LLM Cost Estimator

The LLM Cost Estimator is an interactive Next.js application that helps machine-learning practitioners quickly validate whether a large language model will fit into a particular GPU setup and how much it will cost to run. The calculator combines up-to-date GPU specifications, detailed VRAM breakdowns (weights, activations, KV cache and optimiser state), performance projections and cloud pricing guidance.

Key capabilities

• Automatic Hugging Face introspection – fetch a model configuration from the Hugging Face Hub and auto-populate parameter counts, hidden sizes, attention heads and precision defaults.
• Detailed VRAM analysis – quantify memory consumption for model weights, activations, KV cache and optimiser state with a configurable execution mode (inference or training), precision, overhead factor and batch size.
• Hardware fit recommendations – compare the required VRAM against a curated list of GPUs, highlight whether a selected GPU has enough headroom and propose the closest alternatives.
• Performance estimations – estimate FLOPs per forward pass, tokens per second and milliseconds per token using the selected GPU’s compute throughput and an efficiency factor.
• Cloud cost calculator – explore on-demand pricing across popular AWS, GCP, Azure and independent GPU providers, override hourly rates and receive a total cost projection for the planned runtime.

Getting started

1. Install dependencies:

   npm ci

2. Start the development server:

   npm run dev

3. Navigate to http://localhost:3000 and search for a Hugging Face model such as meta-llama/Llama-2-7b-hf.

Testing

Run the unit test suite to verify estimator calculations:

npm test

Run the full local verification flow:

npm run lint:strict
npm run typecheck
npm test -- --runInBand
npm run validate:math
npm run build
npm run verify:export

The repo baseline is currently Node 24.14.1 with npm 11.11.0 or newer.

Deployment

• The site is built as a static export (output: 'export') and deployed to GitHub Pages from .github/workflows/nextjs.yml.
• Production builds use NEXT_PUBLIC_SITE_URL and NEXT_PUBLIC_BASE_PATH to generate correct canonical URLs, sitemap entries, and static asset paths for the repository Pages URL.

Disclaimer

The estimator uses analytical approximations of transformer memory footprints, throughput and pricing. Results should be treated as indicative; always validate with real workloads before committing to production deployments.

Contributing

Pull requests that improve model coverage, pricing data or UX are very welcome. Please open an issue to discuss substantial changes before contributing.