What’s disclosed

Nebius named NVIDIA as a customer in May 2025 without specifying services or workloads. Their March 2026 strategic partnership covers AI factory design, inference software, infrastructure deployment and fleet management. NVIDIA validated training performance on Nebius production HGX B300 infrastructure using its Performance Benchmarking recipes. Their engineers jointly test and optimize the platform.

NVIDIA names Nebius as a Physical AI Data Factory Blueprint cloud partner for training-data generation, augmentation and evaluation. Nebius manages NVIDIA OSMO across simulation, training and inference. Voxel51 describes their joint loop for finding data gaps, generating missing data and evaluating results before training. NVIDIA and Nebius built an early Blackwell sandbox with orchestration runbooks and multi-node GB200 NVL72 guidance. Nebius announced Hopper access through NVIDIA Brev in June 2025 and Agentic Commerce and Retail Catalog Enrichment blueprints on AI Cloud in June 2026. Their Enterprise Readiness Initiative combines Nebius engineers and NVIDIA expertise to optimize enterprise inference with Token Factory. DataRobot CEO Debanjan Saha describes a validated AI factory delivered with both companies. Nexla describes their joint demonstration converting travel videos into personalized itineraries. NVIDIA also co-developed Revolut’s Nebius-trained PRAGMA model, contributing training infrastructure design and evaluation methodology. Nebius Academy cooperation includes selected NVIDIA Deep Learning Institute courses, joint courses and workshops on Nebius infrastructure. Its Agentic AI course uses Nemotron through Token Factory alongside Tavily. NVIDIA participates in the Nebius AI Builder Program through courses and Nemotron models.

Tavily says NVIDIA uses its search for post-training and evaluation. NVIDIA documents converting Wikidata paths into search riddles, generating Tavily search trajectories and processing them into supervised fine-tuning data. Its Nemotron Super recipe reports about 7,000 high-quality trajectories from 50,000 Wikidata seeds through a hosted Tavily MCP endpoint. Rollouts allow 25 tool-call turns and a 300-second timeout. Complete traces include structured answers, supporting URLs and justification. Tavily also reports science-reasoning examples, BrowseComp evaluation and a finance-agent benchmark. NVIDIA’s Nemotron 3 Ultra paper describes a Tavily search-and-browsing harness that saves results to disk for selective shell inspection across context resets. Its finance evaluation adds Tavily web search as a supplementary tool. Separately, NVIDIA generated about 80,000 research trajectories with its GPT-OSS-120B-powered AI-Q researcher and Tavily retrieval. About 67,000 passed filtering for Nemotron 3 Super fine-tuning. Tavily says that model powered AI-Q’s researcher and subagents in its first-place DeepResearch Bench I and II submission.