The data stack is becoming an AI control market

Capital is rewarding the control plane

The fortnight was not quiet at the top of the stack. On 2 Sep, Databricks completed a $5 billion funding round at a $190 billion valuation, alongside a reported revenue run-rate above $7 billion. On 3 Sep, further coverage said the company had crossed a $7 billion revenue run-rate, while its press cycle also included AgentOps, Genie One, government secure AI and Deloitte’s acquisition of Wavicle to deepen data and AI engineering work around Databricks.

Snowflake had its own public-market moment. On 3 Sep, its Q2 earnings showed adjusted EPS of $0.62 against a FactSet estimate of $0.45, and the company raised its annual revenue forecast. Coverage on 4 Sep and 6 Sep linked the reaction to AI and cloud demand, with the stock reported up 16% after the earnings report and later up about 22% after the raised outlook. On 6 Sep, Snowflake also appointed Adrian De Luca to lead APJ solution engineering, and the same cluster of reporting cited 37% product revenue growth.

The lesson is blunt: investors are not paying for generic data infrastructure stories. They are paying for platforms that can argue they sit where AI workload growth, governance and enterprise distribution meet. Databricks made the private-market case with funding, partner activity and developer surface area. Snowflake made the public-market case with earnings, forecast language and visible customer proof.

AI is being packaged as governance, not magic

Databricks spent the period turning the agent narrative into an enterprise operations narrative. On 2 Sep, it announced new Genie One features aimed at moving insights into action and was covered for work with OpenAI to integrate GPT-5.5 into enterprise workflows. On 3 Sep, it released the Big Book of AgentOps. On 6 Sep, Yahoo Finance covered agent security for the Databricks Unity AI Gateway from the creators of Apache Ranger.

That matters because the category’s centre of gravity is shifting from model access to control surfaces. The Databricks articles on eliminating $1 million in wasted AI agent spend, governance in lakehouse architecture and Apache Ranger security enhancements all point in the same direction. The sales object is no longer just a data platform. It is a managed execution environment for AI systems that finance, security and data teams can tolerate.

Snowflake is using a parallel playbook. On 5 Sep, coverage tied its Q2 revenue jump to AI and cloud demand and referenced CoCo, its AI coding agent. On the same day, Snowflake added Pacific Life to its public customer wall and published a case study on Pacific Life modernising data and AI with Snowflake CoCo. On 31 Aug, it added United Rentals as a customer logo and published a case study on trusted real-time answers across 1,600 plus locations.

The most revealing moves were small and simultaneous

The quieter surfaces moved in formation. Fivetran changed documentation every day from 1 Sep to 5 Sep, adding troubleshooting pages for sources including Gong, MongoDB, Salesforce, SAP SQL Server CDC, Snowflake, GitHub, Slack app syncing, SQL Server query timeouts and Kubernetes deployment permissions. In the same run, it removed older troubleshooting pages tied to destination issues and permissions, while adding repositories including connector_sdk, community_connectors and multiple dbt packages.

dbt Labs made a different kind of content rotation. From 1 Sep to 5 Sep, it added a sequence of Summit pages covering the event experience, agenda sessions, speaker information, deployment optimisation, conversational analytics, scaling on Amazon Redshift and topics such as Dashboards as Code with dbt Charts. Older pages were removed across the same window. This was not a product launch week for dbt Labs. It was a positioning week around the post-AI data stack and the Summit as the distribution vehicle.

Sigma’s messaging moved even faster. On 2 Sep, its hero headline changed from “AI runtime for business” to “Vibe-code enterprise business applications”. On 3 Sep, it changed again to “Go build it. IT approves.”, with the subtext shortened to “Analytics, apps, and agents on one trusted platform.” That two-step matters. The category is still searching for language that lets business users build while reassuring IT that governance has not been bypassed.

Pricing language also tightened. On 31 Aug, Databricks added a new Pay-As-You-Go tier and removed the older “pay as you go” tier. On 2 Sep, it added “Per second granularity” to Pay-As-You-Go and removed “Pay for what you use”. On 4 Sep, Snowflake changed Virtual Private Snowflake from “Contact us” to “Custom” and added that it includes all Business Critical Edition features while being isolated from all other Snowflake accounts. These edits are not theatre. They are the paperwork of enterprise packaging.

Reliability is the tax on consolidation

The same fortnight that rewarded platform breadth also exposed the operational burden of becoming the place where more work lands. Snowflake recorded a critical incident on 31 Aug, INC20000188, affecting services for about 144 minutes before resolution. On 2 Sep, it recorded a major incident, INC20000190, for about 232 minutes. On 5 Sep, another critical incident, INC20000199, affected services for around 156 minutes.

There was broader status-page noise across the category. On 5 Sep, dbt Labs recorded three incidents, including a critical outage in which the dbt Platform was not responding for approximately 91 minutes, plus minor issues around trial onboarding and delayed metadata ingestion with Discovery API latency. On 4 Sep, Sigma reported three incidents, including a critical issue lasting about 80 minutes that prevented GCP-hosted organisations from loading.

Security also stayed visible. On 2 Sep, four CVEs from the previous 30 days were reported as mentioning Snowflake, including CVE-2026-72750, a SQL injection vulnerability in the Snowflake node’s Execute Query operation with a CVSS score of 8.8, and CVE-2026-19594, involving insufficient input sanitisation in the Snowflake Python API. The market context is simple: the more these platforms become AI execution and application layers, the more buyers will price uptime, isolation and vulnerability response into platform selection.