According to a report by LatePost cited by BlockBeats on Aug. 10, Zhipu’s MaaS open platform has nearly 7 million registered users, about 2 million more than in early July. The number of enterprise customers has reached 23,000.
ZCode, the company’s developer product, passed 1 million users within one month of launch. The report said Zhipu’s annual recurring revenue, or ARR, has grown about 15 times so far this year. Market sources said the company’s ARR may now be at $2 billion, though that figure has not been officially confirmed.
Domestic compute expansion as inference demand rises
As demand for model calls climbs, Zhipu is expanding its domestic AI computing infrastructure. The report said the company has put more than 50,000 domestic AI compute chips into service to ease growing inference demand. Earlier market talk had suggested that Zhipu had already built domestic AI computing infrastructure at a scale of 1GW.
Coding demand drives growth
LatePost said Zhipu’s growth has been driven mainly by a surge in demand for coding use cases. After the release of the GLM-5 model, the company’s ARR rose quickly, and its API business has now become its core revenue source.
Zhipu is also working on inference efficiency to improve compute utilization, including through KV cache splitting and inference network architecture upgrades.
Pricing changes and model competition
In July, Zhipu launched a high-speed API that can generate 400 tokens per second, around 8 times faster than its standard API. On July 31, the company reopened purchases for its Coding Plan and sharply adjusted pricing, with the Lite version’s monthly fee raised to 118 yuan.
Industry figures cited in the report said that as AI coding demand keeps growing, inference efficiency will become a key competitive edge for large-model companies. In the same amount of computing power, companies that can generate more tokens will have an advantage in both cost control and commercialization.
Zhipu, Kimi and DeepSeek, among other Chinese large-model companies, are all facing compute bottlenecks and model competition pressure. The market expects several companies to launch next-generation models in August, setting up a more intense race in the sector.