How they use Nebius AI Cloud

Dubformer uses Nebius GPU virtual machines to train speech synthesis and recognition models for its AI dubbing and localization service. The customer story describes NVIDIA A100 and H100 GPUs. Training data spans hundreds of terabytes and streams from Nebius Object Storage to the GPUs. Speech synthesis training uses thousands of hours of audio, while speech recognition training uses hundreds of thousands of hours.

The customer story describes training around the clock and a queue of dozens of experimental tasks. Hypotheses run on Nebius virtual machines alongside production workloads or wait in the queue, depending on data volume. Most experiments concern speech synthesis, while recognition models can take months to train. Although the trained models fit in one GPU’s memory, Dubformer prefers to use eight GPUs simultaneously on an SXM board. The GPUs communicate through NVLink. A Nebius digest also describes model deployment on its platform.