In September 2025, Snowflake announced the Open Semantic Interchange (OSI). It's a collaborative initiative to create a universal, vendor-agnostic specification for semantic layer definitions.
The problem they aim to fix is as follows. Organisations use many tools across their data stack, and each tool defines its own metrics and business logic. OSI aims to create a common format that lets organisations define metrics once. Those definitions then work everywhere. Whether you're in DBT, Tableau, Sigma, ThoughtSpot, or an AI agent.
Snowflake is positioning it as solving "a foundational challenge for AI: the lack of a common semantic standard." The founding partners include Alation, dbt Labs, Hex, Mistral AI, Omni, Salesforce, Sigma, and ThoughtSpot, among others. On paper, it sounds like exactly what the industry needs. Yet, I'm sceptical it will work and not for the reasons you might expect.
Four things give me pause.
The partners don't actually agree on what "semantic" means,
Half the industry isn't in the room, and
The structural incentives favour slow progress.
There's also a chance that by the time they figure it out, LLMs will have made the whole effort irrelevant anyway.
Let me explain.
The Semantic Semantics Problem
The obvious criticisms write themselves: Databricks isn't at the table, neither is Microsoft, there's no public roadmap, and vendor-led standards have a patchy track record. I'll come back to those.
But here's what nobody's talking about: the vendors in this initiative don't agree on what "semantic" actually means.
I've spent time looking at how each OSI partner defines and implements their semantic layer. The differences aren't minor variations in syntax. They're fundamentally different mental models of what the concept represents. Here are some observations.
Transformation vs consumption layer. DBT sees semantics as part of the transformation pipeline, metrics defined alongside models, compiled at build time. Cube, ThoughtSpot, and Omni see it as a consumption layer that sits above the warehouse and generates queries at runtime.
Metrics-first vs schema-first. DBT and Honeydew treat metrics as the primary object; you define what "revenue" means, and the layer figures out the joins. Snowflake, Tableau and Sigma start with logical tables and relationships, with metrics layered on top.
AI-native vs BI-native. Snowflake's Cortex Analyst and ThoughtSpot's Spotter were designed specifically for LLM consumption , YAML schemas optimised for text-to-SQL. DBT, Cube, and Sigma were built for BI tools first and are retrofitting AI support through API integrations.
Catalogue vs layer. Alation, Atlan, and Select Star don't have a semantic layer. They're metadata enrichment tools that sit on top of other people's semantic definitions, cataloguing, building lineage, and governing what others define. They're in OSI to consume and interoperate, not to contribute definitions.
Consumers vs contributors. BlackRock, Blue Yonder, RelationalAI, and Mistral AI aren't semantic layer vendors. They're in OSI because they need semantic interoperability, BlackRock for investment management, Blue Yonder for supply chain, Mistral for LLM grounding. Their presence signals enterprise demand, not product contribution.
The vendors are trying to create an interchange format between systems that don't share a common understanding of the concept they're interchanging. It's like standardising "document interchange" between a spreadsheet, a word processor, a presentation tool, and a database. Yes, they all work with "documents," but the underlying mental models are entirely different. You can export a spreadsheet to PDF, but you'll loose the formulas. You can export a Word doc to plain text, but you'll loose the formatting.
The issue then becomes what exactly transfers when you "interchange" between the tools. The answer, I suspect, is not much beyond the basics. Table names, column descriptions, simple aggregations aren't where the value lies. We want to see magic. For example, if you’ve done some modelling in DBT, I’d love to put Sigma or Tableau on top of that and have them both share the same data model and metrics within their respective capabilities. Imagine how easy it would be to compare BI tools side by side. More on this later.
What do I think “Semantic“ layer means?
I’m out here taking shots but not offering an opinion. A semantic layer codifies how your organisation talks about itself. It maps business terms to data logic, defining what metrics means, which tables to join, which filters to apply, so that anyone asking a question gets a consistent answer, whether they're querying a dashboard, a spreadsheet, or an AI agent.
