Databricks (Mosaic AI)

  Executive Briefing

Databricks is the San Francisco enterprise software company that commercialized Apache Spark and built the modern Lakehouse architecture — and, since mid-2023, it has been systematically transforming itself into one of the most consequential AI infrastructure platforms in the enterprise market. Its AI layer, branded Mosaic AI, spans the full lifecycle from raw data to deployed, governed AI agents, and is notable above all for its refusal to treat data and AI as separate concerns. Where competitors offer standalone model APIs, Databricks bets that the decisive enterprise advantage lies in running model serving, fine-tuning, governance, and orchestration against the same data platform where the training data, feature stores, and business logic already live.

The company's AI ambitions trace directly to the July 2023 acquisition of MosaicML for a reported ~$1.3 billion — the largest AI infrastructure acquisition of that year.¹ MosaicML's founding thesis, that a "mosaic" of efficiency techniques rather than any single approach could dramatically cut the cost of building and owning LLMs, permeates Databricks' current technical posture: a proprietary inference engine with custom CUDA kernels, a serverless-first architecture, and an emphasis on 80%+ cost reductions on open-model serving.² At the helm of Mosaic AI's research organization is Jonathan Frankle, formerly Chief Scientist at MosaicML and now Databricks' Chief AI Scientist, leading a team of 30-plus researchers.

Company-wide leadership remains in the hands of the seven co-founders from UC Berkeley's AMPLab. Ali Ghodsi, CEO and Chairman, steers a company that reported $5.4 billion in annualized revenue as of Q4 FY2026 — growing over 65% year-over-year — with AI products alone accounting for a reported $1.4 billion ARR.³ Matei Zaharia, CTO and creator of Spark, anchors the technical vision. Databricks raised its Series L in December 2025 at a $134 billion valuation, and as of June 2026 is reportedly exploring a further round at $165–175 billion, with CEO Ghodsi indicating an IPO is likely no earlier than 2027.⁴⁵

Mosaic AI's practical significance for AI infrastructure buyers is threefold. First, it is one of the largest neutral serving platforms for open-weight models — hosting Meta Llama, Google Gemma, Alibaba Qwen, and Databricks' own DBRX — with an OpenAI-compatible API, while simultaneously offering gateway access to closed frontier models from OpenAI, Anthropic, Google, and Amazon. Second, its Unity Catalog governance layer applies the same permissions, lineage, and audit framework to models and agents that it applies to data tables — addressing a compliance gap that purely API-focused inference providers cannot fill. Third, Agent Bricks, launched in June 2025, represents a shift from model serving to full agent lifecycle automation, including auto-generated evaluations, synthetic training data, and Test-time Adaptive Optimization (TAO). The result is a platform that positions Databricks not merely as an inference vendor but as the operating system for enterprise AI.

  At a Glance

  Origins & Founding

Databricks was founded in January 2013 by seven researchers from UC Berkeley's AMPLab (Algorithms, Machines, and People Lab): Ali Ghodsi, Andy Konwinski, Ion Stoica, Matei Zaharia, Reynold Xin, Patrick Wendell, and Arsalan Tavakoli-Shirazi.⁶ The company's original purpose was to commercialize Apache Spark, a distributed computing engine Zaharia had created as a faster, more general alternative to Google's MapReduce paradigm. Spark's in-memory processing architecture offered order-of-magnitude speedups over Hadoop for iterative algorithms — particularly machine learning — and the Berkeley team had already demonstrated its research impact before deciding to pursue commercialization.

The founding thesis was that open-source software could serve as both a technical foundation and a distribution moat: Spark's wide adoption in the data engineering community would create an ecosystem of users who might pay for an enterprise-grade managed service. Andreessen Horowitz led a $13.9 million Series A in 2013, validating the bet early.⁶ Over the following years Databricks expanded its platform from Spark compute into a full data analytics stack, eventually articulating the "Lakehouse" architecture — the idea that a single, open-format platform (built on Delta Lake, an open-source ACID table layer over cloud object storage) could replace the traditional two-tier "data lake plus data warehouse" setup.

