Can any on-demand inference be capitalized?

No. Running the model is post-implementation activity, expensed as incurred, or cost of revenue when it feeds a paid product.

ExpenseOn-demand inference

On-demand inference

The posting

Metered, pay-as-you-go inference is consumption in the period. There is no commitment and no asset, so it is expensed as incurred by default ASC 350-40-25. When the inference is a direct input to a product a customer pays for, it moves to cost of revenue rather than operating expense, which changes reported gross margin without changing the cash cost.

Why it is expensed by default

On-demand inference is the running of a system, the post-implementation activity that never capitalizes. You consume tokens and settle the bill; nothing carries a future benefit you control. The only real question is which P&L line it hits.

Expense or cost of revenue

Internal productivity or back-office use: operating expense.
Direct input to a paid product or service: cost of revenue, sizing your AI gross margin.
Never a capitalized asset: running the model is not building one.

Cost of revenueIllustrative example, not client data

CapEx / Balance sheet

OpEx / P&L

No entry

DrCost of revenueX

CrCashX

Production inference serving a paid product. Figures illustrative.

Questions this posts answers

Can any on-demand inference be capitalized?: No. Running the model is post-implementation activity, expensed as incurred, or cost of revenue when it feeds a paid product.

Posts to

Primary sources

Ledger current as of 2026-07-24. A position and a citation, not accounting advice. See how we cite.