Does ASC 350-40 cover fine-tuning a model for internal use?

Yes. Where fine-tuning produces a buyer-owned artifact used internally, the directly attributable development cost is capitalizable once management has committed funding and completion is probable; after adopting ASU 2025-06, only once any significant development uncertainty is also resolved.

Is production inference capitalizable under ASC 350-40?

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

ASC 350-40

ASC 350-40 Internal-Use Software

The posting

ASC 350-40 is the US GAAP home for software an entity builds for its own internal use, which is where most buyer-owned AI lands: eval harnesses, fine-tuned artifacts, prompt registries and integration code. Until ASU 2025-06 is adopted, it splits a project into a preliminary stage (expensed), an application-development stage (capitalized), and a post-implementation stage (expensed). ASU 2025-06, required for annual periods beginning after 15 December 2027, removes the stages and adds a significant-development-uncertainty test to the existing probable-to-complete condition ASU 2025-06.

What it governs in the token stack

ASC 350-40 applies when the AI you spend on is for internal use: you run it, you do not deliver the software itself to customers as a product. That covers the compute and engineering directed at an internal copilot, an eval suite, a prompt registry, integration code and owned pipeline tooling. Where the AI is embedded in a product sold or marketed to customers, ASC 985-20 governs instead.

The three-stage model

Preliminary project stage: conceptual formulation, evaluation of alternatives and vendor selection. Expensed as incurred ASC 350-40-25.
Application-development stage: coding, configuration and directly attributable compute to build the asset. Capitalized ASC 350-40-25.
Post-implementation and operation stage: training users, maintenance and running the system. Expensed ASC 350-40-25.

Only the application-development stage produces a capitalizable asset, and only costs directly attributable to building it enter that asset. This is why the development window matters more than the size of the spend.

How ASU 2025-06 changes it

The pre-amendment text already required, before capitalization starts, that management authorize and commit to funding the project and that it be probable the project will be completed and the software used as intended, alongside completion of the preliminary stage. ASU 2025-06 drops the stage condition, names the probability test the probable-to-complete recognition threshold, and adds that the threshold is not met while significant development uncertainty exists ASU 2025-06. It is effective for annual periods beginning after 15 December 2027, with early adoption permitted at the start of an annual period; until then the stage model applies.

Questions this posts answers

Does ASC 350-40 cover fine-tuning a model for internal use?: Yes. Where fine-tuning produces a buyer-owned artifact used internally, the directly attributable development cost is capitalizable once management has committed funding and completion is probable; after adopting ASU 2025-06, only once any significant development uncertainty is also resolved.
Is production inference capitalizable under ASC 350-40?: No. Running the system is post-implementation activity, expensed as incurred, or cost of revenue when it feeds a paid product.

Instruments and mechanics this standard decides

Primary sources

Ledger current as of 2026-10-23. A position and a citation, not accounting advice. See how we cite.