Snowflake’s annual Summit wrapped its product keynote on June 2 in San Francisco with a clear message: data fragmentation is the single biggest thing holding enterprise AI back, and Snowflake just built the infrastructure to fix it.
With 20,000 attendees in the room and the theme “Making AI Real for Business,” this was less a product launch and more a declaration that the platform era for enterprise AI has arrived.
What Was Announced
Snowflake CEO Sridhar Ramaswamy opened the event in conversation with Anthropic Co-Founder and President Daniela Amodei, signaling just how deeply the two companies have aligned. On day two, co-founder and President of Product Benoit Dageville took the stage to unveil 26 new capabilities across data interoperability, governance, AI development, and security.
The headline announcement was a new open framework for interoperable enterprise data and AI. For the first time, enterprises can work against a single, live, governed copy of their data wherever it lives, whether inside Snowflake, external data lakes, or open systems, without moving or duplicating it.
Here is what that means in practice.
Apache Iceberg v3 support, now generally available. Snowflake now delivers full support for the latest version of the open table format, including more data types, cross-system change tracking, and higher performance on semi-structured data. Horizon Catalog powered by Apache Polaris enables bi-directional read and write access from external engines to Iceberg data managed by Snowflake.
Expanded Horizon Catalog. The catalog now functions as a unified context layer across the enterprise data estate. It captures semantic information, business definitions, and metadata from multiple systems, making that information available directly to AI applications. The result is that your AI agents know not just what your data says, but what it means in business terms.
Cortex Training. This is the biggest new capability for teams building their own AI. Cortex Training gives enterprises access to fully managed GPUs inside Snowflake to fine-tune and train open-weight foundation models, from the Qwen and Mistral families, directly on their proprietary data. No external infrastructure to stitch together, no data moving outside your security boundary. Teams can use techniques like reinforcement learning to build domain-specific models that outperform general-purpose APIs at a fraction of the cost.
Snowflake Datastream. A new fully managed streaming service for Apache Kafka, bringing real-time event data into Snowflake for AI applications that need fresh, continuously flowing context rather than batch updates.
AI Security Posture Management and Data Exfiltration Policies. As AI agents increasingly operate as non-human actors inside business systems, Snowflake added a stack of governance controls: model-level role-based access control, multi-party authorization, AI Security Posture Management, and Cortex Guard to prevent sensitive data from moving through AI model calls.
The Anthropic Connection
The Snowflake and Anthropic partnership continued to deepen. Enterprises are now adopting Claude through Snowflake Cortex AI to run AI agents on their Snowflake data with enterprise-grade controls, model selection across the Anthropic family, and a guarantee that sensitive data stays within the Snowflake environment. No separate API call, no data leaving your governed system.
What This Means for Business
The core problem Snowflake is solving has nothing to do with model capability. Most enterprises have decades of data spread across a dozen systems, each with its own format, its own access controls, and its own version of the truth. Before this week’s announcements, getting an AI agent to reason across all of that required copying data into a central warehouse, normalizing formats, rebuilding governance rules, and hoping nothing drifted. The ETL tax was real, and it was killing AI projects before they started.
The open interoperability framework changes the calculus. Agents can now query across systems without needing a centralised data copy. Cortex Training means companies can build models trained on their actual business context rather than relying on generic frontier models for domain-specific tasks. And the governance stack means security teams can let agents run without losing visibility.
For data teams, Cortex Training in particular deserves attention. The ability to train custom models inside the same platform where your data already lives, with managed compute and no infrastructure overhead, removes one of the last reasons a mid-sized company would need a dedicated MLOps team to run production AI.
For business leaders, the more important number is cost. When you can replace expensive frontier model API calls with a purpose-built model trained on your own data, the unit economics of AI shift dramatically. The question changes from “can we afford to use AI at this scale?” to “how quickly can we move our highest-volume workflows to a custom model?”
The Bigger Picture
It is worth noting what the “Making AI Real for Business” theme actually signals. For the last two years, the dominant enterprise AI conversation was about proof of concepts, pilots, and sandboxes. Snowflake’s Summit theme is not accidental. The vendors who survive the next phase will be the ones helping companies cross the gap from experiment to production at scale.
The combination of open data access, governed AI agents, and custom model training in a single platform is Snowflake’s answer to that challenge. Whether your data team is just starting to build with AI or already running agents in production, the decisions made this week in San Francisco will shape the tools they reach for over the next two to three years.
Enterprise DNA perspective: For EDNA Learn members working with Power BI, Python, and SQL, the Cortex Training announcement is the most practically significant. The ability to train domain-specific models on proprietary data, without standing up external ML infrastructure, makes advanced AI capabilities accessible to data teams that would never have the resources to build and maintain a custom ML pipeline. This is what closing the skills-to-deployment gap actually looks like in 2026.