But consistent doesn't mean identical. The semantic layer generates the query, not the output. A chart needs aggregated totals; a table needs row-level detail; an AI summary needs something else entirely. The semantic layer decides what the metric means. The consuming tool decides how to display it, which means the generated SQL varies depending on what granularity and shape the surface requires. Some processing always has to happen above the semantic layer, in the tool itself. The OSI need to ensure that every partner in the initiative takes the spec and respects the deterministic outcomes the industry expects. How do all these tools consistently communicate granularity, in a way that respects all their interfaces and design philosophies? I didn't sign up for the challenge so I can sleep soundly knowing this isn't my problem to solve.
Who's Not in the Room
Even if the partners could align on what they're building, there's a simpler practical problem.
Databricks is nowhere to be seen. (Databricks Joined on the 26th Jan 2026, beat me to it by 2 days) You cannot claim to be creating a universal data industry standard when one of the two dominant cloud data platforms isn't participating. Their Unity Catalogue has its own approach to semantic definitions, and without Databricks at the table, OSI risks becoming "the Snowflake-aligned standard" rather than a true industry standard.
Microsoft is absent too. No Power BI, no Fabric, no Azure involvement. Given Microsoft's market share in enterprise BI, this is a significant gap. Any organisation running any of the above which is most enterprises, will find themselves outside this "universal" specification.
I'm sure these aren’t the only notable examples. A Snowflake-led coalition of Snowflake-friendly vendors is exactly that. That's not necessarily a bad thing; it's just not an industry standard.
Structural Headwinds
The initiative also faces execution challenges that shouldn't be ignored.
It's solving the easy part first. According to vendors involved, OSI is initially targeting a common format for basic semantic concepts. It's not yet attempting a synchronisation protocol, a common query API, or a way to guarantee consistent numbers across tools. The basics are achievable. Everything after that is what we’re all waiting for.
Partner incentives don't align. As an example, Salesforce is listed as a partner, but Tableau Next has its own semantic model approach baked in, titled Tableau Semantics. Then each partner has their vice. If there's something I've learned about Salesforce since its acquisition of Tableau, it's that they're exceptionally good at selling a vision with the right talking points but if Tableau Next is any indication, it takes a long long time for that vision to become reality.
No clear timeline. When will the specification be ready? When will tools actually support import/export? Nobody seems to know, and the lack of a public roadmap suggests this is still very early.
Semantic lock-in is a feature, not a bug. I mentioned Sigma and Tableau int he context of of DBT modelling. lets expand on that here. If the value in the modern data stack is shifting toward the semantic layer, why would vendors willingly commoditise it? Snowflake wants semantics defined in Snowflake. Tableau wants them in Tableau Semantics. DBT wants them in DBT. A universal format sounds great for customers, but it erodes the switching costs these vendors rely on. Empowering Snowflake's semantic layer to flow freely into any BI tool reduces the differentiation those BI tools can offer around governance, metrics consistency, and upstream integration. OSI asks vendors to collaborate on the very thing that makes their platforms sticky.
That said i have to be fair, momentum is building. OSI expanded its partner list to include Amazon, Google, and DataHub, and held its first working group meeting. The partner list now spans cloud providers, BI vendors, catalogues, and enterprise consumers. That's real weight, it just needs to translate to shipped integrations to avoid becoming just another logo wall.
The LLM Question
One final thought: does any of this matter if LLMs can interpret and translate between formats anyway? If you’d said this to me just a month ago, I would have said “dream on”.
I've worked with LLMs pointing at basic semantic layer definitions. They largely "get it" regardless of syntax. Their approach is entirely different from the one BI tools have to take. An LLM doesn't care whether your metric is defined in dbt YAML, LookML, or Tableau's data model. It can read the intent and work with it. MCPs seem to be the enabling capability here, but better semantic models can’t hurt.
By the time the vendors agree on a specification, build the tooling, and achieve adoption, AI might have made format standardisation irrelevant through the MCP architecture paired with powerful LLMs. The LLM becomes the universal translator.
That's not an argument against OSI. Standards have value beyond technical interoperability. They force clarity of thinking, create shared vocabulary, and provide governance frameworks.
The Bottom Line
OSI is a good idea. Semantic layer fragmentation is a real problem, and a common standard would genuinely help. The pressure to get it right is also very real. Many of the partners in this initiative have already promised an AI-powered nirvana (more on this next week), and 2026 will get slightly awkward when those visions from 2024 don't materialise, particularly for vendors who've already pushed up prices to compensate for capabilities that haven't shipped yet.
I'll be watching with interest. I'm not holding my breath.