MiniMax raises its Alibaba Cloud purchase ceiling to 1.2 billion dollars
MiniMax raised its three-year Alibaba Cloud purchase ceiling by 220 per cent to 1.2 billion dollars as training and inference needs grow. With Chronos-2, Decathlon now forecasts weekly demand for up to 25,000 products per supply zone, while Salesforce uses a new placement setting to spread Agentforce model copies across two availability zones. Production scale is expanding decisions about distribution and resilience alongside spending.
Artificial Intelligence··Night
MiniMax expands its cloud ceiling
MiniMax raised its three-year ceiling for Alibaba Cloud services by 220 per cent to 1.2 billion dollars. The new limit runs through 2028, and the company attributes the change to fast-growing compute needs for training and inference. For 2026, its annual cap rises from 115 million dollars to 300 million dollars; for 2027, from 125 million dollars to 400 million dollars; and for 2028, from 135 million dollars to 500 million dollars. MiniMax had used two-thirds of its original budget by June 2026.[1]
Decathlon takes forecasting to production scale
Decathlon is using the Chronos-2 time-series foundation model to produce weekly sales forecasts for up to 25,000 products per supply zone. CPU inference for 7,000 to 15,000 time series takes 40 to 75 seconds. In the company's own measurement, weighted absolute percentage error at a 12-week horizon fell by 11 points in Southeast Asia and 15 in Latin America. The time needed to bring a new region online dropped from 6 months to 2 to 3 months; the results have not been independently tested.[2]
Salesforce spreads model copies across two zones
While hosting Agentforce models, Salesforce found that the default placement in SageMaker Inference Components did not spread copies evenly across 2 availability zones. A SchedulingConfig setting added to the CreateInferenceComponent interface balances zones within a chosen tolerance and places copies on separate instances for fault isolation. By Salesforce's own calculation, hosting several models on one instance had cut infrastructure cost by 8 times. Models with high-availability requirements are kept at a minimum of 2 copies.[3]
