The data stack becomes an agent stack

Databricks bought itself room to define the category

Databricks set the tempo. On 23 Aug, it completed a $5bn funding round at a reported $190bn valuation, taking its total private capital raised to $25bn. Coverage observed on 22 Aug also reported a revenue run-rate exceeding $7bn. In a category where public-market scrutiny has often compressed ambition into margin talk, Databricks has chosen the opposite posture: more capital, more surface area, and a louder claim on AI infrastructure.

The product message moved in the same direction. On 22 Aug, Databricks changed its homepage headline to “The database your AI agents deserve”, with Lakebase described as “serverless Postgres for applications that scale”. That is not a lakehouse line. It is an applications line. The same week brought press around Document Intelligence for complex document extraction, Feature Store advances for real-time data serving, a Sabesp case study, and an investment in UK-based Advancing Analytics.

The hiring pattern supports the push. On 22 Aug, Databricks had 819 active job postings, including 601 in engineering, 123 in sales and 40 in product. The company is not merely selling a warehouse-adjacent AI story. It is staffing as if the next battleground is the operational runtime around data, models, documents and applications.

The homepage war converged on agents

The notable move this fortnight was not that one vendor adopted agent language. It was that several did it at once. On 22 Aug, Fivetran changed its hero to “Automated data for autonomous agents” and its subtext to “The data foundation for AI”. On the same date, Snowflake moved to “Code Work” and “Bring agentic AI to all your data”. Databricks made the agent database claim. Sigma changed to “AI runtime for business”, with copy positioning itself as the layer to build and scale analytics, apps and agents on trusted data.

That simultaneity matters. The modern data stack is no longer being narrated as extract, transform, model and analyse. The vendors are trying to occupy the layer where agents take action on governed data. The vocabulary is converging faster than the products are being cleanly distinguished.

Sigma showed the instability of that transition in miniature. On 22 Aug, its homepage said “AI runtime for business”. On 23 Aug, it flipped to “Vibe-code enterprise business applications”. On 24 Aug, it reverted to “AI runtime for business”. That 48-hour oscillation is a useful tell. The market wants application-builder upside, enterprise buyers still need a sober control-plane story, and the copy has not yet settled.

Snowflake is making AI economics a front-door issue

Snowflake’s week was less about a single splash and more about removing friction from the AI argument. On 22 Aug, it introduced dynamic model routing, with coverage in SD Times, Techzine Global and Verdict. The feature allows the Cortex AI gateway to select models, and the stated aim in the observed coverage was better AI economics and efficiency.

The developer surface moved too. On 23 Aug, Snowflake’s snowpark-python library reached version 1.54.0, adding functions including `ai_count_tokens` for estimating token counts and `ai_multi_embed` for generating multimodal embeddings. Those are small but telling additions. They turn cost estimation and embedding generation into first-class workflow primitives, which fits the wider push to make AI on enterprise data feel governed rather than experimental.

The commercial surface changed at the same time. On 23 Aug, Snowflake added homepage calls to action for “Why Snowflake” and “View Pricing”, while removing “Sign in” and “CONTACT SALES” from the observed navigation state and adding “start for free”. On 22 Aug, Lee Yong-seok was named head of Snowflake in Korea, with a focus on data and AI growth. The message is consistent: broaden access, explain value, and localise the AI enterprise push.

The rest of the stack is being pulled upward

dbt Labs used the fortnight to concentrate attention on its community and core roadmap. On 22 Aug, its homepage shifted to “Your next level starts at dbt Summit”, with calls to register, view the agenda, read the announcement, start developing and read Fusion docs. The same date brought YouTube content on dbt Core v2, dbt State, faster parsing and a new Roundup format.

That is a different motion from Databricks or Snowflake, but it sits in the same market pressure. If agents are the new buyer-facing story, the transformation layer has to defend its role as the governed place where data products become reliable enough for those agents to use. dbt is not leading with agent copy in the observed events. It is leading with Summit, Core v2 and Fusion documentation.

Fivetran’s repositioning is cleaner and more direct. On 22 Aug, it called itself “The data foundation for AI” and placed autonomous agents in the hero line. At the same time, it had 242 active job postings, primarily in engineering and sales. Separate coverage on 17 and 18 Aug detailed a DOJ probe into Andreessen Horowitz, focused on potential antitrust issues related to board seats at Databricks and Fivetran. The company’s public product message is moving towards AI infrastructure while the governance environment around the category attracts greater scrutiny.