(Yicai) Aug. 27 -- MiniMax’s shares jumped after the Chinese artificial intelligence startup shrank its loss and grew revenue almost four times in the first half of this year after pivoting to enterprise clients.

MiniMax [HKG: 0100] closed 3.8 percent higher at HKD314.60 (USD40.13) per share in Hong Kong today, after surging by as much as 7.3 percent at the market open. The stock was listed in January and has almost doubled in value from its offering price of HKD165.

The net loss narrowed 11 percent to USD358 million in the six months ended June 30 from a year earlier, the Shanghai-based company said in a financial report released yesterday. Revenue surged 283 to USD120 million. By comparison, its income was USD79 million for the whole of last year.

Income from open platform and other AI enterprise services soared more than seven times to USD73.9 million, becoming MiniMax’s top revenue stream, with the growth primarily driven by a growing number of paid and enterprise clients, higher application programming interface call volumes, and rapid adoption of TokenPlan, the company said.

Annual recurring revenue topped USD800 million as of this month, Yan Junjie, founder and chief executive, said at an earnings conference call. While ARR is a projected annual revenue metric based on existing business volume and does not equate to recognized income, the figure signals accelerating commercialization.

Business-to-business accounted for about 80 percent of ARR in the first half, with consumer-facing operations making up the rest, which is a major swivel from the same period a year ago, when B2B contributed around 30 percent and B2C the rest, Yan pointed out.

In the course of just one year, MiniMax’s revenue mix has shifted from being mainly consumer-led to being dominated by enterprise and developer offerings, he noted.

Overseas income topped USD70.8 million, accounting for 61 percent of the total and remaining a core revenue pillar.

However, top-line growth did not erase the entire loss. For AI developers in the model training , product rollout, and infrastructure buildout phase, computing power costs remain a major burden. 

MiniMax dynamically allocates compute resources based on model maturity, business needs, and expected marginal returns, Yan said, adding that text models are the highest priority for compute investment and receive around four times more training resources than video models.

In addition, MiniMax is seeking additional sources of computing power, Yan pointed out. Both the M3 and H3 models are being adapted to run on China-made chips, with a large domestic computing cluster set to go online soon, while the company also aims to further cut per-token inference costs, he said.

The target for M3.1 is to bring inference costs down to about a third of the level when M3 was first launched, Yan stressed. Cost reductions will enable price adjustments that will lower usage barriers, bringing in more users and larger token volumes, likely leading to ongoing improvements in gross margins, he added.

Editor: Martin Kadiev