Snowflake Cortex Aims to Accelerate In-Warehouse AI
Snowflake's Cortex AI platform is being positioned as a secure way to build and deploy Large Language Model applications directly within the data warehouse. Guides highlight its ability to handle end-to-end AI workflows, such as summarization and natural language querying, without moving sensitive data. Because it is built into Snowflake's architecture, Cortex inherits the platform's security and access controls, making it suitable for regulated industries like healthcare.
- Snowflake’s pricing model for Cortex is primarily consumption-based, charging per million tokens for most LLM functions. However, services like Cortex Search also introduce a continuous "serving cost" based on the size of the indexed data, which accrues even when the service is idle. - To enhance its AI capabilities, Snowflake has partnered with Nvidia to integrate technologies like NeMo Retriever directly into Cortex. This allows applications to more accurately retrieve information from data within Snowflake to provide context for generative AI models. - Cortex provides access to a range of third-party LLMs from Google, Meta, and Anthropic, in addition to Snowflake's own open-source model, Arctic. This allows developers to select the most appropriate model for a specific task, such as summarization or sentiment analysis, directly within their SQL or Python code. - For analytics engineering workflows, Cortex is being integrated with dbt to understand data lineage and metadata. This allows AI-powered features to assist with tasks like suggesting model improvements and debugging pipelines by using the context of the entire dbt project. - In healthcare, specific use cases include building chatbots to answer patient questions using retrieval-augmented generation (RAG) on medical documents and extracting structured information from unstructured clinical notes. Innovaccer's Gravity platform leverages Cortex on Snowflake's AI Data Cloud to create longitudinal patient views and accelerate AI adoption for providers and payers. - Competing directly with Databricks, Cortex is positioned as an "AI-as-a-service" offering focused on ease of use for SQL-centric users, while Databricks provides a more engineering-heavy platform for building custom AI and machine learning models from the ground up. - The LLM functions within Cortex became generally available in May 2024, followed by the general availability of Cortex Agents in November 2025. These agents can orchestrate workflows across both structured and unstructured data to answer complex questions. - To prevent harmful or unsafe outputs from the language models, Cortex includes a feature called Cortex Guard, which is built using Meta's Llama Guard model to filter responses.
