Data Warehouses in 2026: The State of the Market

Last updated: 25 September 2026. Verified against vendor sites and announcements on that date.

Quick answer:

The data warehouse market in 2026 is a 4-name race (Snowflake, BigQuery, Databricks, Redshift) sitting on top of an increasingly open storage layer. Ownership consolidated hard in 2024 and 2025: Fivetran and dbt Labs announced a merger, Databricks bought Tabular and Tecton, Metaplane went to Datadog, and a chunk of the “modern data stack” categories collapsed into features of bigger platforms. The interesting frontier is now the AI stack next to the warehouse, not the warehouse itself.

Who bought whom (the 2024-2025 consolidation)

The single biggest change since the last edition of this guide is that the middle of the stack is getting eaten. A partial list of the deals that reshape buying decisions:

• Fivetran + dbt Labs announced a merger in October 2025 at roughly $600M combined ARR. Ingestion and transformation now sold as one story.
• Fivetran acquired Census in May 2025, bringing reverse ETL in-house. It also acquired HVR in 2021 for the enterprise CDC engine.
• Databricks acquired Tabular in 2024 (bringing the founders of Apache Iceberg on board) and Tecton in 2025 (feature stores fold in).
• Metaplane went to Datadog in 2025. Data observability starts becoming a checkbox on general observability platforms.
• Weights and Biases went to CoreWeave in 2025. ML experiment tracking joins the GPU cloud story.
• ClickHouse acquired PeerDB in 2024, tying Postgres CDC directly into ClickHouse Cloud.
• MongoDB acquired Voyage AI in 2025, tying embeddings and rerank models to the operational database.

Two takeaways. Vendors that started as single-primitive tools (reverse ETL, observability, feature stores) either got acquired or expanded into platforms. And the warehouse vendors themselves haven’t been acquired: they’re the buyers.

What technically won

Three things settled decisively.

Separation of storage and compute is now table stakes. Every serious cloud warehouse charges the 2 meters independently. The single-node MPP appliance era is done for new deployments.

Apache Iceberg is the winning open table format. Snowflake, BigQuery, and Databricks all read Iceberg tables natively now, which means storage in your own cloud bucket can serve every engine. Databricks’ acquisition of Tabular in 2024 was the tell: even the vendor pushing Delta Lake accepted that Iceberg had won the standard war.

ELT beat ETL for analytics workloads. Data lands raw in the warehouse first, transformations happen there in SQL, orchestrated by dbt or SQLMesh. Legacy ETL still runs for compliance and cost reasons; nobody starts a new analytics stack that way.

What died (or is dying quietly)

• Standalone Hadoop. Cloudera still exists, but nobody starts a new project on plain HDFS. The lake-plus-format pattern replaced it.
• The dedicated CDP as a category. Segment still ships, but composable CDPs built on the warehouse (Hightouch, Fivetran with Census) took the mid-market conversation. Twilio’s write-down on Segment was public.
• The vector database gold rush. The category still exists but pgvector inside Postgres and native vector search inside every warehouse ate the middle of the market. Rockset was bought by OpenAI and shut down in 2024.
• Standalone data catalogs, at the low end. Snowflake Horizon, Databricks Unity Catalog, and Microsoft Purview cover the basics for teams already on those platforms.

The 4 warehouse names, ranked by fit

Snowflake

The default for teams that want zero infrastructure work. Per-second compute credits, storage billed separately, marketplace of tools. New in 2025: Gen2 warehouses ship with faster compute at their own credit rates and a lower max size (4XL vs Gen1’s 6XL). Native app framework, Snowpark for Python, Cortex for LLM calls on data. If you want the “just works” default, this is it. See the full review.

BigQuery

Google’s serverless warehouse. Priced per TB scanned (on-demand) or per slot-hour (reservations). Best when the team is already in Google Cloud, GA4 is a source of truth, or the workload skews event-heavy. Physical vs logical storage billing at the dataset level is worth understanding (a 14-day lock applies once you switch). See the full review.

