Dremio Made the Data Lake Usable. Then SAP Came Shopping

The least glamorous problem in artificial intelligence is the one that keeps spoiling the demo. A model can write a limerick, summarize a contract and suggest a vacation before lunch. Ask it whether the company missed its margin target in Germany, and suddenly the magic trick needs a catalog, permissions, definitions, reliable tables and a patient data engineer. Dremio lives in that unphotogenic gap between the question and the trustworthy answer.

The Austin-based company, founded in 2015 by Tomer Shiran and Jacques Nadeau, sells an enterprise data lakehouse: a query engine, semantic layer, catalog and governance system that sits over data in cloud object stores, databases and warehouses. Analysts can use SQL and familiar tools such as Tableau or Power BI. Engineers can connect applications. Newer AI interfaces can ask questions in natural language. The pitch is that the underlying data can stay in open formats, especially Apache Iceberg, instead of being copied into one more proprietary box before anybody is allowed to touch it.

That sounds like plumbing because it is plumbing. Dremio's bet was that plumbing could become a strategic product once every large company accumulated several clouds, warehouses, catalogs and decades of awkward history. SAP completed its acquisition of Dremio in July 2026. The price was not disclosed. The reason was plain enough: SAP wanted an open, high-performance route between its own application data, everybody else's data and the AI agents expected to use both.

A lake full of data, and no pleasant way to drink it

Shiran saw the original problem while working at MapR. Enterprises were collecting huge quantities of information, but analysts still waited for specialist teams to extract, transform and load it into warehouses or cubes. The data lake was inexpensive and flexible; the user experience was not. Queries were slow. Copies multiplied. Every new question became a small construction project.

Dremio launched publicly in 2017 as a self-service data platform. Its language evolved with the market - data lake engine, SQL lakehouse, agentic lakehouse - but the job stayed remarkably consistent: shorten the trip from raw enterprise data to a governed answer. Underneath that job were technical assets with unusually durable names. Apache Arrow sped up in-memory data interchange. Apache Iceberg gave giant analytical tables open, dependable structure. Apache Polaris, co-created by Dremio and later accepted as a top-level Apache project, supplied an open catalog.

The distinction from Snowflake or Databricks is not that those platforms cannot work with Iceberg - they can. Dremio's sharper position is to be Iceberg-native and engine-neutral, sitting across an existing estate rather than insisting that one vendor become the estate. Starburst and Trino occupy nearby territory. Cloud-native services such as Athena, BigQuery, Redshift, Fabric and Synapse compete from their own ecosystems. Dremio wins when openness, federation and interactive SQL performance matter at the same time.

The product is a bet that the cheapest data copy is the one you never make.

Reliability broke before the fancy architecture mattered

Maersk offers the useful, slightly embarrassing case study. The logistics company had built what looked like a sensible analytics system. Yet some days it recorded literally zero queries. The first failure was not storage capacity or an exotic algorithm. It was trust. Hours of downtime each week taught users that the platform might not be there when they needed it, so they stopped showing up.

That changed the team's mind about what to optimize. Maersk rebuilt its data platform around Dremio as the query layer, standardized on Iceberg, ran the system on Kubernetes and used Apache Superset alongside Power BI. In six months, the platform became its main destination for end-user SQL. Maersk reported 99.97 percent uptime, only three hours of downtime across ten months, 3,000 direct users, another 5,000 downstream users and 1.6 million daily queries. Two semi-dedicated people maintained it.

The figures belong to one customer and should not be mistaken for a universal benchmark. The more portable lesson is behavioral. Infrastructure adoption is earned by boring reliability before it is accelerated by clever features. An unused platform with elegant diagrams is still unused.

What it costs - and what the sticker leaves out

Dremio Cloud publishes a pay-as-you-go list price of $0.20 per Dremio Compute Unit. A DCU measures engine runtime, including queries, Reflections, background processing and platform overhead. Managed storage is listed separately by region - $23 per terabyte per month in two U.S. AWS regions at the time of publication - with standard network and model-provider charges on top. The 30-day trial includes $400 in credits. Developer support is free on pay-as-you-go; paid tiers range from a percentage of the commitment to a Gold minimum of $80,000 or 25 percent of commitment, whichever is higher. Enterprise contracts and volume discounts require a conversation with sales.

The sticker is only half the model. Dremio argues that direct lake queries remove proprietary storage markups, duplicated datasets and permanent ingest pipelines. Its Reflections feature creates and selects optimized representations behind the scenes, similar in spirit to materialized views, so repeated BI queries do not scan raw files every time. That can save compute, but Reflections also consume compute when refreshed. Large raw scans, heavy ad hoc use and machine-learning workloads still raise the bill. Openness does not repeal physics.

There are more tangible outcomes. Amazon's Supply Chain Finance Analytics team reported moving common query times from about 60 seconds to four-to-six seconds, cutting setup time by 90 percent and eliminating 60 hours of work per project. KION said data-development time fell from hours to under three seconds in a highlighted workflow. NetApp reported queries running 20 times faster. These are vendor-published case studies, useful as evidence of possibility and terrible as a substitute for a proof of concept on your own data.

AI arrived, and the semantic layer got promoted

By late 2025, Dremio had recast the lakehouse for agents. The product added a conversational AI Agent, SQL functions that classify or summarize unstructured material, a hosted Model Context Protocol server, service identities for machine-to-machine work and a VS Code extension. The important part is not the chat window. It is that an agent can inherit the same catalog, business definitions, row filters and column masking that govern a human analyst.

This is where SAP's interest becomes legible. Enterprise models are not starved for fluent prose; they are starved for permissioned context. SAP controls richly defined operational data in finance, supply chain, human resources and procurement. Dremio offers a way to combine that context with non-SAP data without first flattening the world into another closed system. The acquisition moved Dremio from venture-backed challenger to a component in a much larger enterprise distribution machine.

Copy the sequence, not the vocabulary

Founders can steal Dremio's sequencing. Start with a recurring, expensive indignity - here, moving and reshaping data before every useful question. Pair a commercial product with standards customers already want to trust. Make the migration incremental: connect a source, serve one demanding BI workload, measure reliability and cost, then expand. Finally, keep the wedge steady even while the category language changes around it.

The culture offers a smaller steal. Dremio calls its employees “Gnarlies,” after Gnarly the narwhal, which risks sounding like forced startup whimsy. The published values underneath are more adult: clear communication, respect, accountability, curiosity, hard conversations and systems that serve hundreds or thousands rather than one favorite customer. A mascot is optional. Naming the behavior expected when a project gets uncomfortable is not.

When this does not work

Dremio is not a universal escape hatch. If a company's data already fits neatly in one warehouse and the bill is tolerable, federation may add a layer without removing one. Open formats reduce lock-in only when teams preserve interoperability in practice. Poorly partitioned Iceberg tables remain poor tables. Semantic definitions still require owners. Governance cannot be generated by a cheerful agent after the access model has been neglected.

The company also faces a delicate post-acquisition test. Dremio's appeal rests partly on neutrality: one open layer across competing stores and engines. SAP gains the most if that neutrality brings more non-SAP data into its orbit, but customers will watch whether ecosystem promises survive the priorities of a giant platform owner. SAP has publicly committed to continued investment in Arrow, Iceberg and Polaris. The credibility of that commitment will be measured in contributions, compatibility and customer choice.

Still, the arc is instructive. Dremio did not invent the data lake, the warehouse or the AI agent. It found the expensive hesitation between them and kept sanding it down. The names on the slide changed. The user's request did not: let me ask a serious question of the data I already own, and do not make me wait until next month.