AI model marketplaces

State of the art (2026)

The category split into two businesses in 2026. Hugging Face remains the open-model hub – discovery, weights, datasets, demos – the winner-take-most layer, but is still privately held (roughly $4.5B at its last raise) rather than publicly listed. But commercial inference is consolidating onto the hyperscaler-owned marketplaces: AWS Bedrock (30+ models), Google Vertex AI Model Garden and Microsoft Foundry (formerly Azure AI Studio) all offer governed, pay-per-token access to Claude, Llama, Mistral and Gemini behind enterprise IAM and billing. Independent inference clouds – Replicate, Together AI, Fireworks AI, OpenRouter – compete on price and latency below them. Agent stores (OpenAI GPT Store, Copilot Studio) remain a separate, earlier-stage layer tied to their host foundation model rather than an open marketplace.

Hub-vs-marketplace bifurcation

The market runs along a structural split. The hub layer – model discovery, weights distribution, datasets, demos, community – is winner-take-most, and Hugging Face has it (roughly 2.4M models by mid-2026). The commercial inference layer – actually running models in production at low latency and acceptable cost – is fragmenting across Replicate (pay-per-second, broad catalogue), Together AI (fast open-source LLM inference, OpenAI-compatible API), Modal (Python-first serverless), Fireworks AI (enterprise fast inference) and increasingly the hyperscaler-owned marketplaces (AWS Bedrock, Google Vertex AI Model Garden, Microsoft Foundry). Hugging Face is still privately held – its August 2025 round valued it at roughly $7B and it has not filed to go public – so the equity question is unresolved: revenue near $130M is too small to underwrite a multi-billion valuation on the hub alone. The competitive question for the 2026–2028 window is whether Hugging Face can build a defensible inference position against hyperscalers with cheaper compute, deeper enterprise relationships and built-in IAM, billing and compliance.

Hyperscaler marketplaces eat the workload, open hubs keep the discovery

Enterprise inference is consolidating onto hyperscaler marketplaces faster than the open-hub thesis assumes. AWS Bedrock now offers 100+ proprietary and open-source foundation models with managed inference. Google Vertex AI Model Garden similarly bundles Llama, Mistral, Anthropic and Google's own families. Microsoft Foundry (renamed from Azure AI Foundry, itself formerly Azure AI Studio) is the catalogue over OpenAI plus a large third-party set spanning Anthropic, Meta, Mistral, Cohere, DeepSeek, xAI and more. The buyer-side advantage is IAM, billing, data residency and audit alignment with the rest of the enterprise cloud spend – which model-hub platforms cannot match. What the hubs retain is the discovery and demo layer: where a developer first encounters a new model release, evaluates capabilities and decides whether to deploy. That distinction maps cleanly to Hugging Face (discovery) vs hyperscaler marketplaces (commercial inference) and frames the durable revenue split.

Agent marketplaces are a separate, earlier-stage market layer

Agent marketplaces (OpenAI GPT Store, Microsoft Copilot Studio store, Anthropic agent marketplace, Google Gemini Custom) are distinct from model marketplaces - they distribute task-specific agents built on top of foundation models rather than the models themselves. The consumer and prosumer agent-marketplace category is much earlier than the model-marketplace one and is structurally tied to the proprietary foundation models that host it (GPT Store on OpenAI, Copilot Studio on Microsoft+OpenAI, Anthropic agent marketplace on Claude). Open agent marketplaces independent of a single foundation model have not yet emerged at scale. The next 12-24 months determines whether agent distribution converges on the foundation-model-owner platforms or whether an open agent-marketplace layer is viable at all.