The $6.9 billion annual recurring revenue figure for Databricks as of June 2026 signals a platform that has moved well beyond simple data warehousing. I see teams treat the "Mosaic" brand as a single monolithic product, which leads to immediate budgeting errors. The reality is that Mosaic includes distinct tools like AI Vector Search, AI Model Serving, and AI Pretraining, each with different consumption patterns. Confusing these components causes organizations to overestimate costs for basic inference while underestimating the expense of fine-tuning custom LLMs on proprietary data.
Pricing traps in the serverless tier
The serverless architecture fully abstracts the cloud environment by automatically provisioning, scaling, and terminating compute resources without manual tuning. This eliminates the need to size clusters, but it introduces a variable cost structure that many finance teams fail to model correctly. In March 2024, Databricks released the DBRX foundation model, which cost $10 million to create and uses a mixture-of-experts architecture with 136 billion parameters. While DBRX only activates an average of 36 billion parameters per output, the underlying infrastructure required to host and serve such a large model demands significant compute allocation. Teams often assume that because the platform handles scaling, they can ignore usage metrics, yet the separation of compute and storage means you pay for active processing time regardless of how much data sits idle in your own cloud storage using open-source formats like Delta Lake.
I would skip any attempt to predict exact monthly bills based solely on user count, as the platform charges for actual query execution and model inference. The partnership with Anthropic, valued at $100 million, integrates external AI products into the platform, adding another layer of potential cost if you rely on those specific models rather than native ones. Similarly, the September 2025 deal with OpenAI incorporates their LLMs, which may have different pricing tiers than Databricks’ own models. If you build agents using Agent Bricks, introduced at the 2025 Data + AI Summit, you must account for the "harness" overhead that wraps AI models into functional agents. This wrapper adds latency and compute load that does not appear in standard database query logs. You need to monitor the specific workload types, distinguishing between batch analytics and real-time agent interactions, because the latter consumes resources continuously while the former spikes intermittently.
Governance gaps in Unity Catalog
Unity Catalog provides a unified data governance model where cloud administrators configure coarse access control permissions, and Azure Databricks administrators manage team-level privileges. A common mistake is assuming that granting query access to a table automatically secures the AI agents that consume that table. In June 2024, Databricks open-sourced Unity Catalog under the Apache Spark 2.0 license, making it a standard for unified governance. However, when you define business KPIs once as metric views, you must ensure that Genie Agents, which use AI tailored to your organization’s terminology, respect the same ACLs as human users. If an agent queries data through a semantic layer, it bypasses direct table permissions unless explicitly restricted.
The platform allows sharing outside your secure environment via a managed version of OpenSharing, but this feature requires careful configuration to prevent data leakage. Administrators often set up broad permissions for development environments and forget to tighten them for production AI workloads. With Lakewatch, released in March 2026, Databricks offers an AI-powered agentic security platform to automate threat detection, but this tool must be integrated with existing identity management systems to function effectively. The division of responsibility between cloud-native IAM and Unity Catalog ACLs creates a gap where misconfigured roles can expose sensitive data to unauthorized AI models. I recommend auditing every agent’s access path to verify that it inherits the strictest permission level from both the cloud provider and the catalog. Does your current setup distinguish between a human analyst querying a dashboard and an autonomous agent executing a complex multi-step reasoning task?
