https://blog.grevire.dev/en/tags/amd/Recent content in AMD on GrevireBlogHugoenSun, 25 Oct 2026 15:24:22 +0100https://blog.grevire.dev/en/posts/local-ai-cost/Sun, 25 Oct 2026 08:00:00 +0200https://blog.grevire.dev/en/posts/local-ai-cost/Drafted with Mimo v2.5 and corrected by Deepseek V4.1 Correction (2026-10-26). The first version of this article claimed local inference was 10.72× cheaper than the equivalent API. That figure was wrong: I priced every input token at the fresh rate, ignoring cached input, which is billed at roughly a tenth of it and accounts for most of an agent workload. Recomputed properly, the gap is about 2×, not 10.72× — and once you amortize the GPU it disappears entirely.