AI model marketplaces

Core thesis

Finding a model and running it are overlapping services. Hugging Face shows this: it runs the Hub, where models, datasets and demo applications are shared, and it also offers Inference Providers, which sends a customer's requests to outside companies that serve the model. Lasting value in this market depends on who keeps production workloads, who holds the purchasing relationship and whether serving models earns an acceptable margin. Hugging Face has not published revenue, a current valuation or a listing date that we can confirm, so we do not yet know how the money divides between the hub and the companies that run the models.

State of the art (2026)

Model distribution now covers three things: repositories where models are shared, managed services that run them, and tools for tracking models through their life. Several well-known independents now sit inside larger companies. Replicate became part of Cloudflare on 1 December 2025. CoreWeave completed its purchase of Weights & Biases on 5 May 2025. Hugging Face offers routing to outside inference providers alongside its Hub. The size of a catalogue says little about paid use. The fair way to compare services is to take one model and one workload and look at quality, speed, data handling and the total bill.

Open hubs and cloud marketplaces

Open-model discovery and paid inference already overlap, because Hugging Face lets a customer choose a serving provider and routes requests to it. Large buyers often prefer a service that fits their existing identity controls, billing and data-residency rules, and moving a workload later has a cost. Those factors can favour the big cloud providers. We have not seen figures showing what share of spending the cloud providers take, or showing that independent providers cannot compete. Settling that needs financial disclosures and workload data that can be compared across providers.

Agent marketplaces

A model catalogue distributes or serves models. An agent product adds instructions, tools and a workflow on top of a model, and some products do both. Agent marketplaces are at an earlier stage. We do not have reliable figures for their size or for how much money they make, and each product needs to be judged on its own evidence.