The benchmark that keeps coming up
Over the past two years, a specific comparison has shown up repeatedly in data engineering benchmarks: a single, well-tuned machine running an embedded analytical engine keeps pace with, or beats, multi-node cloud data warehouses on the query workloads most companies actually run.
DuckDB, the open-source in-process analytical database, reached the top spot among open-source engines on ClickBench with its v1.4 LTS release, built around a vectorized execution engine that processes data in batches rather than row by row. MotherDuck, the commercial cloud service built on top of DuckDB, publishes its own ClickBench-derived comparisons showing its "Mega" tier finishing a standard benchmark run in 5.9 seconds against 18.7 seconds for ClickHouse Cloud and 32.3 seconds for self-hosted ClickHouse.
Those numbers come from the vendor running the benchmark, which is worth flagging up front. MotherDuck has a direct commercial interest in DuckDB looking fast, and vendor-published benchmarks should always be read with that in mind. But the underlying pattern, that single-node vectorized engines are competitive with distributed systems on datasets that fit on one machine, is not a MotherDuck invention. It traces back to a 2023 argument from MotherDuck co-founder Jordan Tigani, "Big Data Is Dead," which pointed out that most organizations query datasets far smaller than the multi-petabyte scale that distributed warehouses were built for.
What "small data" actually costs
The pricing gap, where it holds up, is large enough to matter for budget conversations, not just architecture debates.
Snowflake's on-demand credit pricing, effective August 2026 for AWS US East, runs $2.00 per credit on Standard, $3.00 on Enterprise, and $4.00 on Business Critical. An X-Small warehouse burns one credit per hour, and each size step doubles consumption, so an XL warehouse consumes 16 credits per hour, roughly $32 to $64 depending on edition. MotherDuck's own comparison puts its Standard tier at roughly 1/13th the hourly cost of a Snowflake XL while delivering, by its measurement, about double the throughput on the same query set. At the high end, MotherDuck's Mega tier lists at $12.00 per hour against a Snowflake 3XL configuration it describes as 91% more expensive per hour for comparable ClickBench performance.
Redshift comes out worse in the same comparison: the fastest configuration MotherDuck tested, a 4-node ra3.16xlarge cluster, runs about $52 an hour, nearly 22 times MotherDuck Standard's cost, for performance MotherDuck's own numbers put slightly behind.
There is also a structural difference in how the meter runs. Snowflake bills warehouse compute with a 60-second minimum each time a warehouse spins up, while MotherDuck bills per second with no minimum. BigQuery uses a different model entirely, charging by bytes scanned rather than by compute time, which independent cost writeups note can produce a "petabyte tax": a single query touching a petabyte of data can run in the thousands of dollars regardless of how simple the query logic is. For teams running frequent, small, interactive queries rather than sustained batch jobs, the billing model can matter as much as raw engine speed.
Self-hosted ClickHouse sits in a different part of the map. Industry cost estimates put a single well-tuned node handling around 10 TB of compressed event data at $1,000 to $5,000 a month in infrastructure cost, competitive with managed options but requiring in-house operational expertise that managed services are priced to remove.
Where this argument breaks down
None of this means distributed cloud warehouses are obsolete, and the benchmark data itself points to where the single-node story stops applying.
The first limit is data volume. DuckDB and similar engines operate on data that fits, or mostly fits, in memory or local disk on one machine. Once working sets move into true multi-terabyte or petabyte territory with heavy concurrent access, distributed architecture stops being a cost inefficiency and becomes a requirement.
The second is concurrency. Benchmarks like ClickBench typically measure single-query or low-concurrency performance. Enterprises running dashboards and BI tools against a warehouse with dozens or hundreds of simultaneous users need the multi-cluster scaling that Snowflake's Enterprise tier and equivalent products in BigQuery and Redshift are built around. A single-node engine that wins on a solo query benchmark can fall over under real concurrent load, and none of the public comparisons cited here isolate that variable cleanly.
The third is governance and access control at organizational scale: row-level security, fine-grained audit logging, and multi-team workload isolation are mature in the established warehouses in ways that newer single-node-first products are still building out.
The practical takeaway
The honest reading of the current benchmark landscape is narrower than "big data is dead," and narrower than MotherDuck's own marketing implies. What the numbers actually support is that a meaningful share of workloads currently running on full-scale cloud warehouses are sized for data volumes and concurrency levels the team does not have, and that the cost of that mismatch is now well enough documented to justify checking. For a team running interactive analytics on datasets in the low terabytes with modest concurrent users, testing a single-node engine against the current warehouse bill is now a cheap, low-risk experiment. For teams at genuine multi-petabyte scale with heavy concurrent BI traffic, the distributed warehouses remain the only option that has been proven to hold up, and no benchmark cited here suggests otherwise.
Sources: MotherDuck: The fastest OLAP databases compared · MotherDuck: Big Data is Dead (Jordan Tigani) · DuckDB Benchmark Suite documentation · ClickBench project · Snowflake pricing guide, effective Aug 7 2026 rates via CostBench · Valiotti Data: ClickHouse vs Snowflake in 2026