# AI Compute Radar

> Measured signals for the AI-model and compute market: Hugging Face
> trending and downloads, OpenRouter API prices, Vast.ai GPU rental rates —
> collected automatically several times a day — plus deterministic engines
> that compute which open models fit which hardware and what a workload
> costs across cloud API, owned hardware and rented GPUs. Every figure is a
> measured reading with a timestamp, a sourced vendor fact, or a declared
> assumption; the methodology page states the rules.

## Core pages

- [Model board](https://aicomputeradar.dev/): open models ranked by measured Hugging Face trending score, with API prices and download counts
- [Model pages](https://aicomputeradar.dev/models): one page per tracked model at https://aicomputeradar.dev/models/<slug> — Heat Score, measured Hugging Face signals with daily readings and the OpenRouter price where one exists; for models with a measured GGUF file also the memory needed at 8K, 32K, 64K and 128K context (weights + context cache + overhead), fit verdicts on every tracked GPU and Mac, and the cheapest measured rental class that runs it comfortably. Models without a published attention shape carry UNKNOWN verdicts rather than guesses.
- [Departures board](https://aicomputeradar.dev/board): every scored model on a live split-flap board — rank, Heat Score, direction — flipping only when a number moves
- [Pick of the week](https://aicomputeradar.dev/pick): every Friday one tracked model, chosen by a published rule (largest counted Heat Score rise among models that run comfortably on a consumer card of up to 24 GB), with the numbers frozen at selection time, a device fit ladder and a written report with caveats; permanent issues at https://aicomputeradar.dev/pick/<year>-w<week>, RSS at https://aicomputeradar.dev/pick/feed.xml
- [Local AI finder](https://aicomputeradar.dev/local): which models run on which GPU or Mac, from measured GGUF sizes and published attention shapes
- [Cost calculator](https://aicomputeradar.dev/costs): the same model priced three ways — cloud API, owned hardware, rented GPU
- [GPU rental prices](https://aicomputeradar.dev/compute/rent): median verified $/h per GPU class, the Rent Index (this week against last week and against the first week collected — described, never forecast) and robust daily history
- [Buy or rent a GPU?](https://aicomputeradar.dev/compute/buy-or-rent): how GPU rental works in plain language, the break-even between a sourced launch price and the measured hourly rate (RTX 4090, RTX 5090), what a working month on an A100/H100/H200/B200 costs, and when each path wins
- [Stack builder](https://aicomputeradar.dev/stack): a deterministic local/cloud stack recommendation from seven questions
- [Local AI tools](https://aicomputeradar.dev/tools): curated runtimes (Ollama, LM Studio, llama.cpp, Jan, Open WebUI)
- [Hardware hub](https://aicomputeradar.dev/hardware): every tracked GPU and Mac with how many models run comfortably, plus the guides below
- [Local AI on a budget](https://aicomputeradar.dev/hardware/budget): which used and entry-level GPUs genuinely run local models, what to avoid, buying checklist — fit counts computed, no street prices
- [Radar Cards](https://aicomputeradar.dev/cards): twelve machines from the 3dfx Voodoo (1996) to the DGX Spark as collectible spec cards at https://aicomputeradar.dev/cards/<slug> — every fact with its status (vendor / review / reported) and source, a hologram drawn from the machine's dimensions and parts list (not a scan), and which of today's tracked models run on it at 8K context through the same fit engine as the finder. Numbers are issued in claim order and kept in the reader's browser; no blockchain, nothing for sale
- [Glossary](https://aicomputeradar.dev/glossary): local-AI terms in plain language (GGUF, quantization, KV cache, unified memory, MoE, …) — one page per term under https://aicomputeradar.dev/glossary/<slug>; Spanish edition under https://aicomputeradar.dev/es/glossary, Portuguese under https://aicomputeradar.dev/pt/glossary (same slugs, hreflang-linked)

## How the numbers are made

- [Methodology](https://aicomputeradar.dev/methodology): what is measured, what is computed, what is assumed — and what this site refuses to publish

## Machine access

- [JSON API](https://aicomputeradar.dev/api/v1): read-only, no keys, CORS-open — models ranked by Heat Score with measured signals (https://aicomputeradar.dev/api/v1/models), hardware-fit verdicts (https://aicomputeradar.dev/api/v1/fit?hardware=rtx-4090), GPU rental prices (https://aicomputeradar.dev/api/v1/prices), hardware profiles (https://aicomputeradar.dev/api/v1/hardware), the pick of the week with its frozen snapshot and report (https://aicomputeradar.dev/api/v1/pick, ?week=<year>-w<week> for past issues). OpenAPI 3.1: https://aicomputeradar.dev/api/v1/openapi.json
- [MCP server](https://aicomputeradar.dev/api/mcp): Streamable HTTP, stateless — tools trending_models, find_fit, gpu_prices, list_hardware, weekly_pick. Setup and terms: https://aicomputeradar.dev/developers
- Every response carries meta.source (live = measured, fallback = demo fixtures) and meta.collectedAt. Data is CC BY 4.0 — attribute "AI Compute Radar (https://aicomputeradar.dev)".

## Notes for agents

- Per-device pages exist at https://aicomputeradar.dev/local/gpu/<slug> for every tracked GPU and Apple Silicon machine.
- Data freshness is visible on every section; the health endpoint at https://aicomputeradar.dev/api/health reports collection staleness as JSON.
- Outbound model links resolve via https://aicomputeradar.dev/go/huggingface?p=<repo>; GGUF links point at the exact quantized file our memory figures were measured from.
- Corrections, feedback, missing models or GPUs: aicomputeradar@gmail.com (see https://aicomputeradar.dev/legal/imprint#contact). A correction that survives checking updates the data.