# Inference Stack

> The reference for how AI infrastructure actually fits together: an interactive navigator for the AI inference stack — application archetype through inference framework, orchestration, hardware abstraction, silicon, and cloud instance — with multi-vendor coverage (NVIDIA, AMD, Google TPU, AWS Trainium/Inferentia, and Intel Gaudi). Also includes a family-level product catalog spanning NVIDIA, AMD, Google, Amazon, and Intel data-center hardware, traced through to where it deploys on AWS, Google Cloud, and Microsoft Azure where that data is available.

This is a structured dataset with a browsable UI on top, built for engineers and agents to query directly rather than scrape the rendered page. Family-level, not exhaustive SKUs — see `/docs/taxonomy-notes.md` for scope notes and open questions. Roadmap-stage products (not yet shipping) are explicitly flagged as such in `availability_status`; spot-check anything you rely on against the linked source before committing spend, since chip/instance availability changes fast.

## Pages

- [/](/): interactive Finder-style column browser — pick an application archetype and trace compatible inference frameworks, orchestrators, hardware abstraction layers, silicon, and cloud instances down to a full deployment path. Multi-vendor: NVIDIA, AMD, Google TPU, AWS Trainium/Inferentia, and Intel Gaudi.
- [/architectures](/architectures): reference architectures — vendor-published blueprints (NVIDIA, AWS, Google Cloud, Azure) and verified production case studies (Character.AI, Fireworks AI, Baseten, Together AI), each citing a real source. Filterable by source type; stack fields link into the stack browser where they match a real node.
- [/catalog](/catalog): every product family — NVIDIA, AMD, Google, Amazon, and Intel data-center hardware — filterable by vendor first, then silicon type, grouped by segment. Per-product deep links: `/catalog?product=<id>` opens that product's detail view directly — treat these as stable, citable URLs.
- [/deploy](/deploy): NVIDIA GPUs plus each cloud's own silicon (AWS Trainium/Inferentia, Google TPU) traced through to the cloud instance families built on them, across AWS, Google Cloud, and Microsoft Azure. Accepts `?provider=aws|gcp|azure` and `?product=<id>` to deep-link a specific provider/product.

## Data (machine-readable)

- [Live query API — products](/api/products): every product family (NVIDIA, AMD, Google, Amazon, Intel), filterable via `?segment=`, `?silicon_type=`, `?status=`, `?vendor=`, `?q=`.
- [Live query API — deployments](/api/deployments): cloud instance families mapped to the silicon inside them (NVIDIA and, where cataloged, each cloud's own), filterable via `?provider=aws|gcp|azure` and `?product_id=`.
- [Live query API — architectures](/api/architectures): reference architectures (vendor blueprints + production case studies), filterable via `?source_type=vendor|case_study`.

## Docs

- [Taxonomy notes](/docs/taxonomy-notes.md): what's deliberately simplified, what's out of scope, open questions for the next pass.

## Notes for agents

- Treat `id` fields (on products, providers, and instances) as stable keys for joining across the API responses; `product_id` on a deployment instance links back to the products catalog.
- `meta.as_of` and per-product `sources` indicate freshness — this dataset tracks a fast-moving market and is refreshed by hand, not continuously.
- This site has no rate limiting in place today; be a reasonable citizen if polling programmatically.