The warehouse ate the modern data stack

If you drew a data architecture diagram in 2016, it had ten logos on it. Draw the same diagram today and most of those logos are tabs inside one product. That consolidation is the biggest structural change in the industry, and it explains a lot of what feels confusing right now.

Every diagram starts the same way

Where does the answer come from?

Every architecture diagram is a route from sources to a person. What changed over ten years is how many boxes sit in between.

The stack we all drew on whiteboards

Around 2016 every verb got its own vendor. Ingest, store, transform, orchestrate, model, visualise, reverse-ETL, catalog. Ten logos, ten contracts, ten places to break.

Then the warehouse started eating

Native tasks replaced orchestrators. Transformation moved inside the engine. Catalogs became a tab. The boxes did not disappear so much as get pulled into the big one.

Now the consumption layer is collapsing too

Semantic layers, embeddings, and agents run where the data already lives. The interesting question is what sits on the right edge: who, or what, actually reads the output.

The “modern data stack” was never really a stack. It was a procurement pattern: buy a best-of-breed tool for every verb, wire them together with a warehouse in the middle, and call the wiring a platform team. It worked because warehouses were cheap and elastic, and because nobody had yet built the verbs into the warehouse itself. That stopped being true around 2022.

Why the collapse happened

Three forces, in order of how much they mattered:

Compute got cheap enough to be a feature. Once transformation could run inside the engine for pennies, “we run your dbt for you” became a checkbox. Orchestrators lost their reason to exist for the median team.
Data gravity beat integration. Every extra hop is a place where freshness, permissions, and lineage break. The tool closest to the data wins by default.
AI needs one place to stand. An agent that answers questions about revenue needs the metric definitions, the tables, and the permissions in one addressable place. Scattered across five vendors, it cannot be trusted. Inside one, it can.

logos on a 2016 reference architecture

illustrative count from the diagram above

logos on the 2026 version

sources, engine, and two consumption surfaces

place metrics get defined

the semantic layer moved inside

What did not collapse

The left edge and the right edge.

Sources are as messy as ever. SaaS APIs still change without notice. Nothing about a bigger warehouse fixes that, which is why ingest is the stage where AI help is still thin.

Consumption is where the fight has moved. If compute, semantics, and models all live in one engine, the differentiation is in how the answer reaches a person. A dashboard is one option. A daily summary in Slack is another. An agent you reply to is a third. The engine vendors know this, which is why every one of them shipped a chat interface in the last eighteen months.

What to do about it

Stop buying middle-of-the-stack tools unless you have a specific reason your engine cannot do the job. The default answer is now “the warehouse does it.”
Invest in the semantic layer early. It is the contract everything downstream, including AI, will depend on.
Treat consumption as a product decision, not a BI license. Ask where your stakeholders already spend their attention, and deliver there.

The stack is not modern anymore. It is just the warehouse, plus whatever sits at the edges. Build accordingly.