anthropic, OpenAI, Google, and Meta.

For investors, the competitive picture is shifting. Z.AI's $1 billion ARR and successful Hong Kong IPO — the first by a Chinese AI model startup — demonstrate growing commercial traction, while Anthropic's $65 billion revenue run rate shows the scale gap remains wide. Chinese labs' cost-efficiency advantages, combined with open-weight distribution, could pressure US frontier model pricing as enterprises adopt dynamic routing to match tasks to cheaper models. Z.AI is already training its next flagship model, which the company hopes will execute complex research tasks stretching over weeks.

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