← Back to feed

How Much Does It Actually Cost to Run a Local LLM? (Euros per Million Tokens, Measured)

L3 · BuilderTutorials & GuidesTowards Data Science$· 7/14/2026

Practical benchmarking data helps builders make informed decisions about local vs cloud inference costs for their AI projects.

AI Summary

A detailed cost analysis comparing the actual energy consumption and operational costs of running various local LLMs on an RTX 3090 GPU versus cloud API pricing. The author measured real electricity costs through precise power sampling and found surprising results - smaller models aren't always cheaper, and three local models actually cost more than cloud alternatives. The article provides concrete measurements in euros per million tokens across eight different models.

Excerpt

I measured the actual GPU electricity for eight local models on one RTX 3090 — and the cheapest wasn't the smallest, nor the priciest the biggest. The post How Much Does It Actually Cost to Run a Local LLM? (Euros per Million Tokens, Measured) appeared first on Towards Data Science.

Read Original
0 upvotes · 0 downvotes · 1 min read

Related Articles

L3 · BuilderTutorials & Guides@simonwillison.net
Simon Willison (@simonwillison.net): Made myself a custom Codex pet today (you'll never guess what it was), and had some fun reverse-engineering how the feature works - it generates animated sprite sheets using gpt-image-2 simonwillison.

Simon Willison shares his experience creating a custom animated pet (a pelican riding a bicycle) for Codex Desktop using GPT-5.6 Sol and gpt-image-2. He reverse-engineers the sprite generation process and documents the step-by-step workflow, including all intermediate images and prompts. The article provides practical insights into using AI image generation for creating game-ready animated sprites.

L4 · DeveloperTutorials & Guides@simonwillison.net
Simon Willison (@simonwillison.net): My parallel agent side-project today was having Claude Code port the new Moebius image pinpointing model to ONNX in order to run it entirely in the browser simonwillison.net/2026/Jun/22/...

Simon Willison documents his experience using Claude Code to port the Moebius 0.2B image inpainting model to ONNX format for browser-based execution via WebGPU. He walks through the agent-assisted conversion process and shares a working demo that allows image inpainting directly in the browser without CUDA dependencies.

L3 · BuilderTutorials & Guides@simonwillison.net
Simon Willison (@simonwillison.net): The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible I told it "For all coding tasks use your judgement to decide an appropriate lower power

Simon Willison shares practical tips for optimizing Claude Code's Fable agent, including letting the AI use its own judgment for testing decisions and delegating coding tasks to lower-power subagents to save tokens. The article provides specific prompt examples and real-world results showing improved cost efficiency while maintaining productivity.

L3 · BuilderTutorials & Guides@simonwillison.net
Simon Willison (@simonwillison.net): New TIL: Using uvx in GitHub Actions in a cache-friendly way I finally found a recipe that I like for running `uvx tool-name` in GitHub Actions without downloading a fresh copy of the package every t

Simon Willison shares a cache-friendly approach for running Python tools with uvx in GitHub Actions, using the UV_EXCLUDE_NEWER environment variable to create deterministic cache keys without needing additional dependency files. The technique allows tools to be cached between workflow runs while providing a simple way to bust the cache when needed. This solves the problem of uvx repeatedly downloading packages from PyPI on every workflow execution.

L4 · DeveloperTutorials & GuidesSimon Willison's Blog
Using uvx in GitHub Actions in a cache-friendly way

Simon Willison shares a cache-friendly recipe for using uvx (the pyproject.toml installer from uv) in GitHub Actions workflows, explaining how to set a UV_EXCLUDE_NEWER environment variable to pin tool versions and cache them efficiently between workflow runs. The technique allows developers to run Python tools in CI/CD without hitting PyPI on every execution, reducing network requests and improving performance.