Capitalizing AI Pretraining Costs: What the Guidance Covers

Under the long-standing ASC 350-40 stage model, you could capitalize costs incurred during the application development stage, once you'd decided the project would be completed and used for its intended function, but not during the preliminary or post-implementation stages; ASU 2025-06 replaces the stages with authorization and probable-completion criteria for annual periods beginning after December 15, 2027. Whether a large-scale pretraining run counts as internal-use software development, versus research and development that should be expensed as incurred, depends on facts specific to your project: is the model built for internal use in your own product, and is there substantive uncertainty about whether the effort will even produce a usable model. This is a real judgment call, and it's one to make with your auditor, not one this article can make for you.

Once you're in the application development stage, compute costs for the actual training runs, the engineering time spent building and tuning the training pipeline, and directly attributable third-party costs tend to qualify for capitalization. Costs from the research phase, exploring whether the approach will even work, evaluating architectures, early experimentation, typically don't, since that phase is squarely research and development under most interpretations.

The line between the two phases is rarely as clean in practice as it sounds on paper. A single training program often moves back into research when an early result looks wrong and the team goes back to testing a different architecture, and that shift should move the accounting treatment with it rather than leaving everything capitalized on the strength of the original plan.

Capitalizing spend doesn't change how much cash goes out the door, but it changes when the expense hits your income statement, which affects reported profitability, loan covenants tied to earnings, and how a training-heavy year looks to a board or an acquirer comparing you against a competitor who expensed the same kind of spend instead. Talk to your tax advisor separately from your auditor, since tax treatment of research costs doesn't always follow the same rules as the book accounting decision above.

The useful life you choose has to reflect how long the model will actually generate value before it's retrained or replaced, not the longest defensible number you can find. For a model in a fast-moving domain where you expect to retrain within a year or two, a short useful life is the honest one, even though it means recognizing the expense faster. A useful life stretched out for a model your team is already planning to replace within a year would not hold up to scrutiny.

Write down, at the time you make the call, which phase each major cost fell into and why, what the intended use case was, and what useful life you assumed and how you arrived at it. Process Street is a reasonable place to keep that decision checklist as a repeatable process for the next model, rather than reconstructing the reasoning after the fact when a new model or a new auditor asks about it.

Record these points at the time you make the decision:

• Which stage each major cost fell into, whether preliminary, application development or post-implementation, and why you placed it there.
• The intended use case for the model when you decided to capitalize, so the completion judgment is documented.
• The useful life you assumed and how you arrived at it, tied to your expected retraining cadence.
• Which research-phase costs, such as early experimentation and architecture evaluation, you separated from later training-run costs, and how.

Most finance and accounting staff have applied this guidance to conventional internal-use software, a CRM build-out or an internal tool, and not to a foundation model training run, so the policy needs actual explanation, not just a memo nobody reads. Trainual is one place to turn that policy into something new team members actually work through, rather than a document that sits in a shared drive until the next audit surfaces a gap in understanding.