The “modern data stack” was useful as marketing in 2019 and useful as architecture in 2022. In 2026 it’s mostly furniture. The interesting question isn’t what’s in the stack — it’s what’s actually settled, what’s still a swamp, and what’s being sold as a category that doesn’t exist. Here’s where we land after twelve months of implementation work.

What has converged

The warehouse-plus-transform layer is done. Snowflake, BigQuery, and Databricks are now interchangeable enough at the workload level that the choice is procurement, not architecture. The interesting variation is governance posture, not engine capability. Storage and compute are separated, costs are billed per second, JSON and semi-structured data are first-class, and Iceberg tables work well enough that picking a “lakehouse vs. warehouse” side has become a vendor-relationship question rather than a technical one.

dbt-style declarative SQL transformation has won the build layer. SQLMesh is the credible competitor and the better tool on several axes (we wrote about that separately), but the category is settled: transformations are version-controlled SQL plus a DAG plus tests plus contracts, executed by an orchestrator that knows enough about your warehouse to materialize incrementally. There is no longer a serious argument for hand-rolled Python ETL frameworks at the analytics layer for the kind of mid-market business we work with.

Reverse ETL is done — Hightouch and Census both work, pricing has stabilized, and “sync this model to that SaaS” is rarely the expensive part of an activation project.

The orchestration layer is roughly done. Airflow is the boring default; Dagster is better for teams that want asset-aware lineage native to the scheduler; Prefect 3 cleaned up Prefect 2’s API churn. Pick one based on which abstraction matches how your team thinks.

What is still a mess

Catalog. Every catalog vendor wants to be the system of record for metadata, lineage, ownership, glossary, access requests, and data product certification. None of them wins on more than two of those simultaneously. Atlan, Collibra, Alation, Secoda, and the Databricks/Snowflake-native catalogs all ship with a default workflow that’s wrong for at least half their customers, and the seat-based pricing models punish exactly the broad, casual usage that would make a catalog useful in the first place. The honest 2026 take is that most catalog deployments are write-only — metadata gets pushed in, almost nobody reads it, and the lineage graph is decorative. The teams that get value out of a catalog are the ones who treat it as a contract surface for a specific stewardship workflow, not as a discovery tool.

Semantic layer. dbt Semantic Layer (née MetricFlow), Cube, Lightdash’s metrics, AtScale, and the BI-tool-native semantic layers (Looker’s LookML, Power BI’s semantic models, Tableau’s data model) are all pursuing the same prize and all making different bets about where the layer should live. The result is that picking a semantic layer in 2026 is picking a political theory of how analytics should be governed: warehouse-resident vs. BI-resident, code-defined vs. UI-defined, single-tool vs. headless. None of the bets is wrong; all of them are exclusionary, because the moment you commit to one you’ve shaped how every downstream consumer asks questions for the next three years. We have strong client-specific recommendations here but no universal answer, and anyone telling you they have one is selling something.

MLOps glue. Feature stores, model registries, experiment tracking, online inference, batch scoring, monitoring — every component has at least three credible vendors, and the stitching between them is bespoke at every client we’ve touched. The “MLOps platform” pitch (a single vendor for all of this) keeps failing because the workloads are too heterogeneous: a retention model and a real-time fraud model and a forecasting batch job have almost nothing in common operationally. The teams that ship reliably treat MLOps as a collection of small contracts between tools rather than a unified platform, and they pay the integration tax knowingly. Anyone proposing to replace that integration tax with a single SKU is proposing a lock-in trade you should evaluate carefully.

What is hype

“AI for analytics” is, in 2026, almost entirely RAG over schemas. Strip the marketing and what most of these products actually do is: index your warehouse’s INFORMATION_SCHEMA, index your dbt manifest, optionally index a sample of query history, and let an LLM write SQL against that context. That is a real and useful product category. It is not the product category being marketed. The marketing claims a system that “understands your business”; the implementation reads column names and guesses. When the column names are good, the SQL is reasonable. When the column names are bad — which is most of the time — the SQL is confidently wrong.

A second flavor of hype is the “AI-generated dashboard” pitch. We have yet to see a deployment where the AI-generated dashboards survived a quarter of real use. They get retired in favor of the dashboards a human analyst made, because the analyst encoded judgment about which numbers matter and the model encoded judgment about which numbers were available. Those are different problems.

The honest version of “AI for analytics” — the version we deploy — is narrow, scoped, and mostly invisible: a query assistant that knows your semantic layer, a data-doc generator that keeps documentation fresh, and a triage agent that routes data-quality alerts. None of it requires a new product category. All of it requires good metadata, which is the same prerequisite the last decade of tooling has had.

The pattern

What’s converged: layers where the workload is well-defined and the contract surface is narrow. What’s a mess: layers where the contract is a political question disguised as a technical one. What’s hype: pitches that claim to solve the political question with a model. The stack will keep evolving, but the shape of the disagreements has been stable for three years and is likely to stay stable for three more.

If you’re picking tools in 2026: be aggressive about converged layers, opinionated about messy ones, skeptical of anything whose demo is a chatbot.