Training builds a model once; inference runs it every time a user asks a question. Over time, inference becomes the dominant share of compute spending for widely used AI products. Its cost per query therefore shapes who earns margin in the AI economy.
Falling unit costs
Hardware improvements, model compression, and competition among model providers have tended to reduce the price per unit of AI output. Falling input costs benefit companies that sell an end product to customers and can hold their price, and pressure companies whose product is the input itself.
Three positions in the stack
- Model providers: high fixed costs, competitive pricing, potential commoditisation.
- Cloud and hardware: paid on volume; benefit if cheaper inference drives more usage.
- Applications with distribution: may gain margin as inputs get cheaper, if customers stay.
What to ask management
For application companies, ask how gross margin responds to changes in model pricing, and whether they are locked into a single provider. For infrastructure companies, ask whether usage growth is outpacing price declines. For model providers, ask what differentiates their offering once comparable capability is available at lower cost.
Cheaper intelligence is a gift to whoever owns the customer relationship.
This framework is a lens, not a prediction. It is intended to help investors place a company within the economics of the stack.
