Same keyway, different keys

Comparisons of Databricks, AWS and Google Cloud usually get framed as a bake-off: which platform is fastest, which is cheapest, which has the better AI story. That framing survives right up until you try to source any of the numbers.

There is a better question, and the last two years have made it answerable. All three now speak the same open table format, so your data is genuinely portable in a way it was not in 2023. What is not portable is the governance model wrapped around it. That is where the switching cost actually lives, and almost nobody markets it, because it is the part you cannot take with you.

What each one actually is now

Worth stating plainly, because all three have renamed things recently and most comparison articles are running on stale vocabulary.

Databricks is a multicloud platform running on all three hyperscalers. The current surface is broad: Delta Lake and Photon underneath, Unity Catalog for governance, Lakeflow for ingestion and orchestration (which absorbed Delta Live Tables as Lakeflow Declarative Pipelines and Workflows as Lakeflow Jobs, both GA on 12 June 2025), Databricks SQL, Mosaic AI, AI/BI Genie, and Lakebase, a serverless Postgres built from the Neon acquisition.

AWS repositioned around SageMaker Unified Studio, GA on 13 March 2025. Read AWS’s own wording carefully: it “brings together functionality and tools from” EMR, Glue, Athena, Redshift, Bedrock and SageMaker AI. It does not replace them. They still exist and are still billed separately. DataZone is not dead either, despite frequent reporting to the contrary; it is the substrate under SageMaker’s domain and governance model, with an explicit upgrade path.

Google Cloud centres on BigQuery, with BigLake for open-format tables and a catalog layer that has been renamed twice: Data Catalog became Dataplex Universal Catalog, and as of 10 April 2026 that became Knowledge Catalog. If you are reading a comparison that says “Dataplex,” it predates April.

The consolidation is real, and it is recent

The last fourteen months of platform strategy

1. 13 Mar 2025

   AWS ships SageMaker Unified Studio

   One workspace over EMR, Glue, Athena, Redshift and Bedrock.

2. 14 May 2025

   Databricks acquires Neon

   Becomes Lakebase. Ghodsi cites agents provisioning most Neon databases.

3. 12 Jun 2025

   Lakeflow reaches GA

   Delta Live Tables and Workflows fold into one brand.

4. 10 Apr 2026

   Google renames Dataplex Universal Catalog to Knowledge Catalog

5. 16 Jun 2026

   Databricks moves on security, agreeing to acquire Panther

6. 17 Jul 2026

   Databricks raises at a $188B valuation

   Coatue-led, per WSJ and Reuters. Up from $134B in December.

Where the lock-in genuinely is

By mid-2026 all three serve the Apache Iceberg REST Catalog protocol. AWS shipped Glue catalog federation for remote Iceberg catalogs in November 2025, explicitly able to query external catalogs including Unity Catalog without copying data. Google’s BigLake metastore speaks the same protocol. Databricks open-sourced Unity Catalog in June 2024 with Iceberg REST compatibility.

So cross-engine reads work. The format war is over and open formats won.

What does not travel across that protocol: row and column level permissions, masking policies, lineage, credential vending, and audit. Each vendor’s catalog owns those on top of the shared metadata layer, and they do not map one to one. Unity Catalog ACLs, Lake Formation tag-based access control, and Knowledge Catalog policies express overlapping but non-identical models.

The practical consequence is that migrating between these platforms is no longer a data migration. It is a governance reimplementation, and that is much harder to estimate because it is not a volume problem. You cannot count terabytes and derive a timeline. You have to re-derive who is allowed to see what, and prove it.

It is also worth noting an asymmetry. Databricks-managed Iceberg tables require Unity Catalog, and write support for external engines into those tables is still labelled public preview rather than GA in Databricks’ own material. Read-out is better supported than write-in, in all directions.

On benchmarks, which you should not trust

The canonical episode here is worth knowing because it set the pattern for everything since.

In November 2021 Databricks announced an audited TPC-DS record, which was real. It separately published an unaudited, TPC-DS-derived comparison claiming large advantages over Snowflake. Snowflake’s founders responded publicly, calling it “a marketing stunt lacking integrity,” re-ran the test on a warehouse half the size Databricks had claimed to use, and reported a cost of $267 against the $1,791 Databricks had attributed to them.

Two parties, one workload, a sevenfold difference

What Snowflake's cost was said to be, by each interested party.

