On July 4th, a report from Goldman Sachs highlighted that a recent call with MiniMax conveyed strong signals regarding its commercialization and technological advancement, with management expressing full confidence in achieving a $1 billion Annual Recurring Revenue target by the end of 2026.
The report identifies the most critical catalyst as the Chinese large language model industry reaching an inflection point in its "price war." As competitor DeepSeek announced price increases for peak usage periods, industry pricing is returning to more rational levels.
MiniMax is achieving gross margins significantly higher than its peers while maintaining highly competitive pricing, such as a blended rate of $0.22 per million tokens for its M3 model. This is attributed to its computational efficiency exceeding 90% utilization, deep integration with domestic chips, and unique "organizational agility." Furthermore, the upcoming launch of its H3 video generation model within weeks is expected to unlock new potential in the multimodal market.
Goldman Sachs maintains a Buy rating with a 12-month price target of HK$860, implying an upside potential of approximately 141% from the current share price of HK$356.80.
Projected Growth in Annual Recurring Revenue
The report states that MiniMax's management systematically outlined the projected growth milestones for its Annual Recurring Revenue during the call.
By the end of December 2025, the figure is expected to reach $100 million. It is forecasted to rise to $150 million by February 2026 and double again from that level by April 2026. Management anticipates further acceleration leading up to the official June 1st launch of the M3 model.
Management explicitly reaffirmed full confidence in hitting the $1 billion ARR target by the end of 2026.
Regarding pricing strategy, the M3 model maintains the same price point as its predecessor, the M2.7. Management emphasized this strategy's sustainability from a gross margin perspective, as upgrades to training and inference architecture have yielded cost savings of over 2x, largely offsetting the increased costs from the doubling of total parameters.
The company also previewed a larger-scale M3 series model for the second half of 2026, aiming to enhance intelligence levels while maintaining robust cost-performance.
This ARR growth trajectory forms the core basis for Goldman Sachs' revenue forecasts. The firm projects MiniMax's revenue will jump from $79 million in 2025 to $300 million in 2026, further reaching $880.1 million in 2027, and surpassing $2.4696 billion in 2028.
DeepSeek Price Increase: A Catalyst for Rational Pricing
The report views DeepSeek's recent pricing announcement as a significant market catalyst with external implications.
DeepSeek announced this week that its V4 official version will launch in mid-July, introducing a peak/off-peak differential pricing mechanism for its API. Peak hour rates will be double the off-peak rates, resulting in a blended pricing of approximately $0.35 per million tokens for the Pro version and $0.12 for the Flash version.
Goldman Sachs interprets this as an early signal of a transition away from the aggressive pricing that has characterized the Chinese AI model sector since late April 2026, where some players operated at zero or negative gross margins. This shift fundamentally reflects the pressure of inference costs being factored into pricing.
In contrast, MiniMax's M3 model offers a blended price of $0.22 per million tokens, presenting a significant competitive advantage in terms of performance-to-price ratio alongside superior gross margins. This advantage stems from a higher proportion of self-built, optimized computing power and an architectural design that enables efficient inference with a smaller number of activated parameters.
MiniMax specifically noted its self-operated compute achieves over 90% utilization. It balances load by serving knowledge workers and developers during peak hours and utilizing idle capacity for experiments and data organization during off-peak periods, thereby supporting cost advantages for long-duration agent workflows.
Upcoming H3 Video Model Launch
Concurrently, MiniMax is preparing to launch its next-generation video generation model, H3, expected for official release "within the next few weeks."
The core upgrades for H3 are reflected in two dimensions: a comprehensive improvement in video generation quality and functional diversity, underpinned by significant architectural advancements, and deep integration with the M3 model architecture. Capabilities from the large language model are embedded into H3's DiT architecture, enhancing its understanding of human actions and basic physical relationships.
Additionally, MiniMax is bringing in vertical domain experts to gradually enter the feature film and series production market, expanding the commercial boundaries of video generation.
Evolving Competitive Landscape
Goldman Sachs finds MiniMax's assessment of China's AI model competitive landscape strategically significant, noting a rapid consolidation from hundreds of players a year or two ago towards leading firms.
During the call, when addressing competition from AI labs under domestic internet giants, MiniMax defined its advantages as an efficient corporate organizational structure, higher infrastructure utilization, rapid model iteration capability, and quick responsiveness to emerging agent opportunities. Examples include the swift commercialization of MaxClaw following the rise of OpenClaw and the rapid deployment of the MiniMax Code product.
Management believes that as competition shifts from "one-time benchmark chasing" to "continuous product iteration and real-world deployment," sustainable ROI will become the core evaluation metric, making organizational agility increasingly valuable in this new competitive paradigm.
Global Infrastructure and Domestic Chip Strategy
Regarding computing infrastructure, MiniMax employs a dual-track strategy, directly leasing capacity from global cloud service providers while also engaging in deep collaboration with emerging cloud service providers.
Currently, its localized inference infrastructure covers over 200 countries and regions globally, with a highly diversified customer base that avoids over-concentration in any single country.
In the Chinese market, MiniMax has achieved a high degree of integration with domestic AI chips for inference tasks. As the capabilities of domestic chips continue to improve, this localization process is accelerating. This approach not only helps reduce reliance on foreign computing power but also builds a degree of supply chain resilience amid broader technology tensions.
Talent and Team Structure
In terms of talent strategy, MiniMax supports intense technological competition with a lean team structure. The company employs between 400 and 500 people, with over 80% engaged in research and development.
Between 300 and 400 employees participate in an Employee Stock Ownership Plan, which covers approximately 7% of the company's equity, using equity incentives to enhance talent retention.
The company continues to recruit fresh graduates from top Chinese universities and overseas institutions. Through its "10X Talent Program," it brings in seasoned experts from vertical industries to translate domain knowledge into capabilities for model training and real-world task optimization.
Comments