Our software and operations stack has been validated across some of the world's most demanding GPU cluster environments. For current delivered and in-development clusters, see the track record on Compute.
One of Asia's top three autonomous driving companies deployed its first thousand-card GPU compute cluster on our cluster management platform, network integration, and ongoing operations services.
A national compute hub serving frontier large model developers. Our platform provided cluster management software and ongoing operations for 4,000 NVIDIA GPU cards in a multi-tenant environment.
Japanese AI research company and developer of the open-source framework Chainer. Built the country's largest GPU cluster — 12,288 GPUs across P100×8 and V100×8 configurations — alongside its own AI chip programme.
Deployed with a leading global short-video platform alongside DriveNets and Broadcom — the first production deployment worldwide of a Scheduled-AI-Fabric built on our DDC-based network architecture.
An e-commerce platform's AI image editing tool accelerated with our inference optimization layer — minimal code changes, direct PyTorch compatibility.
Outcome · Best-in-class acceleration versus comparable engines, with simplified deployment.A carrier's AI custom video ringback tone project — a latency-sensitive, high-volume generative workload — accelerated on our inference platform, with multi-resolution support and LoRA hot-swap across SVD, Diffusers, ComfyUI, and WebUI.
Outcome · 40%+ average performance uplift, ~100% inference speedup for video generation, 10% VRAM reduction.GPU cards under active management
cluster fault recovery time
peak GPU utilization in training
| Metric | Value |
|---|---|
| Total GPU cards managed (cumulative) | 10,000+ |
| Thousand-card clusters deployed | 4+ |
| Ten-thousand-card clusters | 1 |
| Cluster fault recovery time | <1 ms |
| Training GPU utilization (peak) | Up to 55% |
| Ops staffing reduction via automation | ~50% |
| Inference throughput uplift (LLM) | Up to 2.5× |
| Inference latency reduction (LLM) | Up to 2.7× |
| End-to-end inference speedup (diffusion) | Up to 3× |