The AI chapter of the company's story opens with the MosaicML acquisition. MosaicML was founded in 2021 by Naveen Rao (a veteran of Intel's AI division and co-founder of Nervana Systems) and Hanlin Tang, with Jonathan Frankle (a PhD candidate at MIT known for his "Lottery Ticket Hypothesis" work on neural network sparsity) and MIT professor Michael Carbin as founding team members. The company came out of stealth in October 2021 with $37 million led by Lux Capital, DCVC, Future Ventures, and Playground Global, and raised total pre-acquisition funding of under $64 million at a $222 million valuation. MosaicML's central proposition — a mosaic of efficiency techniques could let enterprises build, fine-tune, and own their own LLMs at a fraction of the cost of frontier labs — proved exactly the thesis Databricks needed to extend its data platform into the model era.¹

  History & Timeline

    2013–2018: Spark commercialization and data platform build-out

Databricks spent its first five years productizing Spark and expanding into the adjacent data engineering and analytics layers. It launched the managed Databricks Runtime on AWS and, critically, established the Microsoft Azure Databricks partnership in 2017 — a co-engineering, co-selling arrangement that gave the platform deep integration with Azure infrastructure and introduced Microsoft as an investor. This relationship foreshadowed the enterprise-cloud co-sell model that Databricks later replicated on AWS and Google Cloud. By 2018 the company had established its position as the leading Spark-as-a-service vendor for Fortune 500 data teams.

    2019–2022: Delta Lake, Lakehouse architecture, and the data warehouse war

The 2019 open-source release of Delta Lake — an ACID-compliant storage layer for cloud object stores — changed Databricks' competitive framing from "Spark cloud" to a challenger to traditional data warehouses such as Snowflake and Teradata. Databricks articulated the Lakehouse concept formally in a 2021 CIDR paper, arguing that ACID transactions, schema enforcement, and BI-grade query performance could be delivered on the same open storage layer used for raw data and ML experiments, eliminating redundant data movement. Photon, a vectorized C++-based query engine, was added to deliver warehouse-competitive SQL performance. Series G ($1B at $28B valuation) and Series H (~$1.6B at $38B) rounds in 2021 reflected investor belief in this architecture — and a high-stakes race with Snowflake, which went public in September 2020.

    2023: MosaicML acquisition and the AI pivot

In June 2023 Databricks announced the acquisition of MosaicML for a reported ~$1.3 billion (some sources cite $1.4 billion inclusive of retention packages).¹⁷ The deal closed in July 2023, and MosaicML's technology and team were rebranded as Mosaic AI. Naveen Rao joined as VP of Generative AI; Jonathan Frankle became Chief AI Scientist. The acquisition brought purpose-built LLM training infrastructure (including a 3,072-H100 training cluster), a proprietary composer training library, and enterprise relationships with organizations such as Allen Institute for AI and Replit. Databricks simultaneously deepened its partnership with NVIDIA, bringing NVIDIA in as a strategic investor.

    2024: DBRX, $10B round, and Tabular

In March 2024 Databricks released DBRX, an open-source mixture-of-experts LLM with 132 billion total parameters (36 billion active) trained on 3,072 H100 GPUs over 2.5 months at a reported training cost of $10 million.⁸ DBRX attracted more than 1,000 experimenting customers within two weeks of launch and briefly established Databricks as the operator of one of the strongest openly available base models. In June 2024 Databricks acquired Tabular — the Apache Iceberg table format company co-founded by Apache Iceberg creator Ryan Blue — for a reported ~$2 billion (not officially confirmed), cementing its open-table-format strategy. The company closed a $10 billion funding round at a $62 billion valuation in December 2024.⁴

