Glean's founder says the AI model layer is a commodity. The app layer is where the value is moving.

Arvind Jain runs an enterprise search company that just cleared $200 million in annual revenue. On a recent 20VC interview, he made a case that would have sounded fringe a year ago: the foundation-model layer of the AI stack, the small set of general-purpose models at the top of the pyramid, is no longer where the money is.

The 20VC newsletter for the week of 12 July 2026 and Biggo Finance's recap both carry the same argument. OpenAI and Anthropic built the most capable models, but the durable value in enterprise AI will accrue to the companies that own the data, the workflow context, and the distribution. Jain's framing, "the model layer commoditizes, the app layer wins," is a founder's bet, not a market verdict. But it lines up with where Glean itself has been investing.

Glean is an enterprise AI search and "AI for work" vendor. Sacra pegs the company at roughly $200M ARR, and Perspective AI's research note counts 700 or more enterprise customers. The company's Wikipedia entry places it in the enterprise search category. The financial profile matters because the thesis Jain is selling is, in part, the thesis Glean is built on. If he is right that the durable moat is data and workflow rather than the underlying model, Glean is one of the cleanest public examples of a company trying to live in that moat.

The specific claim doing the work is that "roughly 90%" of AI use cases can now be handled by a wide range of open-weight models (whose weights are publicly downloadable, as opposed to closed API-only systems) and a handful of closed frontier providers. That figure is Jain's framing, not an audited benchmark. He is saying that the marginal enterprise task, summarizing a meeting, drafting an email, searching an internal wiki, no longer requires a frontier closed model to work well. Open-weight releases and a half-dozen API providers can collectively cover the bulk of it.

If the floor is free, the ceiling is the product. That is the mechanism Jain is pointing at. When the foundation layer is commoditized, differentiation, margins, and defensibility migrate up the stack: to enterprise data the model cannot see, to workflow context the model cannot infer from a public corpus, to distribution inside the company's existing tools, and to agentic execution that closes the loop on a task rather than returning a paragraph. The model is the substrate. The product is the connective tissue.

The counterweight Jain himself flags, and which the wire-style "open source is winning" version of this story tends to drop, is selectivity. He notes that AI-driven productivity gains are uneven: customer support agents scale, but the broader claim that "AI makes companies ship faster" does not automatically follow. Support is a high-leverage category because the inputs are bounded, the patterns repeat, and the cost of an imperfect answer is low. Software engineering, sales, and operations look different. A 40% productivity gain in support does not translate into 40% more features shipped.

The other counterweight is the last 10%. Open-weight models vary sharply by language coverage, latency, and regulated-industry compliance. A model that handles 90% of English enterprise Q&A may still be the wrong tool for a HIPAA-bound clinical workflow, a Japanese manufacturing line, or a sub-100-millisecond trading decision. Frontier closed labs continue to push on exactly those edges. The case that the foundation layer is commoditized is not the case that the frontier is.

What to watch in 2026: pricing pressure on the model APIs, especially for the long tail of "good enough" tasks; vertical agents that wrap a workflow rather than a model; workflow depth as a moat, measured in how much of a customer's process the vendor actually owns; and data moats that are not just "we have more documents" but "we have the permissions, the audit trail, and the integration surface to use them inside a regulated job."

Glean's own $200M ARR and 700-enterprise-customer base is a bet that the moat holds. If the bet pays, the next question is which other category, legal, support, sales ops, clinical, gets the same treatment, and whether the model companies become infrastructure providers rather than product companies. The model layer may not be the product. It may be the electric grid.