Computing Power Leasing — Elastic Computing as a Service

For AI training, inference, and high-performance computing scenarios, providing elastic computing power leasing services based on Ascend and NVIDIA GPUs.

Functional Overview

For AI training, inference, and high-performance computing scenarios, we provide elastic computing power leasing services based on Ascend and NVIDIA GPUs. Customers do not need to build their own data centers; they can rent GPU computing power on demand (single card, multi-card, cluster), with billing by the second / by card-hour / by month, supported by 7×24 O&M, achieving 'rent and use immediately, pay as you go'.

Customer Pain Points

  • Small and medium customers cannot afford whole machines: 10,000-card clusters require billions of yuan in investment, and small and medium AI enterprises and research institutions cannot build them on their own, requiring elastic, small-scale computing power leasing.
  • Heavy O&M burden after leasing: Even after leasing computing power, subsequent driver adaptation, fault handling, and performance tuning still require professional teams, which customers lack.
  • Inflexible leasing billing: Under traditional monthly leasing models, customers' computing power remains idle during off-peak periods but still requires full payment, lacking fine-grained billing options such as time-of-day and pay-as-you-go.
  • Difficult cross-region scheduling: Customers' businesses are distributed across multiple locations, requiring task scheduling among different nodes, but there is a lack of unified entry points and resource views.

Solution Advantages

  • Flexible multi-spec leasing: Supports multiple specifications from 128, 256, 512 units to 10,000-card clusters, meeting the needs of large, medium, and scientific research institutions of different scales, with short-term, long-term, and on-demand expansion options.
  • Time-sharing and dynamic pricing: Based on real operating data, provides multiple billing models such as time-of-day (off-peak / peak), by Token consumption, and by card-hour, so customers only pay for actual usage.
  • Full-stack O&M hosting: Provides 7×24 computing power monitoring, driver upgrades, fault self-healing, and performance tuning, so customers do not need to build their own O&M teams.
  • Cross-region unified scheduling: Through the computing power scheduling platform, customers can allocate tasks among multiple intelligent computing nodes nationwide with one click, automatically selecting the optimal resource pool.
  • Security isolation and compliance: Supports physical machine isolation and container-level multi-tenant isolation, meeting high-end requirements such as training data not leaving the domain and model security encryption.

Architecture Diagram

Architecture Diagram