NVIDIA Kumo Relational: Foundation Model for Enterprise Databases Beats Supervised ML Without Training

NVIDIA has introduced Kumo Relational, a pre-trained foundation model that operates directly on relational databases to deliver predictions for tasks such as churn forecasting, demand planning, and fraud detection. Unlike traditional pipelines that demand weeks of feature engineering and per-task model training, Kumo Relational gathers relevant historical context — up to 10,000 rows — and produces answers in a single forward pass.

Across 41 benchmark tasks spanning 15 databases, the model outperforms supervised relational deep-learning approaches and expert-tuned gradient-boosted tree pipelines. On the enterprise-grade SAP SALT benchmark, it exceeds AutoGluon ensembles by 8% in mean reciprocal rank (MRR) out of the box, and by 13% after fine-tuning. The system scales to databases exceeding 500 billion rows.

Why it matters for GPU / AI infrastructure

Kumo Relational signals the same paradigm shift that large language models brought to NLP: a single pre-trained model replaces dozens of task-specific pipelines. For infrastructure teams, this means GPU clusters can serve a unified inference endpoint for diverse predictive queries via NVIDIA's open-source structured-data-models library or as an NVIDIA NIM microservice. The Predictive Query Language (PQL) also makes the model agent-ready, enabling autonomous systems to anticipate business outcomes and act on them.

With no flattening of tables and no manual feature crafting, ML teams can address use cases that were previously too labor-intensive to justify. The model is available today for evaluation and deployment on GPU-accelerated infrastructure.