    2025: Agent Bricks, Lakebase, and Series L

May 2025 saw the closing of the Neon acquisition for a reported ~$1 billion, adding a serverless PostgreSQL engine that Databricks rebranded as Lakebase — a transactional data store purpose-built for AI agents that need OLTP capabilities alongside analytical data.⁹ At the Data + AI Summit 2025 (June 9–12, Moscone Center; 22,000 in-person attendees, 65,000 virtual), Databricks launched Agent Bricks (Beta), MLflow 3.0, AI Runtime GPU serverless, and AI Gateway GA.² September 2025 brought a landmark $100 million multi-year partnership with OpenAI, making OpenAI models natively available on Databricks and — notably — enlisting OpenAI as a Databricks data platform customer for AI training data processing.¹⁰ A parallel $100 million AI integration deal with Anthropic extended Claude models to Databricks customers. Naveen Rao departed in September 2025 to launch a new AI computing startup (Databricks participated as an investor in the new venture), creating a leadership transition within Mosaic AI.¹¹ Databricks' Series L closed in December 2025 at a $134 billion valuation, raising more than $4 billion in equity plus approximately $2 billion in debt capacity — led by Insight Partners, Fidelity, and J.P. Morgan Asset Management, with participation from a16z, BlackRock, Blackstone, Coatue, Goldman Sachs, Morgan Stanley, Qatar Investment Authority, and others.⁴

    2026: $5.4B ARR, positive FCF, and pre-IPO positioning

In February 2026 Databricks reported $5.4 billion in annualized revenue run-rate, growing over 65% year-over-year, with AI products at $1.4 billion ARR and the company generating positive free cash flow over the trailing twelve months — a rare achievement for a company at this scale of investment.³ As of June 2026, the company is reported to be exploring a new fundraise at a $165–175 billion valuation, with CEO Ali Ghodsi indicating that 2026 is a "terrible year" to go public and an IPO is unlikely before 2027.⁵

  What They Offer — Products & Platform

Mosaic AI is Databricks' unified AI development and serving layer, tightly integrated with the Lakehouse/Delta Lake data platform. It is not a standalone inference API: every component is designed to interoperate with the data, governance, and compute infrastructure that enterprise customers already run on Databricks. The platform comprises the following major product components.

Mosaic AI Model Serving is the core inference layer, offering two modes: pay-per-token on-demand endpoints suitable for experimentation and low-commitment workloads, and provisioned throughput endpoints that deliver dedicated capacity with production SLAs.¹² The serving layer handles over 250,000 queries per second across the Databricks fleet. It supports Databricks-hosted open models, external provider models accessed via the AI Gateway, custom ML models (PyFunc, scikit-learn, LangChain), and fine-tuned models — all through a unified OpenAI-compatible API, enabling drop-in migration for teams already using the OpenAI SDK.

Foundation Model APIs provide pay-per-token access to Databricks-hosted frontier and open models from Meta, Google, Alibaba, Anthropic, OpenAI, and Databricks' own DBRX and embedding models. Specific token prices are not publicly listed in a single table; customers are directed to the Databricks pricing page or account teams for current rates.

AI Gateway (generally available as of DAIS 2025) is a centralized governance layer over all model traffic: rate limiting, PII detection, safety guardrails, usage logging, provider fallback, and Unity Catalog permissions. It allows platform teams to enforce company-wide AI policy across both Databricks-hosted and external model calls without requiring individual teams to implement controls independently.

Agent Bricks (Beta, launched June 11, 2025) is Databricks' auto-optimized AI agent builder.¹³ Users specify a task and connect enterprise data sources; Agent Bricks auto-generates evaluations, synthesizes domain-specific training data, and optimizes the quality-to-cost balance using Test-time Adaptive Optimization (TAO) — a proprietary technique that adapts inference-time compute expenditure based on query complexity. Supported use cases include structured extraction, knowledge assistance, text transformation, and multi-agent orchestration.

MLflow 3.0 was redesigned for the generative AI era with agent observability, a prompt registry, and monitoring of agents deployed outside Databricks — including on AWS, GCP, and on-premise environments. MLflow is open-source and one of the most widely adopted MLOps frameworks in the industry.

AI Runtime (Beta as of DAIS 2025, H100s added alongside previously available A10g GPUs) provides serverless GPU compute for training and fine-tuning without long-term capacity reservations. A general availability timeline for H100 serverless had not been confirmed as of the June 2025 announcement.

