<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Daldaltown — Judging AI adoption by the numbers</title><description>Judging AI adoption by the numbers</description><link>https://blog.daldaltown.com/</link><language>en-us</language><item><title>Self-Hosting a 70B Model Breaks Even at 2,131 Tokens per Second, Not 5.6 Billion per Month</title><link>https://blog.daldaltown.com/en/posts/2026-08-sllm-serving-breakeven/</link><guid isPermaLink="true">https://blog.daldaltown.com/en/posts/2026-08-sllm-serving-breakeven/</guid><description>The break-even volume is fixed at 5.60 billion billed tokens per month regardless of workload, which is exactly why that number is useless. The real test is your overnight trough throughput, and once you add a latency target and a standby node, self-hosting stops winning at any volume.</description><pubDate>Sun, 23 Aug 2026 00:00:00 GMT</pubDate><category>sLLM</category><category>TCO</category><category>inference serving</category><category>decision making</category></item></channel></rss>