PlanetGPU Documentation - Data Sources, API, and Methodology

Pricing & data freshness

Live GPU prices are aggregated from public provider catalogs via the open-source gpuhunt project. PlanetGPU caches price API responses briefly, while gpuhunt reloads its catalog during runtime and online providers can change between requests. For each GPU we show the cheapest on-demand and spot rate per provider, the median, and the 90th-percentile (P90) price across all tracked offers. Prices are per GPU, per hour, in USD, and exclude storage, egress, and taxes.

Specifications (VRAM, memory bandwidth, FP16 Tensor TFLOPS, TDP) come from a curated catalog. FP16 figures are dense (no 2:4 sparsity) so vendors compare apples-to-apples. Some outbound provider links are affiliate links marked rel="sponsored nofollow"; they never affect ranking - results are always ordered by price.

Quick Start

PlanetGPU helps you find the cheapest GPU across 12 cloud providers and deploy AI workloads with dstack.

1. Search GPUs

Visit /search to browse GPU cards. Each card shows the cheapest price, number of providers, and regions.

2. Compare GPU power and model fit

Visit /leaderboard to rank the most powerful GPUs, or /find to match an LLM size and precision to GPUs with enough VRAM.

3. Deploy a Repo

Paste a GitHub, Hugging Face, or Docker target on /deploy. PlanetGPU estimates the workload, VRAM footprint, recipe, and suitable GPU families; validate the result before production deployment.

4. Use the API

All data is available via our public API:

GET https://api.planetgpu.com/v1/search?level=gpu GET https://api.planetgpu.com/v1/index/h100 GET https://api.planetgpu.com/v1/leaderboard/models GET https://api.planetgpu.com/v1/hf/models?search=&pipeline_tag=text-generation&sort=downloads&limit=50 POST https://api.planetgpu.com/v1/graph/refresh-llms?limit=500&min_downloads=100&fetch_cards=true

5. Discover LLMs from Hugging Face

Use the /v1/hf/models endpoint to search the Hugging Face Hub for text-generation models sorted by downloads, likes, or trending. Operators can run /v1/graph/refresh-llms daily to update model counters, add newly discovered models, and refresh model cards. The maintained LLM data surface is model size, FP16 VRAM, INT4 VRAM, source, and freshness.

Architecture

PlanetGPU is built on two open-source projects:

gpuhunt — GPU pricing aggregation across cloud providers
dstack — Infrastructure orchestration and deployment configs

Local LLM Agents

If you have an Ollama server running (e.g., on an RTX 5090), set LLM_BASE_URL and LLM_MODEL in your environment. Agents will use your local model for repo analysis and cost optimization.