Item	Reported cost for the same workload
As reported by Databricks the competitor's figure, unaudited	$1,791
As reported by Snowflake re-run on a smaller warehouse	$267

The TPC’s own fair-use policy says derived benchmarks are not comparable to official TPC results, and it has since publicly cited other vendors for exactly this violation. The “Databricks is many times cheaper” line still circulating traces to the disowned comparison, not the audited record.

More usefully: we could not find a single independent, methodologically rigorous benchmark of Databricks against BigQuery and Redshift at realistic scale. The closest is Fivetran’s, which is admirably honest about its own limits (1TB scale, sequential single-stream execution, no tuning) and is still produced by a vendor with a stake in the ecosystem.

Pricing does not compare, and that is structural

Each platform meters something different. Databricks bills DBUs, published from $0.15 per DBU for data engineering, $0.22 for data warehousing and $0.40 for interactive workloads, plus a separate CU rate for Lakebase, all on top of cloud infrastructure you pay for separately in classic deployments. Redshift Serverless bills RPU-hours at $0.375. Athena bills $5 per TB scanned. BigQuery offers per-TB-scanned or slot-based capacity, plus a genuinely unusual choice between logical and physical storage billing.

There is no exchange rate between a DBU, an RPU and a slot. A DBU is a normalised proxy for compute on an instance type you chose; a slot is a unit of parallel query capacity decoupled from storage entirely. Any comparison that produces a single number has made assumptions about workload shape, concurrency and tuning that do the actual work of the conclusion.

We looked for a credible, independently funded like-for-like TCO study. There isn’t one. Every pricing comparison site we found is marketing-adjacent with no disclosed methodology.

The one structural thing worth internalising is that a Databricks bill is effectively two invoices in classic deployments: the DBU charge, and the underlying cloud spend. That surprises people, and it is the most common cost complaint we see from teams new to the platform.

The honest decision framework

Given all that, the useful questions are not about which is best.

Where does your data already live, and what does egress cost to move it? This usually decides more than any feature comparison.

Which governance model can your team actually operate? Unity Catalog is the most coherent of the three if you are all-in on Databricks, and it is also the one that most strongly assumes you are.

What is the shape of your workload? Heavy Spark-based transformation and ML has a genuine affinity with Databricks. SQL-first analytics on data already in BigQuery has little reason to leave. Neither of those is a benchmark claim; they are statements about where the friction is lowest.

Who is going to run it on Tuesday? A platform your team can operate at 2am beats a platform that scored better in a comparison written by someone who will not be on call.

Aeolus view. For a startup or scale-up, this decision is usually over-thought. Your data is probably already in one cloud, your identity provider is already there, and the marginal capability difference between these platforms will not determine whether your AI initiative works. What will determine it is whether your metrics are defined once and your pipelines are trustworthy, which is platform-independent work. Pick the one your team can run, keep your tables in Iceberg so the format stays portable, and spend the argument budget on the governance model, because that is the part you would actually have to rebuild.

A note on the numbers

Databricks is private, so its figures are self-reported rather than audited. The solid anchor is its December 2025 announcement of a revenue run rate above $4.8 billion growing 55 percent year over year, more than 20,000 organisations, and a $134 billion valuation. Higher figures circulate, including a $6.9 billion run rate reported by CNBC in June 2026 and the $188 billion valuation reported in July 2026, and those come from credible outlets but are still company-disclosed rather than filed.

For a like-for-like comparison, Snowflake’s audited public filings are more useful: $1.33 billion product revenue in the quarter ending 30 April 2026, up 34 percent, with net revenue retention of 126 percent. AWS reported $37.6 billion in Q1 2026, up 28 percent. Alphabet reported Google Cloud at $24.8 billion for Q2 2026, up 82 percent.

Also worth flagging: every “X percent of enterprises have adopted a lakehouse” statistic we traced came from vendor-commissioned surveys, mostly from vendors selling lakehouse products. There is no neutral adoption figure we could find.

Where this leaves you

The comparison you were probably looking for, a table declaring a winner, would have been built on numbers nobody can source. What we can say with confidence is narrower and more useful: the data format question is settled and portable, the governance question is not, and the cost question cannot be answered generically.

If you are choosing between these platforms, or suspect you chose wrong, we are glad to look at it with you. The answer is often that the platform is fine and the problem sits a layer above it.