The MosaicML integration and model democratization
Databricks completed its acquisition of MosaicML to provide a platform for training, customizing, and deploying generative AI applications. This acquisition, valued at roughly $1.3 billion, occurred on June 26, 2023. The company aims to reduce the price of training and customizing large language models through hardware and software efficiency improvements. MosaicML engineers previously developed the MPT-7B model, which reached over 3 million downloads. Databricks combines its AI-ready data management technology with the MosaicML large language models platform to allow businesses to build low-cost language models using proprietary data.
The integration focuses on three specific goals for enterprise generative AI. First, the company works to democratize model capabilities by lowering costs and increasing access to modeling architectures. Second, the team works to make generative AI models function for enterprises by incorporating custom business data like customer accounts or orders. Third, the company unifies the AI and data stack by putting data at the center of the machine learning lifecycle. This includes upstream data preparation such as cleaning, featurization, and embedding. Databricks also uses data to create a positive reinforcement loop by learning from how models perform with generated data.
The talent migration and hiring trends
Databricks is the primary destination for data infrastructure talent during the 2026 industry reshuffle. While search trends mention layoffs, the company has no mass layoff memo or WARN filing in 2026. A small number of targeted role eliminations involving fewer than fifty positions occurred in March across product management and post-acquisition redundancy roles. These cuts represent a tiny fraction of the workforce compared to the 840 open roles listed on the company careers page as of May 2026. The company’s revenue run-rate crossed $5.4 billion in April 2026, showing 65 percent year-over-year growth.
The company sources much of its talent from competitors and previous acquisitions. Databricks hires many senior engineers from Snowflake and Confluent because those employees possess the specific domain knowledge required to manage large-scale enterprise data migrations and complex real-time streaming workloads that demand high availability and strict governance. The company also absorbs founding teams through its acquisitions, such as the MosaicML generative AI infrastructure team, the Tabular Iceberg creators, the Neon serverless Postgres team, and the Tecton engineers. The 840 open roles include approximately 290 field engineering positions, 245 engineering roles, 130 sales positions, and 55 product roles.
The engineering and executive leadership
Ali Ghodsi serves as the chief executive officer and co-founder of Databricks. He joined the company in 2016 and specializes in distributed systems and big data. Matei Zaharia, a co-founder, serves as the chief technology officer and created Apache Spark. He also started the MLflow machine learning lifecycle project. Reynold Xin, a co-founder, works as the chief architect and designed components like GraphX and Project Tungsten. Ion Stoica, a co-founder, serves as the executive chair and specializes in cloud computing and distributed systems.
Patrick Wendell leads engineering for the company’s AI efforts, including products like Genie and Unity Gateway. He manages partnerships with AI model providers and focuses on large-scale analytics systems. Hatim Shafique serves as the chief operating officer and previously held leadership roles at AppDynamics. David Conte has led the financial functions as chief financial officer since October 2019. The company employs more than 17,300 people across nineteen countries.
Agent Bricks and the platform expansion
Databricks expanded Agent Bricks into a full developer agent platform at the 2026 Data + AI Summit. This platform addresses the difficulty of running agents safely at enterprise scale by managing token capacity, deployment, security, evaluation, monitoring, and cost control. Agent Bricks provides model choice across all major frontier providers. The platform supports agent frameworks including Claude Code SDK, LangGraph, Agno, and CrewAI. The 2026 expansion includes support for Kimi from Moonshot AI and Grok from xAI.
The platform includes a meta-orchestration layer called Omnigent, which is an open-source project under Apache 2.0. Omnigent allows teams to compose and share agent workflows without being locked to one tool. Databricks manages a version of Omnigent in Beta that provides shared history, remote access, and governance through the Unity AI Gateway. The platform also includes Agent Memory Services built on Lakebase and a Databricks Sandbox for secure agent code execution. You already know that building a chatbot is simple, but building an agent that follows business rules requires deep integration with governed data.
