Frequently asked questions
The questions a controller actually types when a CFO asks whether AI spend can go on the balance sheet. Each answer states the position and the governing standard, and rigorously separates what a standard requires from what is a matter of judgement. Where the answer depends on judgement, it says so.
Answers
Can AI spend go on the balance sheet instead of the P&L?
Some of it. Qualifying development-phase costs can be capitalised under IAS 38, and qualifying application-development costs under ASC 350-40. Research, exploration and pre-training are expensed.
Is fine-tuning an LLM capitalisable?
It can be. Fine-tuning a base model for a defined internal use can reach development phase under IAS 38 or the probable-to-complete threshold under ASC 350-40, so costs from the recognition point may be capitalised if the tests are met.
Do GPU hours go on the balance sheet?
Only if consumed developing a qualifying asset after the recognition tests are met. GPU hours spent on research or pre-training are expensed, and committed unconsumed compute is a prepayment.
Does IFRS or US GAAP apply to me?
The framework follows your reporting basis, not a choice. IFRS reporters apply IAS 38; US GAAP reporters apply ASC 350-40 or ASC 985-20. A dual-reporting group must satisfy both.
What did ASU 2025-06 change?
It replaces the ASC 350-40 stage model with a probable-to-complete recognition threshold and adds a significant-development-uncertainty condition that defers capitalising novel or unproven software. It is effective for annual periods beginning after 15 December 2027.
Why can a funded novel model still not be capitalised?
Because completion and intended use must be probable, not merely funded. Significant development uncertainty over novel functionality holds probability below the threshold under US GAAP, and technical feasibility is unmet under IFRS.
What useful life can I defend for an AI model?
Usually a short one, supported by your own model-replacement history and release cadence. The life is reviewed at least annually and should be evidenced rather than assumed.
Are third-party AI API fees capitalisable?
No. Metered access to an external model is a service contract, so the fees are expensed. Only qualifying implementation costs of a related cloud arrangement can be capitalised under ASU 2018-15.
Is data labelling capitalisable?
It can be, when directed at a qualifying development-phase build. Labelling during research, exploration or ongoing operation is expensed.
Can I reinstate cost I already expensed once the model qualifies?
No. Expenditure recognised as an expense in earlier periods cannot be capitalised later, which makes the recognition date consequential.
What does capitalising do to EBITDA?
It lifts current-year EBITDA because the spend leaves the P&L, then becomes an amortisation charge in later periods. The cash spent is unchanged.
What evidence does an auditor expect?
Contemporaneous artefacts for each recognition criterion: a feasibility prototype, a funding minute, a deployment plan, a business case, resource evidence, and time and compute allocation records supporting the attributable cost.
- S1IAS 38 Intangible Assets, IFRS Foundation (IFRS). https://www.ifrs.org/issued-standards/list-of-standards/ias-38-intangible-assets/
- S2Handbook: Software and website costs (ASC 350-40 internal-use software), KPMG (US GAAP). https://kpmg.com/us/en/frv/reference-library/2026/handbook-software-website-costs.html
- S5ASU 2025-06 on internal-use software costs, BDO (US GAAP). https://arch.bdo.com/new-asu-on-internal-use-software-costs-guidance