Pretraining vs fine-tuning
Pretraining builds a model from raw tokens and consumes roughly six FLOPs per parameter per token; fine-tuning adapts existing weights with a small fraction of that compute. Fine-tuning buys capability cheaply but inherits the base model's ceiling, license, and roadmap.
Training vs inference economics
Training is a large one-time (and repeated-experiment) cost; inference is a recurring cost that scales with demand. High-volume products are often dominated by inference, which is why serving efficiency and hardware ownership matter more as token volume grows.
Utilization and total cost of ownership
A cluster at 30% utilization costs the same to own as one at 60% but delivers half the work. TCO includes hardware, power, cooling, networking, engineers, and the opportunity cost of capital, not just the sticker price of accelerators.
Custom silicon, trust, and control
Custom chips can improve performance-per-watt and secure capacity, but demand enormous fixed costs, compiler complexity, and fab dependence. Labs also integrate for reasons beyond unit economics: supply assurance, data and IP control, recruiting signal, and iteration speed. The honest quiz answer is usually 'all of the above, weighted by your scale.'