Top AI Product

When Zhipu AI submitted its IPO hearing file in Hong Kong, the document looked, at first glance, like a familiar success story. Rapid revenue growth, strong positioning in China’s large language model market, and a long list of technical achievements. But once you sit with the financials for a while, the filing starts to feel less like a celebration and more like a stress test for the entire Chinese LLM industry.

Zhipu’s revenue growth is real. The company generated RMB 57 million in 2022, RMB 124 million in 2023, and RMB 312 million in 2024, with revenue in the first half of 2025 already reaching RMB 191 million. That is more than 130 percent compound annual growth over two years. By Chinese enterprise software standards, this is exceptional momentum, and it explains why Zhipu is now the largest independent general-purpose model provider in the domestic market by revenue.

But revenue is only half the story, and arguably the less important half.

Losses have expanded at a much faster pace. Net losses widened from RMB 144 million in 2022 to RMB 788 million in 2023, then surged to nearly RMB 3 billion in 2024. In just the first half of 2025, losses already exceeded RMB 2.3 billion. The filing makes it clear that these losses are not accidental or temporary. They are the direct result of sustained spending on model training, research teams, and compute infrastructure, with research and development costs alone reaching more than RMB 2.1 billion in 2024 and continuing to rise sharply in 2025 .

What stands out is not that Zhipu is losing money. Any frontier AI company will. What stands out is the ratio. In 2024, the company lost close to ten yuan for every yuan of revenue it generated. Even as revenue accelerates, the gap between income and cost is widening rather than narrowing.

This is where comparison with OpenAI becomes unavoidable.

OpenAI is also deeply unprofitable, but the structure of its losses looks very different. Public disclosures and partner reporting suggest OpenAI’s annual revenue is measured in billions of US dollars, driven by consumer subscriptions like ChatGPT Plus, enterprise plans, and a globally dominant API business. Its losses are large, but they sit on top of enormous demand and relatively strong unit economics. OpenAI is burning cash because it is scaling faster than supply efficiency can keep up.

Zhipu, by contrast, is burning cash while demand remains narrow and structurally constrained.

The difference comes down to how money is made. Zhipu’s revenue is still heavily enterprise-led, tied to private deployments, customized solutions, and institutional clients. These contracts are real and often politically or strategically important, but they do not scale cleanly. Margins are thin, sales cycles are long, and each new client often requires bespoke work. There is no consumer flywheel quietly generating high-margin recurring revenue at scale.

At the same time, compute costs in China remain a structural burden. Even when capacity is available, fragmentation and efficiency gaps mean that each incremental improvement in model capability carries a heavy cash penalty. As models grow stronger, the cost curve does not flatten fast enough to offset revenue growth.

None of this undermines Zhipu’s technical credibility. The models are competitive, the research is serious, and the team clearly understands the global frontier. But the IPO filing strips away a comforting assumption that many in the industry have held: that once Chinese models reach technical parity, economic parity will naturally follow.

The numbers suggest otherwise.

Zhipu’s IPO is not a warning about execution or ambition. It is a signal that China’s LLM industry has reached a turning point where capability is no longer the bottleneck. Efficiency is. Until Chinese large model companies find a way to turn strong models into scalable, high-margin products rather than capital-intensive projects, the gap with US leaders will persist, regardless of benchmark scores or leaderboard rankings.

The models are catching up. The business models are not.