AI Functions in SQL enable batch inference directly within SQL queries, allowing data engineers to run model calls at scale without leaving the analytics workflow. Databricks reports this runs up to 3x faster and 4x lower cost than competing batch inference solutions based on internal benchmarks, with multimodal support for text and images.

AI Search (Public Preview) is a vector search service scaled to billions of vectors, enabled by a separated compute-and-storage architecture that Databricks reports delivers 7x lower cost than its previous architecture. It targets retrieval-augmented generation (RAG) applications where index size and query cost are the primary constraints.

Lakebase is a serverless PostgreSQL database (built on technology acquired from Neon, May 2025) processing over 10,000 queries per second, designed for AI agents that require OLTP transactional capabilities alongside the analytical data available in the Lakehouse.⁹ Lakebase closes a gap that previously forced customers to run separate transactional databases outside the Databricks environment.

Genie is a conversational AI assistant for data analytics within the Databricks platform, enabling business users to query data in natural language. Databricks positions Genie alongside Lakebase as the next major growth vectors after core data warehousing.

Unity Catalog is the data and AI governance layer spanning tables, models, functions, agents, and endpoints. It applies consistent access control, lineage tracking, and audit logging across the full AI product stack — a differentiator for regulated industries where model governance is as important as model performance.

MCP Support: Databricks has integrated the Model Context Protocol (MCP), with MCP servers hostable via Databricks Apps. Databricks provides hosted MCP servers for Unity Catalog functions, Genie, and AI Search, enabling agentic frameworks that implement MCP to connect to Databricks resources without custom integrations.

  Technology & Infrastructure

Databricks' inference infrastructure is built primarily on NVIDIA H100 Tensor Core GPUs for production serving and training, with NVIDIA A10g GPUs available in the AI Runtime serverless beta tier.² NVIDIA is both a strategic investor in Databricks and an integration partner through the NVIDIA AI Enterprise software stack. The training cluster used for DBRX — 3,072 H100s connected via 3.2 TB/s InfiniBand — gives a reference point for the scale of infrastructure Databricks operates directly.⁸

The most technically significant infrastructure investment is Databricks' proprietary inference engine: a custom in-house serving stack with private optimizations and custom CUDA kernels tuned specifically for Meta Llama and other popular open-source LLMs. Databricks reports this engine delivers up to 1.5x higher throughput than vLLM-v1 on common workloads, and credits it with enabling an 80% cost reduction on Meta Llama 3.3 serving alongside 40% faster response times relative to prior infrastructure.² This is a meaningful departure from the industry norm of deploying vLLM or TGI; Databricks is investing in inference optimization at a layer most enterprise AI platforms outsource.

Multi-cloud deployment is table stakes for the enterprise market. Model serving endpoints are regionally distributed across AWS, Microsoft Azure, and Google Cloud Platform regions, and the platform's storage and compute layers are designed to run on all three clouds without architectural changes. This cloud neutrality is central to the enterprise pitch: customers are not locked into a cloud vendor's AI stack.

The software stack reflects Databricks' open-source heritage: Apache Spark (co-created by the founding team) handles distributed data processing; Delta Lake provides ACID storage; MLflow (now at version 3.0) handles experiment tracking, model registry, and agent observability; Photon is the vectorized SQL query engine with planned NVIDIA GPU acceleration; and Lakebase (Neon-based) adds serverless PostgreSQL for transactional workloads. AI Search uses a separated compute-and-storage architecture that enables cold-start scaling to billions of vectors without maintaining a persistent in-memory index.

The total fleet handling model serving processes over 250,000 QPS and serves more than 10,000 enterprise organizations globally.²

  Model Catalog & Performance

Databricks operates one of the broadest multi-vendor open-model catalogs in the enterprise AI market. Models are hosted directly on Databricks infrastructure (Foundation Model APIs) or accessed via the AI Gateway from external providers. The catalog as of mid-2026 includes the following primary models.

Databricks-hosted open models:

Databricks-hosted frontier models (via partnership):

External models via AI Gateway (pass-through):

OpenAI GPT-4o and o-series, Amazon Bedrock (Nova, Titan), Google Vertex AI (Gemini 2.0), and Cohere Command-R are accessible through the AI Gateway with Databricks-level governance, logging, and access controls applied at the gateway layer.

