What the GTT Teaches
The fleet investigated why large models were timing out and why Q2 produced Thai text. The answer was not in the models. It was in the physics of the machine — an iGPU, unified memory, and two inference servers fighting over the same 34GB pool.