AI data stack starts selling control, not magic

The clearest signal in data and analytics this week was not another model claim. It was the move to make AI look governable. On 25 Aug, Snowflake’s CEO argued publicly that trusted data, rather than model benchmarks, will define enterprise AI. On 26 Aug, Snowflake joined Reuters to offer AI-ready trusted news content on its Marketplace, ended the use of service-account passwords, and was reported to be acquiring Natoma for secure connectivity, with no deal value disclosed.

Databricks moved in the same direction from a different starting point. On 26 Aug and again in coverage on 28 Aug, it introduced Governance Hub for account-level governance across Databricks estates. Its homepage also changed on 26 Aug from “The database your AI agents deserve” to “One database for AI, apps and agents”, while “The Databricks Data and AI Platform” replaced the broader “The Databricks Platform”. That is a tighter sell: not just agents, but the data system agents run on.

The market implication is simple. Enterprise AI positioning is settling around custody, policy, lineage and operational trust. Snowflake framed it as trusted data. Databricks framed it as governed estates plus a database for AI, apps and agents. The language differs, but the direction is the same.

Databricks had the louder fortnight. Funding reports appeared repeatedly between 24 Aug and 29 Aug, including a $5bn round at a $190bn valuation, with one 24 Aug report also citing a revenue run rate exceeding $7bn. On 24 Aug, Databricks was also reported to have acquired Electric to enhance its ability to create faster AI agents.

The product surface widened at the same time. On 24 Aug, Databricks released databricks-dspy 0.1.0 and launched repositories including databricks-ai-bridge, zerobus-sdk, appkit, devhub, databricks-sdk-java and databricks-vscode. On 25 Aug, it released databricks-sql-go 1.14.0, raising the minimum Go version to 1.25.0 and addressing OSV-Scanner findings. On 26 Aug and 29 Aug, more repositories appeared, including databricks-jdbc, sdk-js, databricks-sdk-go, databricks-sdk-java and databricks-sql-kernel-bindings.

The pricing copy moved too. On 27 Aug, Databricks added Learn Spark to the Free Edition tier and removed Professional Data and AI tools from that tier. The Pay as you go plan changed from “Custom” to “Contact us”, while per-second granularity was added and “Pay for products you use” was removed from observed copy. The pattern is not accidental. Databricks is courting developers with SDKs and connectors, while tightening the commercial packaging around serious usage.

Snowflake’s week was strong on enterprise AI narrative and busy on developer artefacts. On 25 Aug, Snowpark Python API 1.54.0 added ai_count_tokens and ai_multi_embed. On 26 Aug and 27 Aug, Snowflake added repositories including cortex-training, snowflake-connector-net, homebrew-snowconvert-ai, gosnowflake, snowflake-sqlalchemy and snowflake-ingest-java. On 29 Aug, further repositories appeared, including libsnowflakeclient and ArcticInference.

The same period also brought operational noise. On 24 Aug, Snowflake’s status page showed a major incident, INC20000173, resolved after about 101 minutes. On 25 Aug, a critical incident, INC20000175, lasted around 42 minutes before resolution. On 28 Aug, Snowflake reported one major incident, INC20000182, resolved after approximately 81 minutes, alongside a minor incident that was identified. Also on 24 Aug, two articles detailed vulnerabilities in Snowflake’s internal systems, including one article that said an AI agent found a flaw that was exploited.

For a company leaning into trusted data and enterprise AI, these are material facts rather than a narrative break. Trust is now both the marketing promise and the operating burden. Snowflake also changed Virtual Private Snowflake pricing copy on 25 Aug, with observed language both adding and removing references to all Business Critical Edition features. In a category where security and governance are becoming the product, small packaging edits and status-page events carry more weight than usual.

Sigma Computing showed the most visible messaging volatility. On 26 Aug, its homepage changed from “AI runtime for business” to “Vibe-code enterprise business applications”. On 27 Aug, it changed back to emphasise “AI runtime for business”, replacing the vibe-code framing. Around that move, Sigma launched AI-heavy content: six blog posts on 26 Aug, four more on 27 Aug, two on 28 Aug, and then removed the “Sum: The New Analytics Skill Set - What Matters in an AI World” resource on 29 Aug.

That 24-hour homepage reversal is the sort of timing gap operators usually miss if they only read launches. It matters because AI positioning is still being tested in public. “Runtime”, “vibe-code”, “agents”, “self-service BI” and “governed apps” are not interchangeable. They imply different buyers, different proof points and different competitors.

dbt Labs made the same contest more concrete in packaging. On 26 Aug, dbt Wizard was added to the Starter tier, dbt Copilot code generation was removed from Starter, dbt Copilot advanced was added to Enterprise, and dbt Copilot was removed from Enterprise in observed pricing copy. Its content shifted towards summit pages, AI-augmented analytics and trust: on 27 Aug it added “Scaling AI is Easy; Trusting It is Hard”, and on 28 Aug it added “Why Your AI Pilot Stalled at the Context Gap”.

Fivetran, by contrast, looked more operational than rhetorical. On 24 Aug, dbt_shopify 1.9.2 added DuckDB as a supported destination. On 27 Aug, dbt_zendesk 1.7.1 did the same. On 29 Aug, dbt_quickbooks 1.9.1 added DuckDB support. The documentation stream also skewed practical, with pages on live mode sync, Chorus troubleshooting, Shopify HTTP 403 errors, Snowflake save button issues and daylight saving time discrepancies appearing across 25 Aug to 29 Aug.