How they use Nebius AI Cloud

TheStage AI uses Nebius GPU VMs to test and optimize inference. Nebius’s customer story describes uploading models, analyzing speed, cost and quality trade-offs through TheStage, and deploying optimized models on Nebius. TheStage keeps several H100 nodes available for experiments, develops inference-acceleration methods and fine-tunes client models there. One foundational-model optimization project used a single GPU node for evaluations and experiments. Nebius says TheStage selected it after testing providers for training stability, with uninterrupted GPU operation requiring no special MLOps expertise. Nebius also describes Stable Diffusion v1.5 H100 tests using UNet quantization and structured sparsification for INT8 and sparse computation, reducing GPU requirements.

In May 2025, the founder reported live diffusion inference on Nebius Blackwell infrastructure. With early access to NVIDIA HGX B200 on Nebius AI Cloud, TheStage benchmarked optimized diffusion models. A May interview reported about 22.5 iterations per second for FLUX.1 versus 6.5 on H100 with PyTorch bf16. The article reported 1024×1024 generation in 0.3 seconds with FLUX.1-schnell versus a previous record of 0.6 seconds, and 1.85 seconds with FLUX-dev versus rivals’ 3.1 seconds. These tests used Nebius B200 infrastructure. Nebius reported up to 1.6× latency reduction from B200 alone and up to 3.5× with TheStage’s compiler optimizations. Integration into cloud platforms including Nebius was underway. In August 2025, the CEO reported initial inference results about 3.5 times faster.

In January 2026, the founder credited Nebius with a training cluster for TheWhisper-Turbo 0.8B, a speech transcription and translation project. The release reported English and multilingual Open ASR Leaderboard word error rates of 6.00 and 4.74. It supports 11 languages, word-level timestamps and real-time streaming. ANNA compression uses mixed INT8 modes on NVIDIA and mixed 4-bit and 8-bit weights on Apple. In May 2026, announcing wider platform availability, he credited Nebius with GPU infrastructure support from day zero. For TheStage’s Gemma4 E2B/E4B mobile-inference compression release, he thanked Nebius for GPU infrastructure and said evaluations ran on Modal. E2B shrank from 9.3 GB to 1.4 GB. In its case study of optimizing Recraft’s 20B and Red Panda diffusion models, TheStage also thanked Nebius for GPU infrastructure supporting its research.