Every two years since 2018, someone has launched a model marketplace with the intent to be “the App Store for AI.” None of them have worked, except one. Here’s why, and what it implies for the infrastructure layer of enterprise AI.
The pattern
The pitch is always the same: enterprises want to swap models in and out based on task and cost. A marketplace abstracts the underlying model, provides a single API, and lets customers choose the best model for each use case. Beautiful in theory. Graveyard in practice.
The failed examples follow a consistent path. Launch with a wide catalog. Attract developer attention. Struggle to get enterprise procurement buy-in. Fail to differentiate from the cloud providers (AWS, Azure, GCP) who offer the same models with better SLAs and existing enterprise agreements. Pivot or shut down.
Three examples worth examining briefly: AI21 Labs’ enterprise marketplace pivot, Cohere’s brief experiment with a model discovery layer (now scrapped), and Scale AI’s attempted marketplace before they focused on data. All smart companies. All faced the same structural problem.
The Hugging Face exception
Hugging Face works because it solved a different problem. It’s not a procurement marketplace — it’s a collaboration and distribution platform for the model development community. The stickiness comes from the community: researchers upload models, developers build on them, companies fine-tune them, the whole ecosystem creates value that no single participant could replicate. Network effects are the moat, not the catalog.
The enterprise deployment layer of Hugging Face (the “Inference Endpoints” product) is a separate business with a different value proposition: run this community model in your cloud environment, with security guarantees. That’s infrastructure, not a marketplace.
What actually makes AI infrastructure sticky
The consistent pattern across durable AI infrastructure businesses: stickiness comes from data, not from models. The infrastructure that persists is the infrastructure that sits between your proprietary data and the model layer, not the infrastructure that abstracts one model from another.
Vector databases have won on this logic. Your embeddings are valuable. Migrating them is painful. The database that holds them becomes sticky. Fine-tuning infrastructure wins on the same logic: your labeled training data is valuable, and the tooling that manages, versions, and deploys it gets embedded in your workflow.
Model-agnostic proxies — tools that swap models in and out transparently — have not found durable enterprise revenue, because the model is not the sticky part. The data and the fine-tuned behavior are the sticky part.
What comes next
The next wave of enterprise AI infrastructure will be built around agent memory and context management — specifically, the tooling that maintains state and context across agent sessions, task chains, and multi-agent workflows. That tooling is deeply embedded in enterprise data, it accumulates value over time, and it’s genuinely hard to migrate. That’s a marketplace that won’t die.