# AI Infrastructure

Engineering notes, architectural deep-dives, and practical playbooks from the Devforth team.

[All](/insights/)[Architecture & Systems](/insights/architecture_and_systems/)[AI Infrastructure](/insights/tech_article/)[DevOps & Reliability](/insights/devops_and_reliability/)[OpenSource Contributions](/insights/opensource/)[Engineering Playbooks](/insights/engineering_playbooks/)

## Latest

[

![Claude Code /usage significantly overestimates output tokens](https://static.devforth.io/posts/logo-svg/2026/2ea9635b-9a22-4a17-842c-fc407a2da1bf-logo-preview-new.svg)

AI Infrastructure Jun 27, 2026

## Claude Code /usage significantly overestimates output tokens

Claude Code’s local /usage may be inflating output tokens by millions. We traced JSONL duplicates and live HTTPS traffic to find what’s really happening.

](/insights/claude-code-usage-significantly-overestimates-output-tokens/)[

![Self-hosted GPT: real response time, token throughput, and cost on L4, L40S and H100 for GPT-OSS-20B](https://static.devforth.io/posts/logo-svg/2026/e2f694c5-cda5-4e61-9cd9-4bae8491e83b-logo-preview-new.svg)

AI Infrastructure Feb 20, 2026

## Self-hosted GPT: real response time, token throughput, and cost on L4, L40S and H100 for GPT-OSS-20B

We benchmarked modern open-source LLMs across several popular GPUs (L4, L40S and H100, RTX 4090) to measure real-world context limits, throughput, latency, and cost efficiency under varying levels of concurrency — as close as possible to real production conditions. Here we share the results.

](/insights/self-hosted-gpt-real-response-time-token-throughput-and-cost-on-l4-l40s-and-h100-for-gpt-oss-20b/)[

![LLM Terminology Guide: Weights, Inference, Effective sequence length, and Self-Hosting Explained](https://static.devforth.io/posts/logo-svg/2026/535346aa-a7d2-48e7-8b1f-6b70a4c95649-logo-preview-new.svg)

AI Infrastructure Feb 16, 2026

## LLM Terminology Guide: Weights, Inference, Effective sequence length, and Self-Hosting Explained

A clear guide to generative AI and LLM terminology. Learn how model weights, quantization, inference, context length, batching, sampling and many more — including how to evaluate vendor APIs and self-host models like GPT-OSS-20B.

](/insights/llm-terminology-guide-weights-inference-effective-sequence-length-and-self-hosting-explained/)[

![GPT-J is a self-hosted open-source analog of GPT-3: how to run in Docker](https://static.devforth.io/posts/logo-svg/2026/d836a55f-0d3d-4206-a07c-765cd2ce29ee-logo-preview-new.svg)

AI Infrastructure Oct 21, 2022

## GPT-J is a self-hosted open-source analog of GPT-3: how to run in Docker

Learn how to setup open-source GPT-J model on custom cheapest servers with GPU. Try to run the text generation AI model of the future and talk to it right now!

](/insights/gpt-j-is-a-self-hosted-open-source-analog-of-gpt-3-how-to-run-in-docker/)

[Image Licensing & Usage Policy](/legal/image-license/) • [Core Engineering Intelligence](/insights/core_engineering_intelligence) • [Coding Agents for Development](/agents-for-code/)
