Elastic Compute Leasing for an AI Company

Solution

Flexible Leasing Model: Customers sign a monthly base-compute contract that guarantees minimum resources, while dynamic scaling is also supported—when a training job starts, the platform automatically allocates the required number of GPUs from the resource pool and releases them after the training ends, charging only for actual GPU-hours used.
Dedicated Compute Pool: Chinasoft International provisions a physically isolated Ascend compute zone for the customer at the Xi'an AI Computing Center, where data storage, model training, and inference services all run within the dedicated environment to meet data-security and compliance requirements.
Full-Stack Managed Operations: Chinasoft International provides 7×24 compute monitoring, driver upgrades, automatic fault recovery, and performance tuning, so customer algorithm engineers can call compute resources directly via API or Jupyter Notebook without managing underlying infrastructure.
Integrated Model Deployment: After training is completed, models are automatically packaged and deployed to the inference service cluster in the same dedicated zone, enabling seamless training-to-inference handoff.

Implementation Results

Compared with self-purchased hardware of the same scale (depreciated over three years) or monthly reserved leasing, the elastic leasing model lets the customer pay only for actual training time, cutting monthly compute spend from RMB 480,000 to RMB 170,000.
Chinasoft International provides pre-configured deep-learning framework images (PyTorch, TensorFlow, MindSpore) and distributed-training optimization, shortening the model iteration cycle from one week to four days.
The customer does not need to hire IT operations staff; the compute platform is ready to use out of the box, with mean time to recovery (MTTR) of less than 30 minutes.
Monthly active API calls grew from 200,000 to 1.5 million, while inference service availability remained above 99.95% and service was never interrupted due to insufficient compute.
During an urgent regulatory compliance check that required re-validating all historical models within 48 hours, the customer scaled up to a 200-GPU cluster with one click and completed the task six hours ahead of schedule.

Customer Testimonial

Chinasoft International's elastic compute leasing lets us focus all our energy on model algorithms instead of struggling with GPU drivers and network tuning. Without this flexible compute service, we could not have completed four major version iterations in 18 months.