Open-weight models promoted by Chinese AI developers are facing worsening profitability despite drawing technical attention.

Business Insider reported on Monday that open-weight AI, unlike open-source software that releases code, has a structure in which inference costs rise as usage increases.

Conventional software has almost no additional distribution costs once built. AI, by contrast, continues to require semiconductors, power and data centre resources for each response. Arjun Bhatia (아르준 바티아), an analyst at William Blair, pointed out that it is not easy to find a path to monetise open-weight models.

Z.ai posted about $107 million in revenue and nearly $500 million in losses last year. It drew attention in the industry by releasing its open-weight model GLM 5.2 last month, but its shares have fallen more than 40 percent over the past month. MiniMax also recorded $79 million in revenue and $250 million in losses last year, and its shares have fallen more than 50 percent over the past month.

Open-weight models disclose trained parameters externally so other companies can download and run them. Actual operations are often handled by large cloud companies such as Amazon, Microsoft, Google, Oracle and Alibaba, or by specialised providers. Companies can also run models themselves. Stable revenue is generated only when model developers provide their own inference services, but Business Insider reported that Western corporate customers generally do not choose this approach due to data security concerns.

Moonshot AI is not free of these limitations. The company drew attention for strong performance after releasing its open-weight model Kimi K3 last week, but it halted new customer registrations within days of the launch due to a lack of computing resources.