Snowflake and Anthropic expand into a $200 million agentic-data partnership
Claude moves deeper into the enterprise data warehouse, with SQL-fluent agents that are audit-logged by default.
The deal makes Claude the default agentic model for Snowflake's managed workloads — data-aware, schema-grounded, and governed at the query level. The two companies describe the expanded partnership as a commitment of two hundred million dollars over an unspecified multi-year term, with co-engineered integrations landing across Snowflake's Cortex AI stack through the second half of 2026.
Enterprise buyers have spent the past year asking a narrow version of the same question: which agent can I trust inside my warehouse. The answer most vendors have offered is a wrapper — a model bolted to a retrieval layerA system that searches a database or document store for relevant information to feed into a language model, anchoring the model's responses in factual data., with governance reconstructed around it. Snowflake and Anthropic are offering something structurally different: Claude instances that run against audited schemas, with every query logged, every tool call attributed, and every output traceable to the row it came from.
The technical detail that matters is row-level attribution. According to briefings shared with early-access customers, Claude's responses inside Snowflake now carry provenance metadataData that tracks the origin and history of a piece of information. In enterprise software, it allows an output to be traced back to the specific document or database row it was generated from. by default — a feature the companies describe as audit-complete rather than opt-in. For regulated industries, that is the distinction between a pilot and a production rollout; several large banks and a pair of pharmaceutical customers are reportedly moving from the former to the latter on the back of it.
The winners are Snowflake, whose AI narrative required exactly this kind of anchor tenant, and Anthropic, which deepens its enterprise flank against an OpenAI product suite that has so far led with developer reach rather than data-governance rigour. The losers are the independent agentic-analytics startups whose addressable market just contracted, and the Databricks-Mosaic stack, which now has a more specific competitive shape to respond to.
What the expansion opens is a template: frontier modelA highly capable, large-scale artificial intelligence model that matches or exceeds the state of the art at the time of its release., plus warehouse-native governance, sold as a single contract. What it forecloses is the assumption that enterprises will tolerate indefinite ambiguity about where their data goes when an agent touches it. The next round of enterprise AI deals, the ones still being negotiated this quarter, will be priced against this one.