Everything Snowflake Announced at Summit 2026: CoCo, Datastream, Iceberg V3, and Cortex Training

Beyond new surfaces, the product gained true agentic autonomy. The new Automations feature executes recurring development tasks on a schedule, while Cloud Agents are serverless, event-driven workers that react to data changes or business triggers—such as alerting a team when inventory drops below a threshold—without any local client needing to remain open .

On the business-user side, the company also rebranded Snowflake Intelligence into Snowflake CoWork, a personal AI agent now available across an iOS mobile app, a Slackbot integration, and a Microsoft Excel extension .

Snowflake Datastream: A Fully Managed Kafka-Compatible Streaming Service

To solve the long-standing problem of operating and governing separate streaming infrastructure alongside the cloud data platform, Snowflake introduced Snowflake Datastream. The service is a Snowflake-native, fully managed streaming engine that speaks the full Apache Kafka protocol .

Existing Kafka producers and consumers can connect with a configuration change, eliminating the need for separate brokers, connectors, or additional clusters to manage . Streaming data lands directly as governed Snowflake tables or Apache Iceberg tables within Snowflake's security perimeter . Snowflake is explicitly positioning Datastream for the real-time data market, which it estimates at a $128 billion total addressable opportunity . The service is currently in private preview .

Apache Iceberg V3 and Bidirectional Interoperability

Snowflake announced the general availability of Apache Iceberg V3 support, claiming the broadest Iceberg feature set on the market . The centerpiece of the announcement is full bidirectional interoperability powered by the Snowflake Horizon Catalog, which embeds the open-source Apache Polaris catalog .

This means any Iceberg REST-compatible engine—such as Apache Spark, Trino, or Flink—can read and write Snowflake-managed Iceberg tables, and Snowflake can similarly read and write tables managed by external catalogs . Write access from external engines was in public preview as of the Summit, while read access had already reached general availability . To support this open architecture, the company also introduced Snowflake Storage for Apache Iceberg Tables, a new managed tier for open-format data .

Cortex Training: Bring Your Own Model, Managed by Snowflake

Rounding out the Summit was Cortex Training, an expansion of the existing Cortex Fine-tuning service into full custom model training. The service allows enterprises to fine-tune open-weight foundation models—including Qwen, Mistral, and Meta's Llama—on fully managed GPU infrastructure inside Snowflake's governed environment .

Organizations can use parameter-efficient fine-tuning (PEFT) techniques and even reinforcement learning on proprietary data without moving information to external systems or managing distributed GPU clusters . Snowflake handles the infrastructure sourcing and scaling . The service sits alongside a growing model marketplace in Cortex AI, which already includes models from Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, and the newly announced SpaceXAI models .

The Strategic Bet: The Agentic Enterprise

The connective tissue across every Summit announcement was Snowflake's vision of the agentic enterprise—a state where governed enterprise data is seamlessly connected to AI agents that can reason, act, and automate . By splitting its agent portfolio into CoCo for developers and CoWork for knowledge workers, adding governed real-time streaming with Datastream, and making training custom models a first-party managed service, the company is positioning the AI Data Cloud not just as a place to store and query information, but as the runtime for autonomous business logic .