Earlier Databricks-heritage open models — MPT-7B and MPT-30B (released by MosaicML in 2023 before the acquisition) — remain available in the ecosystem but are not primary catalog entries.

Cross-links: DBRX Instruct, Meta Llama 3.3 70B, Llama 4 Maverick.

  Pricing & Performance Position

Databricks offers two primary pricing modes for model serving. Pay-per-token (on-demand) pricing requires no upfront commitment and is suited for experimentation and variable workloads; specific per-token rates are not published in a single public table and customers are directed to the Databricks pricing page or account teams.¹² Provisioned throughput provides dedicated capacity billed per compute unit with performance guarantees, suited for production workloads with predictable SLAs and the ability to deploy fine-tuned models.

Databricks' public cost-efficiency claims are notable in context:

• 80% cost reduction on Meta Llama 3.3 inference relative to prior serving infrastructure, attributed to the proprietary custom-kernel inference engine.²
• 40% faster responses on Llama 3.3 versus prior serving.
• Up to 1.5x throughput improvement over vLLM-v1 on common workloads from custom CUDA kernels.
• 3x faster and 4x lower cost for AI Functions batch inference versus competing batch inference products (Databricks internal benchmark).
• 7x lower cost for AI Search at billions-of-vectors scale versus the prior AI Search architecture.

These figures are Databricks-reported and should be independently validated for specific workload profiles. The proprietary inference engine is the clearest technical differentiator on cost: most comparable enterprise AI platforms deploy vLLM or TGI rather than investing in custom serving infrastructure.

For enterprise buyers, the total cost of ownership calculation extends beyond per-token pricing to include the value of unified governance (avoiding a separate MLOps platform), the cost savings from serverless scaling-to-zero on the AI Runtime, and the reduced data-movement costs from running inference adjacent to data already in the Lakehouse.

  People & Leadership

Databricks' executive leadership is defined by an unusual degree of founder continuity. All seven UC Berkeley AMPLab co-founders remain in senior roles more than a decade after founding.

Ali Ghodsi (Co-Founder, CEO, Chairman) is the public face of the company and its commercial strategist. He studied parallel computing at Chalmers University and KTH before joining the AMPLab; his research focused on resource management in distributed systems (Dominant Resource Fairness, Mesos). He has led Databricks through every major fundraising round and has consistently deferred the IPO in favor of continued growth investment.

Matei Zaharia (Co-Founder, CTO) is the creator of Apache Spark and one of the most cited figures in distributed systems research. He has maintained a joint appointment at Stanford (previously MIT) while serving as CTO. His technical vision shapes Databricks' open-source strategy and research priorities, including the development of DBRX.

Ion Stoica (Co-Founder) holds a UC Berkeley professorship and co-created Apache Spark alongside Zaharia. He plays an advisory and technical leadership role and is closely involved in infrastructure research, including the distributed inference work that underpins Model Serving.

Jonathan Frankle (Chief AI Scientist) joined via the MosaicML acquisition, where he was Chief Scientist. His academic work — particularly the Lottery Ticket Hypothesis on neural network pruning, published during his MIT PhD — established him as one of the leading researchers on model efficiency. He leads Mosaic Research, a team of more than 30 researchers, pursuing work in reinforcement learning, LLM inference optimization, and enterprise agent systems including KARL (an enterprise RL agent) and Omnigent (a meta-agent harness).¹¹

Naveen Rao (Co-Founder of MosaicML, former VP Generative AI at Databricks) departed in September 2025 to launch a new AI computing startup, in which Databricks participated as an investor.¹¹ His departure created a leadership transition within the Mosaic AI product organization; a formal replacement or restructured reporting had not been publicly documented at research time.

Reynold Xin, Patrick Wendell, and Andy Konwinski remain in senior technical and strategic roles. Arsalan Tavakoli-Shirazi leads go-to-market and field engineering.

