Models & Infrastructure Tools

Models and infrastructure is the builder's category: where AI models get hosted, served, connected, and measured. If you're shipping an AI product rather than using one, these are your suppliers.

Hugging Face is the gravity well — the hub where open models, datasets, and demos live, plus hosting on top. Replicate and fal.ai sell serverless inference: call a model via API, pay per second or per generation, never touch a GPU. Modal and Baseten are deployment platforms for when you outgrow serverless and want your own endpoints with real scaling controls. Pinecone is the vector database most retrieval-augmented apps started on. LiteLLM is the open-source gateway that normalizes a hundred model APIs behind one interface. Heurist and x402 round out the infrastructure edge.

Unusually, this category also includes the scorekeepers: LMArena runs the crowdsourced model leaderboard, Artificial Analysis benchmarks price and performance across providers, and Stanford HELM and SWE-bench publish rigorous academic and coding evaluations. That mix is deliberate — choosing infrastructure without benchmarks is guessing.

The differentiators are practical: cold-start latency, GPU pricing, developer experience, and lock-in. Costs range from free (the benchmarks) to enterprise contracts. The sound pattern: prototype on serverless inference, validate with public benchmarks plus your own evals, and only then commit to dedicated deployment.