NVIDIA’s Auto Ontology got its first write-ups at the end of October. Free, open source. It sits between an AI agent and your data warehouse and tells the agent what your data means: which column is revenue, which “active customer” is the approved one, where a number comes from. So the agent writes its SQL against those definitions. No more guessing from column names.

Decent news on its own. What made me sit up is the timing. NVIDIA was the sixth vendor in one month to move on the same idea.

What is NVIDIA Auto Ontology?

In plain words: a business glossary that an AI agent has to use. Auto Ontology reads your database metadata and documentation and proposes a starting ontology, the business terms and how they map to tables and columns. NVIDIA describes it as a starting point “that business experts can test, refine, and certify” (NVIDIA). AI agents reach it through MCP, a command line or an API.

It reads from Databricks, Snowflake, Postgres and DuckDB, keeps the ontology in Postgres next to the catalog it describes, and defaults to NVIDIA’s own Nemotron models, though any OpenAI-compatible model works. The repository has a line I wish more vendors would print: “a revenue metric answered faster is worth nothing if the number is wrong” (GitHub).

Two things to keep next to the launch. The 90% figure, less time to create an ontology, is “in NVIDIA testing”. Their own numbers. And accuracy? No results yet. The repository points to a separate evaluation repository for benchmarks such as BIRD. Harness, no scores.

Who else announced this, and when?

I asked ins7ghts, which reads what vendors and practitioners publish about data and AI every day, for the articles on ontologies and semantic layers since late September. Between the explainers and the job posts, six announcements stood out. I opened each one at the source. Microsoft’s is an addition: it has had Fabric IQ, ontology included, since late 2025, and in Barcelona it added a new way to build those ontologies.

One detail I only caught because the articles sat side by side. Dell’s launch on 6 October already listed NVIDIA Auto-Ontology as part of its platform. The first write-up of the open-source release came on 28 October, three weeks later.

Why are they all building the same thing now?

Because the AI agents arrived before the meaning did. An AI agent that writes SQL will happily join the wrong tables and hand back a number with total confidence. Every vendor in that table sells the same fix, more or less: approved definitions the AI agent must use, and an answer that shows its work.

Now read the last column again. Only NVIDIA says who approves: business experts certify. Oracle and SAP say your own teams define or build the definitions. Microsoft, Dell and Zoho do not say. Where anyone says, it is a person. It reminds me of my DJ years. Anyone can cut a rough mix in an afternoon now, the software does half of it. Somebody still has to stand in front of the speakers and say: this one goes on the record. Once in a while the rough cut was the hit. Most of the time it was not, and the sound engineering and production made it. People underestimate that part. The rough mix got cheap. The sign-off did not, so that is where the cost sits now.

Why I am researching this

Because I ran into the same problem in my own work, on a much smaller scale, and it took me a year to see it.

Ins7ghts turns what the field publishes into answers. Early on we said things about what our data could show, and measured afterwards. In one year we revised those claims about ten times. So now we keep a claims register: every sentence we are allowed to say about our data, the test it passed, and the limits that travel with it. Claims that failed stay on the list, so nobody reuses them by accident.

Reading the NVIDIA repository, I realised that register is the same thing these vendors are building: certified definitions, with their source, that a writer (human or AI) has to use instead of guessing. Ours is for claims, theirs is for revenue and customers. Writing a claim down is quick. Getting it through a test before we may say it, that is the slow part. (I wrote more on that side in data governance for AI and governing agentic AI.)

What I am watching next

Three questions. I will write up what I find.

First, accuracy. NVIDIA ships the benchmark harness, not the scores. The first published result from an independent team, on a real company’s data rather than a benchmark, will tell more than any launch.

Second, the sign-off itself. If every vendor needs a person to certify, the real product is the approval workflow: who approves a definition, how fast, and what happens when finance and sales disagree on what a customer is. That is lineage and provenance applied to meaning.

Third, whether definitions travel. Microsoft used the same Barcelona announcement to commit to Apache Ossie (incubating), a vendor-neutral format for exchanging semantic definitions, formerly Open Semantic Interchange, and Microsoft engineers contributed a Power BI converter (Snowflake). Converters from Ossie to dbt’s semantic layer are already merged (Apache Ossie). If your definitions stay locked inside one vendor’s layer, you will sign them off six times.

I follow this every morning in Ins7ghts Daily. The research shows up there first, before it becomes an article like this one.

How this was read

Six announcements between 28 September and 28 October 2026, found by asking ins7ghts for articles on ontologies and semantic layers, each opened at its source on 3 November 2026. The last column quotes or summarises what each source says about who defines or approves the definitions, and “not described” means the announcement does not say. Oracle’s line on who maintains the rules is an analyst’s reading in InfoWorld. Dell’s is NAND Research’s report of the launch. NVIDIA’s 90% figure is NVIDIA’s own. No ins7ghts count or claim is used in this piece.