  Funding, Ownership & Business

Databricks has raised a reported total of more than $19 billion across approximately twelve funding rounds since 2013.⁴⁶ The trajectory reflects the company's evolution from a Spark compute startup to one of the most highly valued private software companies in history.

Key financing milestones:

• Series A (2013): $13.9M — Andreessen Horowitz.⁶
• Series B–D (2014–2017): ~$233M total — NEA, a16z, Microsoft (alongside the Azure Databricks partnership).
• Series E–F (2019): ~$540M.
• Series G (2021): $1B at $28B valuation — Franklin Templeton, Morgan Stanley.
• Series H (2021): ~$1.6B at $38B valuation.
• Series I (2023): NVIDIA joined as a strategic investor alongside the MosaicML acquisition period.
• December 2024: $10B round at $62B valuation.
• Series L (December 2025): More than $4B in equity plus $2B in debt capacity ($7B total). Led by Insight Partners, Fidelity, J.P. Morgan Asset Management. Participating investors include a16z, BlackRock, Blackstone, Coatue, GIC, MGX, NEA, Ontario Teachers' Pension Plan, T. Rowe Price, Temasek, Thrive Capital, Winslow Capital, Goldman Sachs, Morgan Stanley, Neuberger Berman, and Qatar Investment Authority.⁴
• June 2026 (reported): Exploring a new pre-IPO round at a $165–175 billion valuation; unconfirmed at research time.⁵

The business model is consumption-based SaaS: customers pay per token for on-demand inference, per compute unit for provisioned throughput, and on a usage basis for the broader data platform. Enterprise licensing, annual contracts, and cloud marketplace listings (AWS Marketplace, Azure Marketplace, Google Cloud Marketplace) round out the go-to-market. Revenue metrics as of Q4 FY2026 (February 2026 report): $5.4B ARR, 65%+ YoY growth, greater than 140% net revenue retention, more than 800 customers at greater than $1M ARR, more than 70 customers at greater than $10M ARR, and positive free cash flow over the trailing twelve months.³

  Customers & Partnerships

Databricks serves more than 20,000 organizations worldwide as of early 2026, with 60%+ of Fortune 500 companies as customers.³ Pre-acquisition MosaicML customers included the Allen Institute for AI, Replit, Hippocratic AI, Generally Intelligent, and Scatter Labs — an enterprise and research-organization mix that demonstrated the platform's applicability across domains.

The company's most strategically significant partnership is the September 2025 OpenAI agreement — a reported $100 million multi-year deal that makes OpenAI models (with GPT-5 designated the flagship) natively available on Databricks, while also enlisting OpenAI as a Databricks customer that uses the data platform to process AI training data.¹⁰ This bidirectional arrangement is unusual: Databricks is simultaneously a distribution channel for OpenAI and an infrastructure provider to OpenAI. The parallel Anthropic partnership (reported $100 million AI integration deal) makes Claude models available directly on Databricks infrastructure, reflected in the databricks-claude-* model catalog entries.

Additional strategic partnerships include:

• Google Cloud: A four-year Gemini integration partnership; Gemini 2.0 models available via AI Gateway.
• AWS: "Buy with AWS" Marketplace listing and advanced generative AI partnership; Amazon Bedrock models (Nova, Titan) accessible via AI Gateway.
• NVIDIA: Strategic investor; NVIDIA AI Enterprise software integration; GPU supply partnership for H100 training and serving infrastructure.
• Meta: Databricks among the launch partners for successive Llama releases; deep integration of Llama 3.x and Llama 4 series into Foundation Model APIs.
• Alibaba: Qwen model family available in Foundation Model APIs under preview and general access tiers.

  Competitive Position

Databricks Mosaic AI occupies a distinctive position in the enterprise AI infrastructure market — neither a pure inference API provider nor a hyperscaler, but an integrated data-plus-AI platform that leverages its existing data estate to make model serving, governance, and orchestration stickier than standalone alternatives.

