The original headline is: "Why enterprise AI inference should be treated as cost of goods sold, not capex"

Short summary

Enterprise AI transforms software from a fixed cost into a variable one, billing per token like a taxi meter rather than a license. The author argues against buying GPUs to convert variable costs back to fixed, since models spoil rapidly and clusters sit 70-80% idle. The correct reframe: treat inference as cost of goods sold, apply gross-margin discipline, and manage unit economics per feature like any other per-transaction input cost.

• •AI inference is a variable per-token cost, not a depreciable license
• •Owning GPUs trades usage-based metering for utilization risk—worse for spiky demand
• •Reframe inference as COGS and apply existing gross-margin discipline to AI features

Generated with AI, which can make mistakes.