MiniMax Defies Expectations with 700% Enterprise Revenue Surge, Quietly Becoming the AI 'Shovel Seller'
In the minds of many, when MiniMax is mentioned, the first things that come to mind are perhaps still Talkie, the AI companion that once went viral overseas, or Hailuo AI and Xingye, which are well-known to domestic users. For a long time, this large model unicorn was often labeled by the outside world as a "consumer-facing company.
The situation has undergone earth-shattering changes over the past year.
MiniMax's latest 2026 mid-year financial report shows that, compared to a year ago when the consumer side accounted for about 70% and far led the enterprise side's roughly 30%, its revenue structure has completely inverted, presenting a situation where the B2B side holds absolute dominance.
MiniMax took a year to complete a structural migration and has now become a computing infrastructure company driven by a global developer ecosystem, multimodal MaaS, and Agent infrastructure.
According to disclosed data, MiniMax's overall revenue scale has achieved a massive increase, realizing a total revenue of approximately $120 million in the first half of 2026, a year-on-year increase of 283.1%. Its half-year revenue alone reached 1.
5 times its total revenue for the entire year of 2025.
Among this, B2B business revenue, primarily consisting of the open platform and other enterprise services, reached as high as $73.9 million, a year-on-year surge of over 700%. Its proportion of total revenue skyrocketed from 30.
3% in the same period last year to 63.4%, becoming the company's largest revenue pillar in one fell swoop.
The roaring growth of the enterprise market has brought a significant increase to MiniMax's ARR (Annual Recurring Revenue), reaching $800 million as of August, with the To B business accounting for about 80% and To C only about 20%.
(It is worth noting that ARR is not an audited figure under Generally Accepted Accounting Principles and inherently contains some fluidity, but it remains a valuable consideration metric for AI model and SaaS companies.)
In addition, overseas markets contributed about 60% of MiniMax's revenue in the first half of the year, once again highlighting its globalized nature.
Alongside the revenue surge, sales and distribution expenses decreased by 17.9% year-on-year to $27 million, while gross profit grew by 464.8% year-on-year to $20.81 million.
In terms of losses, the net loss for this period decreased from about $400 million in the same period last year to approximately $360 million.
This means that MiniMax has completely bid farewell to its early stage of relying on user acquisition through paid traffic for consumer-facing pan-entertainment applications. After fully pushing its foundational and multimodal model capabilities, it has shifted to natural network growth driven by the enterprise and developer ecosystem.
Let's make a simple horizontal comparison:
- According to a report by The Information, DeepSeek achieved $70.7 million in revenue in the first seven months of this year. Excluding the bustling month of July, MiniMax's first-half revenue is roughly estimated to be about twice that of DeepSeek.
- Although DeepSeek's net revenue is lower than MiniMax's, its first seven months achieved 10 times the total revenue of 2025, a growth rate much higher than MiniMax's.
- DeepSeek's net loss for the first seven months was $106 million, while MiniMax's net loss for this period is about 3.4 times that of DeepSeek.
New and Old Consensuses
MiniMax revealed that a fault-like surge in Token consumption is an important reason for the explosive revenue growth this period.
In July of this year alone, Token consumption across all MiniMax platforms reached exactly 20 times that of January. To date, its global enterprise customers and developers have exceeded 2 million, 10 times the number at the end of 2025.
The reason for such an exaggerated, non-linear leap in Token consumption lies in the migration of the model consumption paradigm.
CEO Yan Junjie revealed that the paradigm of model consumption has fully transitioned from "direct conversations between humans and AI" to "multi-turn cascading interactions between Agents and AI."
The growth in inference demand driven by Agents is significantly faster than the growth in human user numbers and message counts. This directly leads to an exponential explosion in Token consumption per user.
This is already an "old consensus" in the rapidly changing model industry.
The more important new consensus is that post-training has become the new ceiling for Scaling.
Judging from recent public information on Zhipu's GLM 5.3, DeepSeek V4 Flash (0731 version), and Gemini 3.x Flash, an increasing number of model vendors are heavily relying on post-training during minor version updates, or even relying solely on post-training, to significantly improve model performance on specific mainstream tasks.
Yan Junjie also pointed out this shift in the new industry consensus during the earnings call: the Scaling of current large models is rapidly expanding from singular pre-training to mid-training, post-training, and reinforcement learning.
In the post-training phase, massive synthetic data generation, environment rollouts in reinforcement learning (inference/generation), and trial-and-error verification of generated tasks are all essentially inference computing. The lower the inference cost per unit of computing power and the larger the throughput, the larger the trajectory and experimental scale the model can handle during post-training. This is why a group of model vendors, including MiniMax, are investing an increasingly higher proportion of computing power in the mid-training and post-training phases.
Yan Junjie pointed out that under the same computing power scale, the multiplier of inference computing efficiency improvement shows a basically linear correlation with the multiplier of post-training scale expansion. This sentence can be roughly understood as: the harder the push in post-training, the greater the improvement in inference efficiency, and the better the final effect presented by the model.
In other words, regardless of whether people think the Scaling Law has hit a wall or whether model parameters have reached their limit—post-training...