Shares of the Chinese AI startup surged after it reported a reduction in losses and a significant increase in revenue for the first half of this year, driven by its renewed focus on enterprise customers. The company’s stock climbed 3.6 percent to HKD314 (roughly $40.07) during midday trading in Hong Kong today, after earlier jumping by as much as 7.3 percent.
In its recent financial results, the Shanghai-based firm revealed that its net loss was reduced by 11 percent to $358 million for the six months ending June 30, compared to the previous year. Revenue skyrocketed by 283 percent to approximately $120 million, compared to $79 million for all of last year.
The revenue from open platform and other AI business services increased more than sevenfold to $73.9 million, becoming the company’s leading income source. This growth was largely fueled by an increasing number of paying enterprise clients, higher API call volumes, and rapid adoption of TokenPlan, the company stated.
As of this month, annual recurring revenue (ARR) surpassed $800 million, according to Yan Junjie, the company’s founder and CEO, during an earnings call. While ARR is an estimation based on existing business activity and isn’t recognized as actual income, it indicates an acceleration in commercializing their offerings.
In the first half of the year, approximately 80 percent of ARR came from business-to-business operations, with the remaining 20 percent from consumer-facing services. This marks a dramatic shift from the same period a year ago, when B2C made up around 70 percent and B2B only 30 percent. The company successfully transitioned its revenue model from consumer-oriented to primarily enterprise and developer-driven within just a year.
International revenue reached over $70.8 million, representing 61 percent of the total and remaining a key pillar of the company’s income.
Despite this rapid revenue growth, losses still persisted, mainly due to high costs associated with ongoing model training, product launches, and infrastructure development. MiniMax manages its computing resources dynamically, allocating more power to models in advanced stages of development, based on business needs and expected returns. Text-based models are prioritized, receiving roughly four times more training resources than video models.
The company is also expanding its computing capacity by integrating support for domestic chips alongside its M3 and H3 models, with a large local data center set to launch soon. Efforts are underway to further reduce per-token inference costs, which is expected to improve affordability and increase user adoption.
Yan emphasized that the goal for the M3.1 model is to reduce inference costs to about one-third of the original levels seen when M3 was first introduced. Such cost reductions plan to enable price adjustments that will lower barriers for users, attract more customers, and handle larger token volumes. This, in turn, is anticipated to drive ongoing improvements in gross margins.