The Cost of a Token

The tensoreconomics piece is the best published derivation of cost per token from first principles. It connects the bandwidth arithmetic from the KV-cache quest to dollars, a translation most engineers never learn to make. InferenceMAX is the methodology reference: cost claims as Pareto frontiers across hardware with stated assumptions, no single-number cherry-picking. Read it as a template for making a cost claim you'd defend under cross-examination.

Then the capstone: a cost model for the hardware you actually run, from amortized hardware and power, through measured throughput, to $/M tokens per model and quantization. Almost nobody walks into an interview with a cost model built from real bills and real measured throughput. That's what makes the write-up at the end one of the strongest artifacts on this roadmap.