Metadata tools did not fail because the idea was wrong. They failed because description does not pay the bills.
Start with the business fact, because it is the one buyers feel first. The standalone data catalog never became the large, durable software category its early valuations implied. Collibra peaked at a $5.25 billion valuation in November 2021 and has since rebranded away from "metadata" toward "data intelligence." Alation, the other category-definer, never went public, was last valued at $1.7 billion in November 2022, and has cut headcount since. Waterline Data, an early machine-learning catalog, was absorbed by Hitachi Vantara in 2020. The clearest signal of all came from Gartner: in August 2021 it scrapped its Magic Quadrant for Metadata Management Solutions entirely and replaced it with a Market Guide that opened on the line "Traditional metadata practices are insufficient."
That is the headline. The rest of this piece explains why it happened, where metadata tools still earn their keep, and why the operational center of the data stack has shifted to the semantic layer. We will be precise, because a "decline" narrative is easy to overstate. Metadata is not dead. The category that sold metadata as a standalone governance product is the thing under pressure.
What metadata tools were, and why they got funded
For roughly a decade, the enterprise answer to "we have too much data and nobody understands it" was to buy a catalog. Alation, Collibra, Waterline, Informatica, and later Atlan and data.world all sold variations of the same promise: discover what data exists, document where it came from, tag what is sensitive, define what terms mean, and make all of it searchable.
This was a real problem and these were real products. Gartner published a Magic Quadrant for Metadata Management Solutions annually through its final edition on 11 November 2020, and being named a Leader in it helped Collibra, Alation, and Informatica raise large rounds and win large enterprises. Billions of dollars in venture funding and enterprise spend flowed into the category between roughly 2015 and 2021.
The structural flaw: metadata describes, it does not enforce
Here is the technical reason the business model strained. A catalog is descriptive, not operational. It sits beside the data and tells you things about it. It does not participate in what happens to the data.
Consider a policy. In a catalog, a policy is a record: "this column is PII, restrict access." Whether that restriction is actually enforced depends on a completely separate system, the database or the warehouse, and on a human reading the tag and configuring the control correctly. The distance between the description in the catalog and the enforcement in the data platform is exactly where governance fails. Nothing in the catalog compiles into the query. Nothing breaks when reality drifts away from the documentation.
That gap produces a predictable decay pattern. Definitions are written once and go stale. Lineage graphs break as pipelines change. Glossaries diverge from how teams actually calculate revenue. Sensitivity tags lag behind new tables. Because nothing downstream depends on the metadata being correct, keeping it correct is pure cost with deferred benefit. The catalog becomes shelfware.
Why the ROI never closed
For an economic buyer, the failure mode is mundane and brutal: slow time to value plus low adoption equals weak renewal. Catalogs required heavy upfront modeling and ongoing manual curation by data stewards who often lacked the authority to enforce anything. Industry write-ups on governance failure converge on the same point: programs die not from bad technology but because the tooling lived outside the daily workflow, so people ignored it. A catalog that is not used cannot be renewed at the price it was sold at.
The market data reflects the squeeze. The standalone data catalog market sat somewhere between $1 billion and $2 billion in 2024, small for a category that absorbed so much capital, and analyst growth estimates diverge wildly, which itself signals an unsettled category. The vendor outcomes tell the rest of the story: no catalog-first company reached a large public exit. Several were absorbed into larger platforms.
The consolidation wave of 2025
The clearest evidence that standalone metadata struggled to stand alone is that, in 2025, much of the category got acquired into something bigger. ServiceNow bought data.world. Salesforce agreed to buy Informatica for roughly $8 billion. Collibra spent the year acquiring rather than being acquired, bolting on data-access governance (Raito) and unstructured-data capabilities (Deasy Labs) and leaning hard into the "data intelligence" and AI-governance narrative. Even Alation, the catalog archetype, spent 2025 buying an AI-agent startup, Numbers Station, to reposition around agentic workflows.
Detachability is the through-line. A catalog attaches to your metadata, and metadata can be re-pointed at a new tool. That makes catalogs replaceable, and replaceable products have weak pricing power and weak retention. This is the bridge between the technical story and the business story: the same property that makes metadata passive (it sits beside the data) makes the vendor detachable (it sits beside the stack). Enforcement equals stickiness. Description equals churn risk.
The turning point: LLMs made "query the data" everyone's job
The shift accelerated in 2023 and 2024 for a specific reason. Large language models turned "ask the data a question" from a task for trained BI analysts into a mainstream expectation. And LLMs exposed the catalog's core weakness instantly.
Point an LLM at raw warehouse tables and ask it a business question, and accuracy is poor: on the BIRD benchmark, text-to-SQL execution accuracy started around 40 percent in March 2023, and on the harder, enterprise-realistic Spider 2.0 benchmark the best model still reached only 31 percent as of April 2025. Ground that same model in a governed semantic layer, where revenue, churn, and active customer are precisely defined, and the picture changes completely. dbt Labs reran its own benchmark in 2026 and found that for questions within a well-modeled semantic layer's scope, leading models return correct results 100 percent of the time, while raw text-to-SQL on those same models hovered in the mid-60s and produced inconsistent results run to run.
