July 2026 Summaries

The blog explores the economics and technicalities of deploying large language models (LLMs) using open-weight models, specifically focusing on the financial and mathematical aspects of LLM inference. It discusses the process of self-hosting models like Kimi K2.6 using NVIDIA's B200 GPU, detailing the costs, memory, and performance considerations involved. The text explains how model architecture, quantization, and GPU specifications impact memory requirements and inference speed. It delves into concepts such as arithmetic intensity, memory vs. compute bounds, and batching, outlining how these factors influence inference efficiency and costs. Through a series of formulae and load test results, the blog provides a detailed analysis of how to optimize LLM deployment for cost-efficiency and performance, while emphasizing the importance of load testing and collaboration with inference providers. The author also highlights the complexities of self-hosting, advising most teams to consider inference providers unless significant cost savings can be achieved.