Enterprise data is inherently relational. Customers place orders that contain products, support tickets link to accounts, transactions connect devices, merchants, and locations, and knowledge graphs span entities across the business. Traditional machine-learning approaches flatten these relationships into hand-crafted features, discarding much of the signal that lives in the connections themselves. Graph machine learning changes that equation by treating tables and foreign keys as a native graph and learning directly over the structure.
NVIDIA Kumo brings graph neural networks (GNNs) and related structured-data foundation models into the Snowflake AI Data Cloud. Running as a Snowflake Native Application on Snowpark Container Services (SPCS) or via connectors and notebooks, Kumo lets data scientists and AI teams build predictive models on multi-table relational data while keeping data, training, and scoring inside the governed Snowflake perimeter. The combination removes a major barrier to production graph ML: the need to extract, transform, and move large relational datasets into external graph platforms.
This post examines the technical architecture of the Kumo–Snowflake integration, performance and accuracy advantages, real-world use cases, comparisons to traditional approaches, implications for data science and AI teams, and what the partnership signals for enterprise ML in 2026.
Why Graph ML Matters for Relational Enterprise Data
Most business prediction problems are multi-table by nature:
- Will this customer churn in the next 30 days? (customers + orders + support interactions + product usage)
- Which products should we recommend next? (users + items + interactions + attributes)
- Is this transaction fraudulent? (accounts + devices + merchants + historical patterns)
- What is the likely lifetime value or next best action?
Feature engineering pipelines that attempt to capture these relationships are expensive to build, brittle to maintain, and often incomplete. Graph ML models the entities as nodes and the relationships (foreign keys, events, interactions) as edges. Message-passing GNNs then learn embeddings that incorporate neighborhood context, often delivering higher accuracy with far less manual feature work.
Kumo specializes in this paradigm. It represents Snowflake tables as a graph, uses predictive query language (PQL) to define the prediction task in a declarative way, and trains or applies models that operate over the full relational structure—optionally enriched by LLM context from unstructured data via Cortex or other models.
Technical Architecture: Kumo Inside the AI Data Cloud
Deployment options
- Snowflake Native App on SPCS — Kumo services run as containers inside the customer’s Snowflake account. Training jobs, model artifacts, and predictions stay within the account boundary. Snowflake’s control plane schedules and isolates the containers; data access uses native SQL/Snowpark interfaces.
- Direct connectors and notebooks — Teams can create graphs from Snowflake tables via Graph.from_snowflake(), train or score in Snowflake Notebooks (including GPU-backed Container Runtime), and publish models to Snowflake Model Registry.
- Hybrid patterns — Some deployments train inside Snowflake and serve online predictions from a customer-controlled VPC while still keeping source data inside Snowflake.
Core workflow
- Connect to Snowflake tables/views (automatic discovery of columns, types, primary keys, and foreign-key relationships is supported).
- Define a graph (tables as node sets, relationships as edges).
- Express the prediction task with a predictive query (e.g., “PREDICT COUNT(ORDERS.*, 0, 90, days) > 0 FOR EACH CUSTOMER …”).
- Train or apply a model (graph transformers / GNNs that learn from the multi-table structure).
- Write predictions back as Snowflake tables and monitor with built-in evaluation and explainability.
Because sampling and context generation happen close to the data, complete source tables do not need to be loaded into application memory; relevant neighborhoods are retrieved for each prediction entity.
Security and governance
Data never leaves the customer’s Snowflake account in the Native App model. Access is governed by existing Snowflake roles and privileges. Model artifacts and prediction tables are ordinary Snowflake objects subject to the same controls, tagging, and auditing as any other enterprise data.
Performance and Accuracy Advantages
Graph ML on relational data frequently outperforms single-table or heavily engineered baselines because the model can learn which neighboring entities and paths matter. Kumo’s design emphasizes:
- Elimination or drastic reduction of manual feature engineering.
- Learning embeddings across multiple tables in one model rather than stitching separate models.
- Time-aware sampling and predictive queries that respect temporal order (critical for churn, propensity, and fraud).
- Explainability that traces predictions back to contributing columns and raw records.
- Integration with LLMs (including Cortex) so unstructured text can enrich the graph signal.
NVIDIA has also open-sourced elements of the Kumo family (structured/relational models, engine libraries, forecasting and anomaly variants), enabling broader experimentation while production deployments continue to benefit from the managed Snowflake-native path. Snowflake customers in advertising, energy, financial services, and retail are already using these capabilities.
Benchmark-style claims around the Kumo Engine highlight substantial speedups for multi-hop context generation compared with general-purpose engines on large graphs, underscoring that the stack is engineered for production-scale relational prediction rather than research prototypes.