A searchable glossary cannot fix this. An agent does not read your business glossary and absorb wisdom. It needs definitions it can compile through. That is the semantic layer's entire reason to exist.
What a semantic layer actually is, and why it wins
A semantic layer is a typed, versioned, executable definition of business meaning, entities, metrics, relationships, and constraints, that sits between data sources and anything that queries them. The difference from metadata is not cosmetic. It is architectural:
- Operational enforcement. The definition is compiled into every query. You cannot compute revenue a second way, because the only path to "revenue" runs through the definition.
- Continuous validation. A mismatch raises an error instead of silently returning a wrong number. Failures surface at compile time, not in a board deck three weeks later.
- Lower manual burden. Definitions evolve through code review in version control, not through a steward manually re-tagging assets in a separate UI.
That last point is the economic punchline. Metadata is governed by adding human effort over time. Semantics are governed by changing code, which teams already do. One scales with headcount. The other scales with engineering practice.
The vendor evidence: everyone is building semantic-layer-first
Follow where the building is happening. dbt rose from a transformation tool into a semantic layer by acquiring Transform and its MetricFlow engine in 2023, then re-open-sourced MetricFlow under Apache 2.0 in 2025 after concluding a closed semantic layer would not be trusted. In October 2025, Fivetran and dbt Labs agreed to merge into a roughly $600 million ARR entity explicitly positioned as "the data infrastructure for agents you trust." Snowflake shipped Cortex Analyst in August 2024 and native Semantic Views, treating the semantic model as a first-class, governed database object. Databricks open-sourced Unity Catalog in June 2024 and added governed Unity Catalog Metrics. Cube was built semantic-layer-first as an API-first headless layer; AtScale and Timbr launched as semantic platforms, not catalogs.
The tell is unmissable: new entrants launch as semantic layers, while incumbent catalogs scramble to bolt on "active metadata" and AI-readiness messaging. When the challengers and the incumbents are both running toward the same architecture, the architecture has won the argument.
The honest part: what metadata still does well
A fair analysis has to resist the temptation to declare metadata dead, because that claim is false and any buyer who has run a regulated data estate knows it. Several metadata functions remain genuinely valuable and are not replaced by a semantic layer:
- Sensitivity and PII classification. A human-curated, searchable, policy-bound classification of what is sensitive is a regulatory requirement under GDPR and CCPA. A metric object does not do this.
- Discovery and lineage. "What tables exist and where did they come from?" is a legitimate question, especially in data mesh estates with distributed ownership.
- Business glossary. A human-readable, version-controlled definition of what "customer" or "churn" means in plain language is a different artifact from a compiled metric, and useful for humans aligning on meaning.
- Access request workflows. "Who can access what" is a governance function that catalogs often hosted well as a user experience.
The honest frame is complementarity, not combat. Metadata describes. Semantics enforce. A mature estate needs both. What changed is which one sits at the center.
Why this is now structural, not a fad
Three forces lock the shift in place. The LLM inflection made deterministic, policy-aware query interfaces a requirement rather than a nicety. The data mesh inflection pushed ownership outward, which means enforcement has to move from centralized human stewardship to compile-time rules that travel with the definition. And the regulatory inflection raised the stakes: the EU AI Act's Article 10 data-governance obligations for high-risk systems become enforceable on 2 August 2026, and they demand auditable, versioned governance, not tags in a search index.
Then there is the irony worth sitting with. In November 2025, Gartner revived the Magic Quadrant for Metadata Management Solutions after a four-year absence, describing the market as shifting "from augmented data catalogs to metadata 'anywhere' orchestration platforms," and naming Informatica, IBM, Alation, Atlan, and Collibra as Leaders. Metadata is not disappearing. It is being absorbed into operational systems and made active. The standalone catalog as the center of governance is what ended. Metadata as a capability inside a semantic and governance fabric is being reborn.
Where Colrows fits
Colrows is built for this world, and it is worth being precise about the claim. Colrows is a semantic layer, not a replacement for every metadata function. PII classification, discovery, and a human-readable glossary still matter, and often live best in dedicated tools. What Colrows provides is the operational core: entities, metrics, and relationships as first-class objects; definitions that are executable and versioned; governance enforced at compile time; lineage that is inherent in metric dependencies rather than reconstructed after the fact.
The one distinction that matters most for buyers: the warehouse-native semantic layers stop at their own boundary. Snowflake Semantic Views govern Snowflake. Databricks Unity Catalog Metrics govern the lakehouse. That is powerful inside one platform and silent outside it. Colrows is semantic-layer-first across the estate, built for multi-source, multi-warehouse environments where meaning has to stay consistent no matter where the data lives. That is the gap warehouse-native semantics leave open, and it is the gap that matters as estates get more federated, not less.
The bottom line
The center of the data stack moved from storage to compute to meaning. Metadata tools helped us describe data, and description was necessary but never sufficient. The market learned, expensively, that a system nothing depends on will be neglected, and a neglected system will not renew. Semantic layers win because they are depended upon: every query and every agent compiles through them. That technical fact (enforcement beats documentation) is the same as the economic fact (stickiness beats detachability). Buy metadata tools for what they are still good at. Build your governance on the layer that the queries cannot route around.
