MiniMax Raises Three-Year Alibaba Cloud Procurement Cap by 220% to $1.2 Billion
Nashnova编辑部
AI company MiniMax has raised its three-year Alibaba Cloud spending cap by 220% to $1.2 billion — after burning through two-thirds of its 2026 budget in just six months, compute costs are outrunning revenue growth.
A 220% increase — how big is the new cap?
MiniMax's three-year Alibaba Cloud procurement framework rises to $1.2 billion, a 220% increase over the original agreement. The revision was disclosed via Hong Kong Stock Exchange filing on August 27.
Year by year: 2026 jumps from $115 million to $300 million, 2027 from $125 million to $400 million, 2028 from $135 million to $500 million — nearly triple the original plan across all three years.
This means → MiniMax expects compute consumption to accelerate each year, not grow linearly: the 2028 cap is 67% higher than 2026's.
Two-thirds of the annual budget gone in six months — where did it go?
By end of June 2026, MiniMax had consumed two-thirds of its original 2026 budget, with half the year still remaining.
In plain terms = the spending plan set at the start of the year was nearly exhausted by mid-year, forcing an emergency cap increase.
API call budgets surged in parallel — from $650,000 to $7.5 million for 2026 alone, a tenfold-plus jump. The three-year API cap now stands at $62.5 million, nearly 20 times the original limit.
Where is the money coming from — can revenue keep up?
MiniMax reported $116.6 million in first-half 2026 revenue, up 283% year-on-year. Enterprise sales grew 700% and were the core driver.
The product lineup includes M-series large language models (foundational AI models for text generation), the H3 video-generation model, and the consumer app Hailuo AI.
This means → the enterprise boom brings revenue and higher inference compute demand — every new enterprise client calling the model requires more real-time processing power on the back end.
Training and inference both burning — how does the math work?
Compute demand is rising on two fronts: training — next-generation model R&D requires larger-scale resources; inference — enterprise call volumes are climbing, driving sustained real-time compute consumption.
In plain terms = training is "the cost of building new models," inference is "the cost every time a customer clicks." Both are growing at once.
This reflects a critical question: whether the compute budget can keep pace with business expansion will directly determine MiniMax's ability to convert growth into sustainable margins — the faster the growth, the faster the burn.
市场有风险，内容仅供研究参考，不构成投资建议。