Real-World Use Cases
Fraud and risk
Multi-hop patterns (device → account → merchant → prior disputes) are natural graph problems. Kumo models can surface risky entities or transactions by learning from the full relationship neighborhood rather than isolated features.
Recommendations and personalization
Item-to-user and item-to-item predictions become declarative graph queries. Historical interaction graphs plus product and user attributes drive ranked recommendations that update as new events arrive.
Customer propensity and LTV
Predict which customers will purchase again in a future window, or estimate lifetime value, by combining order history, engagement, and support data in a single graph model.
Knowledge graphs and entity resolution
Linking records across systems and inferring missing relationships benefits from the same relational inductive bias.
Operational and supply-chain forecasting
When entities (SKUs, locations, suppliers) and their interactions form a graph, structured forecasting models can incorporate relational context beyond pure time-series approaches.
Comparisons to Traditional Graph and ML Approaches
| Approach | Strengths | Limitations vs. Kumo on Snowflake |
|---|---|---|
| Manual feature engineering + tabular models | Familiar tooling | High maintenance cost; loses multi-hop signal |
| External graph databases + GNN frameworks | Flexible research stack | Data movement, security boundary, ops overhead |
| Single-table Snowflake ML / time-series functions | Fully native, simple | Limited multi-table relational depth |
| Generic AutoML on flattened tables | Fast to try | Still requires feature pipelines |
Kumo’s differentiator is running sophisticated graph learning inside the AI Data Cloud with declarative predictive queries, native governance, and minimal data movement. It complements rather than replaces Snowflake’s broader ML surface (Container Runtime, Model Registry, Cortex, feature stores).
Implications for Data Scientists and AI Teams
Productivity
Data scientists spend less time on feature pipelines and more time on problem formulation, evaluation, and business impact. Predictive queries make the modeling intent explicit and reviewable.
Governance and compliance
Models trained and scored inside Snowflake inherit existing access controls, masking, tagging, and audit capabilities—critical for regulated industries.
MLOps
Models can be versioned in Snowflake Model Registry, scheduled via tasks or ML Jobs, and monitored with the same observability patterns used for other production workloads.
Collaboration with agents
As enterprise agents proliferate, graph-derived predictions (risk scores, recommendations, propensities) become high-value tools those agents can call—grounded in the same governed data.
Actionable insights for leaders
- Identify 1–2 high-value multi-table prediction problems currently limited by feature engineering effort.
- Pilot Kumo as a Native App or in Snowflake Notebooks on a well-understood domain (e.g., churn or recommendations).
- Establish clear success metrics: lift vs. current models, time-to-first-model, and operational cost of maintaining features.
- Align security and platform teams early so SPCS compute pools, roles, and network rules are ready.
- Treat graph predictions as first-class data products that downstream applications and agents can consume.
- Combine Kumo relational models with Cortex for hybrid structured + unstructured signal where text adds value.
- Monitor the open Kumo model family for opportunities to accelerate specific tasks while keeping production paths governed.
What This Signals for Enterprise ML in 2026
Three trends converge in the Kumo–Snowflake integration:
- Relational data is the new frontier for foundation models — After text and images, structured multi-table data is receiving serious foundation-model investment (Kumo Structured/Relational and related open releases).
- Native > bolted-on — Enterprises prefer ML that runs where the data and governance already live. SPCS and Native Apps make this practical for sophisticated techniques that previously required external platforms.
- Agents need reliable predictive tools — As agentic systems move into production, they require accurate, low-latency predictions grounded in enterprise relationships. Graph ML supplies exactly that class of signal.
Organizations that treat multi-table relational prediction as a core capability—rather than a series of one-off feature projects—will move faster from data to decisions. The Snowflake + NVIDIA Kumo combination provides a governed, scalable path to do so inside the AI Data Cloud.
Conclusion
Graph machine learning unlocks predictive signal that lives in the relationships among enterprise entities. NVIDIA Kumo brings production-grade GNNs and structured foundation models to Snowflake, allowing teams to train and score multi-table models without extracting data from the AI Data Cloud. Declarative predictive queries, native security, explainability, and integration with the broader Snowflake ML surface dramatically lower the barrier to accurate relational predictions for fraud, recommendations, propensity, and beyond.
For data science, AI, and platform leaders, the opportunity is concrete: replace brittle feature pipelines with graph-native models that learn directly from the data’s natural structure, keep everything inside existing governance boundaries, and feed higher-quality predictions to both human decision-makers and emerging enterprise agents. In 2026, that combination of accuracy, governance, and operational simplicity is becoming table stakes for serious enterprise ML.