Snowflake reported FY2026 product revenue of $4.47 billion, a 29% increase. In the quarter ended April 30, 2026, product revenue jumped 34% to $1.334 billion. This growth surpassed management’s own guide by $69 million. Management tied this reacceleration to AI consumption. Databricks reports an AI products line at a $1.4 billion run-rate with 65% annual growth. I observe that both companies reach similar scales by converging on a single product category: governed analytics plus AI on cloud object storage. Snowflake was a SQL data warehouse; Databricks was a Spark-based data lake. By 2026, Snowflake runs open-format tables and a full AI stack, while Databricks runs a SQL warehouse product at a run-rate above $1 billion. Snowflake serves 13,328 customers, including 790 from the Forbes Global 2000. Databricks maintains a $5.4 billion revenue run-rate and net retention above 140% with over 800 customers spending over $1 million a year. The market for data and analytics software exceeds $650 billion by 2028 according to IDC. I see that the demand for large language models shifts focus from simple chat to the data plumbing required for retrieval-augmented generation. You already know the basics, so I will focus on the consumption metrics. Snowflake ended its fiscal year with $4.4 billion in cash. Databricks reached a $134 billion valuation following a Series L funding round in late 2025. The convergence of these platforms means the distinction between a warehouse and a lakehouse has effectively vanished.
Scaling inference with Arctic
I examine the efficiency of Snowflake Arctic against Databricks deployment models. Arctic utilizes a Mixture-of-Experts architecture with 17B active parameters. This design provides advantages during inference. At a batch size of 1, Arctic achieves a throughput of over 70 tokens/second with FP8 quantization. This performance makes it viable for interactive serving. Arctic has up to 4x less memory reads than Code-Llama 70B and up to 2.5x less than Mixtral 8x22B.
| Metric | Snowflake Arctic | DBRX / Llama 3 70B |
|---|---|---|
| Active Parameter Comparison | 50% fewer than DBRX | N/A |
| Parameter Comparison | 75% less than Llama 3 70B | 100% |
| Batch Size 1 Throughput | 70+ tokens/second | N/A |
Databricks deploys DBRX as an NVIDIA NIM microservice. I find that Snowflake Arctic focuses on business tasks like SQL generation and object creation. Arctic excels in query writing, query explanations, and Snowpark scripts. It also facilitates data sharing in cleanrooms. Snowflake worked with NVIDIA to provide preliminary implementations of Arctic for interactive inference. While Arctic scales for enterprise concurrency, it requires significant GPU power for large batch sizes. Databricks uses Mosaic AI for model deployment and monitoring. The MosaicML acquisition allowed Databricks to position itself as a platform where enterprises train custom models on proprietary data. At large batch sizes, Arctic incurs 4x less compute than CodeLlama 70B and Llama 3 70B. This efficiency helps manage costs during high-volume inference. Databricks continues to expand its partnership with NVIDIA to improve its GPU-backed consumption.
Deployment capability becomes the moat
The competition moves from architecture to deployment capability. Snowflake builds its strategy around the context layer through Cortex and CoCo. These tools aim to make AI answers more trustworthy by managing semantics and verified queries. Databricks pursues a different path by integrating the full lifecycle through Mosaic AI and Unity Catalog. Databricks uses Lakebase to provide a managed Postgres service that integrates directly with the Lakehouse. This collapses ETL pipelines for users building AI-powered apps. However, Databricks’ user experience with Spark-based pipelines still requires significant platform engineering attention to manage GPU scheduling and driver compatibility. I see that the competitive battle between these two companies shifted from the architectural debate that defined the early days of the lakehouse versus warehouse discussion to a sophisticated competition for which platform becomes the default substrate for AI-era data workloads. While Databricks manages model training and deployment through Mosaic AI, Snowflake targets the context layer by providing access to foundation models from providers like OpenAI and Anthropic through the Snowflake platform. The Fivetran-dbt merger in June 2026 created a single ingestion layer for trusted AI agents, aimed at simplifying data infrastructure. The ability to move from pilot to production depends on whether a platform manages the data cleaning and governance steps required for an agent to function. The market landscape for 2026 shows that the most successful users pair platform adoption with phased migration, prioritizing governance-critical workloads before performance optimizations. This approach prevents the duplication of data and reduces the governance tax paid when using multiple catalogs. Can enterprises achieve this without a custom orchestration tier?