Its most direct competitor is Snowflake Cortex AI, which is pursuing an analogous strategy from the data warehouse side: adding LLM inference, fine-tuning, and agent capabilities to an existing enterprise data customer base. Both companies are expanding from their data platform strongholds into AI, and both emphasize governance and data integration as differentiators over raw inference price. Snowflake has historically had broader enterprise penetration; Databricks has a stronger open-source ecosystem and a more capable standalone AI infrastructure (custom inference engine, MLflow, Unity Catalog AI governance).

Hyperscalers — AWS SageMaker and Bedrock, Azure AI Foundry, Google Vertex AI — offer AI serving embedded in cloud infrastructure with deep pricing leverage and existing enterprise relationships. Databricks' counter is cloud neutrality (runs on all three) and the argument that cloud-native AI tools require data movement off the Lakehouse, whereas Databricks keeps compute and data co-located.

Pure inference API providers (Together AI, Fireworks AI, Groq, Cerebras) offer lower raw inference cost and lower latency for single-model serving, but lack the governance, fine-tuning workflow, and data integration that regulated enterprise buyers require. Databricks claims to close the cost gap with its custom inference engine while adding these enterprise layers.

Differentiation summary: Unified data-plus-AI governance via Unity Catalog; proprietary inference engine with custom CUDA kernels (1.5x vLLM throughput); open-source ecosystem (Spark, Delta Lake, MLflow) as distribution and lock-in mechanism; model-agnostic serving (open + closed models via a single API); and the OpenAI and Anthropic partnerships that provide access to frontier closed models within the same governance perimeter. Acknowledged risk areas: Naveen Rao's departure created uncertainty around AI strategy continuity;¹¹ hyperscalers have deeper infrastructure pricing leverage; Snowflake's enterprise penetration remains an ongoing sales motion challenge.

  Notable Events

• July 2023: MosaicML acquisition closes for a reported ~$1.3B; rebranded as Mosaic AI.¹
• March 2024: DBRX released as open-source MoE LLM (132B total / 36B active params); trained on 3,072 H100s for a reported $10M; 1,000+ experimenting customers within two weeks.⁸
• June 2024: Tabular acquired for a reported ~$2B (not officially confirmed) — co-founded by Apache Iceberg creator Ryan Blue.
• December 2024: $10B funding round at $62B valuation.
• May 2025: Neon acquisition for a reported ~$1B; Lakebase product launched.⁹
• June 9–12, 2025: Data + AI Summit 2025 at Moscone Center, San Francisco (22,000+ in-person; 65,000+ virtual); Agent Bricks Beta, MLflow 3.0, AI Runtime GPU serverless, and AI Gateway GA announced.²
• September 2025: Naveen Rao (VP Generative AI, MosaicML co-founder) departs to launch a new AI computing startup; Databricks invests in the venture.¹¹
• September 25, 2025: OpenAI partnership announced — $100M multi-year deal; GPT-5 designated flagship for Databricks customers.¹⁰
• December 2025: Series L closes at $134B valuation with more than $7B total capital raised.⁴
• February 2026: Databricks reports $5.4B ARR, 65%+ YoY growth, and positive free cash flow.³
• June 2026 (reported): Exploring new fundraise at $165–175B valuation; IPO guidance points to 2027 at earliest.⁵

  Outlook & Roadmap

Databricks is executing a multi-vector expansion strategy that positions AI as the primary growth engine, with the data platform as the durable moat that keeps customers integrated.

Agent Bricks and the agentic shift: The launch of Agent Bricks in June 2025 marks the clearest articulation of Databricks' long-term AI product thesis — that the competitive advantage in enterprise AI lies not in offering cheaper or faster tokens but in automating the agent lifecycle: evaluation generation, training data synthesis, and inference-time optimization through TAO. If this thesis proves correct, Databricks becomes the platform where enterprises build and own their AI agents, not merely consume model APIs.

Lakebase and Genie as growth levers: CEO Ghodsi has pointed to Lakebase (serverless Postgres for AI agents needing transactional state) and Genie (conversational analytics) as the next major product-led growth opportunities beyond data warehousing. Lakebase in particular closes a gap that previously required customers to maintain separate OLTP infrastructure outside Databricks.

