How they use Nebius AI Cloud

Mirage, formerly Captions, trained its audio-to-video foundation model on a Nebius cluster. GeekWire's interview overview describes the model as capturing voice, motion and emotion. It credits Nebius's scalable infrastructure, reliability and hands-on support with helping Captions train large models and accelerate research and development. In its Q1 2025 earnings call, Nebius said Captions scaled GPU training for Mirage using its infrastructure and accelerated time to market. In announcing the Mirage name, Gaurav Misra said the company was building a video foundation model intended to turn an idea into fully usable, cohesive output, starting with short-form video.

Mirage's team says cluster reliability makes long training runs more predictable and gives researchers more time to test new models. Drew Jaegle, then head of AI, described experiments spread across hundreds of GPUs and the disruption that even a short outage can cause. He said Nebius cluster stability helps the team plan large experiments without constantly adjusting for potential failures. Co-Founder and CEO Gaurav Misra credited automated fault handling and a low incident rate with giving the team more time to iterate on models.

In the March 2025 interview, Jaegle described building training infrastructure over the preceding quarter and supporting single training runs across thousands of GPUs. He credited Nebius with helping recover from failures and providing support around the clock for the New York team. As Captions explored scaling inference, he said Nebius provided flexible ways to scale its model and quickly test it in different settings.