CourseLarge Language Models: The Complete Course · Module 13: Deployment, Operations, and Economics · part 91 of 103
Part 91 · Module 13: Deployment, Operations, and Economics

Lesson 2: Running the model yourself, and what it costs

22 min read·13 Oct 2026
Module 13 · Lesson 2 · About 45 minutes · Needs: Module 1 Lesson 2 (memory for weights and the KV cache), Module 3 Lesson 1 (prefill, decode, the KV cache), Lesson 1 of this module

The situation

It is Tuesday 25 August 2026, week 13. Yesterday's routing rule sends T-1017, Jonas Feldmann's German duplicate-charge ticket, only to a model Quillpath runs itself, because his contract promises EU processing (simulated, Lesson 1). The notebook model was too small: it called the ticket low priority.

Tomasz, who owns the platform, asks the practical questions. Which GPU? Which server? Can the weights shrink, and what breaks? How many tickets at once? What does it cost next to the API? This lesson answers each with numbers, on T-1017.

T-1017 EU clause Quillpath's own model nowhere else 1. Memory 2. Server 3. Precision 4. Batching 5. Bill
T-1017 may only go to a model Quillpath runs (blue: Quillpath's system), so five questions follow: memory, server, precision, batching and the bill.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.