Genie and the necessity of business context
The Genie product suite focuses on providing business context to AI through Genie One, Genie Agents, and Genie Ontology. Genie One provides a conversational tool over the Model Context Protocol (MCP) within the Unity Gateway. This service allows clients like Claude or ChatGPT to ground data questions in trusted insights while enforcing Unity Catalog permissions. Genie Code enables end-to-end document processing pipelines by orchestrating parsing, classification, extraction, and validation. This tool works with unstructured PDFs, contracts, or invoices to evaluate extraction quality against labeled data.
Genie Ontology attempts to close the gap between a table’s accuracy and the real business meaning of data. This meaning often lives in dashboards, SQL queries, or notebooks rather than in the data itself. Genie Agents let teams create domain-specific agents grounded in this trusted context. Genie App Builder allows users to build governed data and AI apps using natural-language prompts inside an App Space. These apps scale to zero when idle and operate within a governed boundary.
The real-time lakehouse architecture
Databricks is moving real-time capabilities closer to the lakehouse through Lakehouse//RT, Lakebase, and Lakeflow. Lakehouse//RT uses the Reyden engine to provide millisecond performance for operational analytics and BI workloads without moving data away from the lakehouse. This architecture reduces the need for separate serving layers that cause governance drift and engineering overhead. Lakebase supports transactional workloads and allows users to manage snapshots programmatically through a snapshot schedule API.
Lakeflow provides tools for ingestion, transformation, orchestration, and pipeline development. The company uses Lakeflow to bring transactional and analytical processing closer to the governed lakehouse. This connection is vital for AI agents that require current data to remain useful for fraud, supply chain, or finance use cases. The company also provides the Shopify connector in Lakeflow Connect for ingesting e-commerce data like products and orders.
Model availability and retirement schedules
The model landscape on Databricks includes both proprietary and third-party models. The company manages the retirement of older models to ensure users move to more capable versions. For example, the DeepSeek V4 Pro (0813) model retires on October 30, 2026, and users should migrate to DeepSeek V4.1 Flash. The Moonshot AI Kimi K2.7 model also retires on October 30, 2026, with Kimi K3 as the replacement.
| Model Name | Availability Type | Current Status or Replacement |
|---|---|---|
| Meta-Llama-4-Maverick | Pay-per-token | Replacement for DBRX |
| Llama-3.1-70B | Fine-tuning | Replacement for DBRX |
| DeepSeek V4.1 Flash | Foundation Model API | Available |
| OpenAI GPT-6 Sol | Foundation Model API | Available |
| OpenAI GPT-6 Luna | Foundation Model API | Available |
| Anthropic Claude Opus 5.5 | Foundation Model API | Available |
| Moonshot AI Kimi K3 | Foundation Model API | Replacement for Kimi K2.7 |
The DBRX model, which was a Mixture-of-Experts model with 132 billion parameters, reached its end of life in 2025. Users must migrate DBRX workloads to Llama models to avoid service interruptions. The DBRX retirement forces engineers to re-validate prompts and downstream logic tuned to the specific output style of the original model.
Governance and the enterprise verdict
Governance at Databricks extends into the AI runtime through Unity Catalog and the Unity AI Gateway. Unity AI Gateway manages models, agents, MCP services, skills, tools, cost controls, routing, and monitoring. The company also provides Metastore-level ABAC policies in Beta, which allow governance teams to attach row filters or column masks at the metastore level. This prevents the need to replicate policies for every individual catalog.
The company uses Unity Gateway to add coding-agent activity to unified tracing, which shows the tools and skills used alongside model requests. Databricks also provides automatic identity management for Entra ID and Okta support to assist in regulated environments. Will the reliance on third-party model providers eventually erode the margins Databricks builds through its platform? Databricks remains the most effective destination for companies that need to move AI from experimental pilots to production workloads using governed, real-world data.
