NVIDIA Announces Kumo for Snowflake

NVIDIA announced its Agent Knowledge Foundation on October 21, 2026, including Kumo, a family of open models, libraries and tools for predictive AI on structured data. Snowflake customers can use the NVIDIA Kumo Relational model and Kumo Engine, an open-source package for training graph neural networks (models that learn from connected data), with Snowflake ML. The setup runs in Snowflake Notebooks on Container Runtime, allowing data science teams to train models, score results and keep data inside Snowflake’s security and governance perimeter.

Kumo learns from relational structure and temporal patterns, such as tables, the keys connecting them and when events occur, rather than requiring teams to flatten connections into one wide table. The approach applies to customer churn, fraud detection, demand forecasting and product recommendation. NVIDIA says GPU acceleration supports learning directly from connected tables, keys and timing patterns without moving data outside the governed environment.

Teams install the Kumo package in a Snowflake Notebook and can train a model to predict which active customers will buy again in the next 90 days, using four Snowflake tables and a few lines of Python. Snowflake Notebooks connect to governed data, define prediction targets and show results; GPU compute pools provide NVIDIA GPU infrastructure through Snowflake’s container runtime. NVIDIA’s verified agent skills for Snowflake CoCo can turn a natural-language prediction task into a runnable notebook, from connecting data and GPU training to writing predictions back to Snowflake.

Snowflake says customers including Affirm, National Oilwell Varco (NOV), VideoAmp and Yieldmo are already using Kumo with Snowflake ML. NOV’s vice president of finance business operations said Kumo accelerated work across different business areas, while CoCo made it easier to explore and build where the data resides. Yieldmo’s vice president of product said its Companion Audiences product builds act-alike audiences from existing performance, lets buyers set a similarity threshold and choose whether it leads or supports a campaign, and allows managed-service buyers to train on conversion data with no seed spend. Snowflake says Kumo workloads can be orchestrated with ML Jobs within its existing security perimeter.

NVIDIA announced its Agent Knowledge Foundation on October 21, 2026, including Kumo, a family of open models, libraries and tools for predictive AI on structured data. Snowflake customers can use the NVIDIA Kumo Relational model and Kumo Engine, an open-source package for training graph neural networks (models that learn from connected data), with Snowflake ML. The setup runs in Snowflake Notebooks on Container Runtime, allowing data science teams to train models, score results and keep data inside Snowflake’s security and governance perimeter.

Kumo learns from relational structure and temporal patterns, such as tables, the keys connecting them and when events occur, rather than requiring teams to flatten connections into one wide table. The approach applies to customer churn, fraud detection, demand forecasting and product recommendation. NVIDIA says GPU acceleration supports learning directly from connected tables, keys and timing patterns without moving data outside the governed environment.

Teams install the Kumo package in a Snowflake Notebook and can train a model to predict which active customers will buy again in the next 90 days, using four Snowflake tables and a few lines of Python. Snowflake Notebooks connect to governed data, define prediction targets and show results; GPU compute pools provide NVIDIA GPU infrastructure through Snowflake’s container runtime. NVIDIA’s verified agent skills for Snowflake CoCo can turn a natural-language prediction task into a runnable notebook, from connecting data and GPU training to writing predictions back to Snowflake.

Snowflake says customers including Affirm, National Oilwell Varco (NOV), VideoAmp and Yieldmo are already using Kumo with Snowflake ML. NOV’s vice president of finance business operations said Kumo accelerated work across different business areas, while CoCo made it easier to explore and build where the data resides. Yieldmo’s vice president of product said its Companion Audiences product builds act-alike audiences from existing performance, lets buyers set a similarity threshold and choose whether it leads or supports a campaign, and allows managed-service buyers to train on conversion data with no seed spend. Snowflake says Kumo workloads can be orchestrated with ML Jobs within its existing security perimeter.