Research roadmap: Mosaic Research (30+ scientists led by Jonathan Frankle) is pursuing reinforcement learning for enterprise agents (KARL), a meta-agent harness (Omnigent), and continued LLM inference optimization. MLflow 3.0's support for monitoring agents outside Databricks signals an ambition to become the observability layer for enterprise AI regardless of where models are deployed.

Frontier model access: Continued expansion of the model catalog through partnerships with OpenAI (GPT-5 as flagship), Anthropic (Claude Opus, Fable, Sonnet series), Google (Gemini), Meta (Llama 4 and beyond), and Alibaba (Qwen) — all within the Unity Catalog governance perimeter. The strategic bet is that model-agnostic access plus superior governance is more durable than any single model partnership.

MCP and agentic interoperability: The MCP integration positions Databricks to benefit from the broader agentic ecosystem adopting the Model Context Protocol as an interoperability standard, potentially making Databricks data and AI assets accessible from any MCP-compatible agent framework.

IPO and capital structure: CEO Ali Ghodsi indicated in June 2026 that 2026 is a "terrible year" to go public given crowded market conditions (including SpaceX and other high-profile listings), with an IPO likely no earlier than 2027. The reported exploration of a new pre-IPO round at $165–175 billion would provide additional capital for M&A — Databricks has stated it intends to continue acquisitions with the Series L capital — and extend the runway before public markets.⁵

Open questions: DBRX's competitive position relative to current-generation models (Llama 4, GPT-5) has likely evolved since its March 2024 release; the AI Runtime H100 serverless GA timeline remains unconfirmed; and the organizational structure of Mosaic AI post-Naveen Rao has not been fully publicly documented.

---

  References

1. Databricks acquires MosaicML — Press Release
2. Mosaic AI announcements at Data + AI Summit 2025
3. Databricks surpasses $5.4B revenue run-rate — Press Release
4. Databricks Series L funding — CNBC
5. Databricks reportedly eyes $175B valuation — TechFunding News
6. Databricks — Wikipedia
7. Databricks acquires MosaicML for $1.3B — TechCrunch
8. DBRX — Wikipedia
9. Databricks launches Lakebase — SiliconAngle
10. Databricks and OpenAI partnership — Press Release
11. Databricks AI strategy after Naveen Rao — InfoWorld
12. Databricks Model Serving product page
13. Databricks launches Agent Bricks — Press Release
14. Foundation Model APIs overview — Microsoft Learn
15. Databricks $4.8B ARR run-rate — Press Release

  References

1. MosaicML acquisition announcement and TechCrunch coverage — Databricks PR; TechCrunch. Some sources cite $1.4B inclusive of retention packages. ↩ ↩² ↩³ ↩⁴

2. Mosaic AI announcements, Data + AI Summit 2025 — Databricks Blog. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. Revenue and ARR figures — Databricks Press Release, Feb 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

4. Series L funding details — CNBC, Dec 2025. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

5. $165–175B valuation report and IPO timeline — TechFunding News. Reported; unconfirmed at research time. ↩ ↩² ↩³ ↩⁴ ↩⁵

6. Databricks founding and early funding history — Wikipedia: Databricks. ↩ ↩² ↩³ ↩⁴

7. MosaicML acquisition consideration — $1.3B or $1.4B inclusive of retention; discrepancy across sources, not officially clarified by Databricks. ↩

8. DBRX training details — Wikipedia: DBRX. Training cost of $10M is Databricks-reported. ↩ ↩² ↩³ ↩⁴

9. Lakebase and Neon acquisition — SiliconAngle. Neon acquisition price of ~$1B is reported, not officially confirmed. ↩ ↩² ↩³

10. OpenAI partnership — Databricks Press Release. ↩ ↩² ↩³

11. Naveen Rao departure and AI strategy continuity — InfoWorld. ↩ ↩² ↩³ ↩⁴ ↩⁵

12. Model Serving pricing modes — Databricks Model Serving. Specific per-token rates not publicly listed; check Databricks pricing page directly. ↩ ↩²

13. Agent Bricks launch — Databricks Press Release. ↩