By SnowflakeLearning Objectives
- Understand what Snowflake Cortex AI is and how it integrates AI into data warehousing
- Identify the key Cortex AI features including LLM functions, agents, fine-tuning, and the Arctic model
- Evaluate Snowflake's competitive position against Databricks, AWS, and Google BigQuery
What Is Snowflake Cortex AI?
Snowflake Cortex AI is a suite of AI and machine learning capabilities built directly into the Snowflake data platform. Instead of exporting your data to a separate AI tool, Cortex AI lets you run large language models, build AI agents, and perform machine learning — all using SQL queries on data that never leaves Snowflake.
This "AI on governed data" approach is especially appealing to regulated industries like finance and healthcare, where data residency and security requirements make it impractical to send data to external AI services.
💡Key Concept
SQL-First AI: Cortex AI functions are called directly in SQL queries. For example, SELECT AI_SUMMARIZE(document_text) FROM contracts summarizes every contract in your table — no Python, no API keys, no data export. The AI runs where your data already lives.
Core Features
Cortex AI Functions (GA November 2025)
Built-in SQL functions that bring AI capabilities directly to your queries:
| Function | What It Does |
|---|---|
| AI_COMPLETE | Chat and text completion using hosted LLMs |
| AI_CLASSIFY | Text and image classification |
| AI_SUMMARIZE | Automatic text summarization |
| AI_TRANSCRIBE | Audio and video transcription |
| AI_EMBED | Generate vector embeddings for semantic search |
| AI_SIMILARITY | Compare text similarity scores |
| AI_REDACT | Detect and redact personally identifiable information (PII) |
Third-Party Model Access
Cortex AI hosts major frontier models that run directly on your Snowflake data:
- OpenAI GPT-5.2 — same-day preview access announced
- Claude 4 Opus — Anthropic's flagship model
- Llama 3 and Mistral — open-source options
Cortex Agents (GA November 2025)
AI agents that orchestrate across structured and unstructured data. They combine Cortex Analyst (for SQL/analytics questions) with Cortex Search (for document retrieval) to answer complex business questions that span databases and documents.
Cortex Search
A vector and semantic search service for building RAG (Retrieval-Augmented Generation) applications. Index your documents, PDFs, and unstructured data — then query them with natural language, all within Snowflake.
Cortex Fine-Tuning
Customize open-source models (like Llama 3) on your proprietary data without exporting it. GPU-accelerated training runs inside Snowflake, billed per token processed.
Cortex Code
An AI coding agent launched in November 2025 that understands your enterprise data context. Now expanding beyond Snowflake-native SQL to support dbt and Apache Airflow. Already has 4,400+ users.
Snowflake Arctic
Snowflake's own open-source model family:
- Arctic LLM — 480 billion parameter Dense-MoE hybrid (only 17 billion active at inference time), trained for under $2 million. Excels at SQL generation, coding, and instruction-following. Apache 2.0 license.
- Arctic Embed — a family of text embedding models in five sizes for vector search and RAG.
Pricing
Cortex AI uses Snowflake's credit-based, per-token billing — not a separate subscription.
Standard Snowflake credits cost roughly $2 to $4 each depending on your edition and contract. A 30-day free trial with $400 in credits is available for new accounts.
⚠️Warning
Cortex AI costs can escalate quickly on large datasets. A single poorly-scoped query against millions of rows can cost thousands of dollars. Always filter your data before running AI functions, and monitor usage through Snowflake's cost management tools.
Snowflake Cortex vs. Competitors
| Platform | Approach | Best For | Lock-in Risk |
|---|---|---|---|
| Snowflake Cortex AI | SQL-first AI on governed data; multi-cloud | Teams wanting AI inside their data warehouse | Low (runs on AWS, Azure, and GCP) |
| Databricks | Full lakehouse + custom AI/ML at scale | Organizations needing end-to-end data engineering + AI | Medium (Spark ecosystem) |
| AWS SageMaker/Bedrock | Broadest managed AI services catalog | AWS-committed organizations | High (AWS-specific) |
| Google BigQuery ML | Serverless analytics + AI functions | GCP-committed teams wanting fast analytics | High (GCP-only) |
Snowflake's key differentiator: AI runs directly on governed data with no data movement required — critical for regulated industries where data residency matters.
Company Details
| Detail | Info |
|---|---|
| Company | Snowflake Inc. (NYSE: SNOW) |
| Founded | 2012 |
| CEO | Sridhar Ramaswamy (since February 2024; ex-Google Ads SVP) |
| Headquarters | Bozeman, Montana (principal office) with major Bay Area presence |
| Employees | ~7,000+ |
| Market Cap | ~$60 billion |
| Product Revenue (FY2026) | $4.72 billion (+30% year-over-year) |
| Total Customers | 13,300+ |
| Million-Dollar Customers | 733 (+27% year-over-year) |
| AI Feature Users | 9,100+ accounts; 6,100+ weekly active AI users |
| Recognition | Fortune Future 50 ranked number 1 (2025) |
| Website | snowflake.com |
Strengths
- No data movement — AI runs directly on governed data inside Snowflake; critical for regulated industries
- SQL-first approach — data analysts and SQL developers can use AI without learning Python or ML frameworks
- Multi-cloud — runs on AWS, Azure, and GCP, reducing vendor lock-in
- Frontier model access — GPT-5.2, Claude 4 Opus, Llama 3, and Mistral available alongside Snowflake's own Arctic model
- Massive enterprise adoption — 13,300+ customers including hundreds of Fortune 500 companies
- Cortex Agents — orchestrate across structured databases and unstructured documents in a single query
Limitations and Considerations
- Credit-based pricing complexity — per-token costs vary by model and can escalate quickly on large datasets
- Narrower AI scope — focused on analytics and data tasks; not a general-purpose AI development platform like Databricks or SageMaker
- Arctic model limitations — strong at SQL and code but less competitive than GPT-5.2 or Claude 4 for general reasoning
- Snowflake dependency — Cortex AI only works within the Snowflake platform; you must be a Snowflake customer
- Fine-tuning constraints — limited to supported open-source models; you cannot bring arbitrary custom models
Key Takeaways
- Snowflake Cortex AI brings AI directly into the data warehouse — SQL functions, LLM access, agents, fine-tuning, and vector search, all on governed data that never leaves the platform
- Particularly valuable for regulated industries (finance, healthcare) where data residency and security requirements prevent sending data to external AI services
- Hosts frontier models (GPT-5.2, Claude 4 Opus) alongside Snowflake's own Apache 2.0 licensed Arctic model family
- 13,300+ customers and $4.72 billion in product revenue make Snowflake one of the largest data platforms; Cortex AI is its fastest-growing feature set