Databricks

Grew out of Spark and lakes, now sells warehouse-style SQL on top of Delta and Iceberg. Best when machine learning is the point, when data is genuinely petabyte-scale, or when you want the storage format to stay open. The Free Edition replaced Community Edition in 2025; Mosaic AI covers the LLM stack. Learning curve is steeper than Snowflake, and cluster right-sizing is a discipline. See the full review.

Amazon Redshift

Still fine, still the natural pick for AWS-native shops. RA3 nodes separate storage and compute (finally), and Serverless removes the cluster management overhead. Underrated for workloads with predictable throughput; can be over-configured for spiky analytics. Ranks 4th here mostly because Snowflake and BigQuery caught up on the AWS side.

The supporting cast

Beyond the big 4, these are the warehouse-adjacent tools worth knowing in 2026:

• ClickHouse and Apache Druid: real-time analytics on event streams, sub-second dashboards on billions of rows.
• Apache Pinot and StarRocks: user-facing analytics at high concurrency.
• MotherDuck and DuckDB: the embedded revolution. A shocking amount of “warehouse” work now fits in a laptop process.
• Firebolt: tuned for low-latency workloads with predictable pricing.
• Microsoft Fabric: OneLake-based bundle that consolidates Azure Synapse, Power BI, and Data Factory into one SKU.
• Iceberg and Delta Lake: not warehouses, but the storage formats every warehouse now reads.

Full comparison in the directory and in the tools chapter.

The AI stack next to the warehouse

The most interesting change in 2026 is not inside the warehouse. It’s next to it.

Every major warehouse vendor added vector search in the last 24 months. Snowflake Cortex, BigQuery vector search, Databricks Mosaic AI, Redshift with Bedrock integration. That means RAG against your own data can happen in-warehouse without moving anything.

Meanwhile a real “AI data stack” category emerged: LangSmith, Langfuse, Arize, Braintrust for LLM observability and evals; Pinecone, Weaviate, Qdrant, Milvus for dedicated vector databases; LlamaIndex, Haystack, Unstructured for RAG pipelines and document parsing. Feature stores (Feast, Tecton, Hopsworks) sit on the warehouse and feed ML pipelines.

The strategic implication is that the warehouse is now the ground truth for both analytics and AI. Which makes owning the warehouse decision more important, not less.

The semantic layer comeback

A quiet story worth flagging. Semantic layers spent 2020-2024 as a marketing category nobody actually deployed. In 2025 that changed, driven by AI.

When you point an LLM at “revenue last quarter” against raw tables, it invents definitions. Point it at a governed semantic layer and it picks from real, versioned metrics. Snowflake Semantic Views, Databricks Metric Views, dbt’s Semantic Layer via MetricFlow, and Cube all shipped serious upgrades in the last 18 months. Expect this layer to become a required component of any AI-first analytics stack.

What to watch through 2027

1. Iceberg as the shared substrate. If Snowflake, Databricks, and BigQuery all read the same Iceberg tables, the fight moves entirely to compute, tools and governance. Buyers win.
2. The Fivetran + dbt Labs integration. Ingestion and transformation as one product changes the buying pattern for the modern data stack.
3. AI-native BI. Databricks AI/BI (Genie), ThoughtSpot, and Snowflake’s Cortex Analyst are all betting that dashboards get replaced by chat interfaces on top of a semantic layer.
4. Real-time as the default. ClickHouse Cloud and streaming SQL databases like Materialize and RisingWave are pulling the “batch vs streaming” question toward the streaming side for anything that touches product.
5. Regulatory push. EU AI Act enforcement and sector-specific rules (finance, health) are turning catalogs and lineage from nice-to-haves into audit requirements.

The honest picture

The data warehouse is not a solved problem, but it’s a mature one. The interesting decisions today are less about which warehouse and more about how it connects to the rest of the stack (ingestion, transformation, semantic layer, AI). The vendor that wins the next 5 years won’t be the one with the fastest engine. It’ll be the one whose ecosystem is easiest to compose without lock-in.

Next chapter: The 2026 warehouse pricing playbook, where we get into the actual cost mechanics for each of the 4 warehouses.