NVIDIA introduced Kumo Relational, a pre-trained foundation model that makes predictions directly on enterprise relational databases without requiring task-specific training, feature engineering, or manual table flattening. The model processes up to 10,000 historical examples as context before making predictions on outcomes such as churn, demand, or fraud in a single forward pass. By operating directly on database structures—including tables, foreign keys, and timestamps—Kumo Relational preserves fine-grained structural signals that traditional approaches lose when flattening data. The model demonstrates substantial performance improvements over production machine learning pipelines. Across 41 tasks spanning 15 databases, Kumo Relational outperforms supervised deep learning models and expert-engineered gradient-boosted tree systems, with gains of nearly 4 AUROC points over XGBoost on classification tasks and 7 MRR points on the SAP SALT enterprise benchmark. When fine-tuned, it exceeds AutoGluon ensembles by 8 percent in mean reciprocal rank, suggesting it could replace many existing ML systems in production environments. The release reflects a broader shift in enterprise machine learning, similar to how large language models transformed natural language processing. NVIDIA designed Kumo Relational to scale to databases with over 500 billion rows and work with both human analysts and AI agents through its Predictive Query Language interface. The model is available as an open-source library and through NVIDIA's cloud NIM service.