The capitalise-versus-expense decision for AI spend.
AI capitalisation is recording internal AI investment and spend as an intangible asset on the balance sheet rather than expensing it through the P&L. Under IFRS, IAS 38§57 permits it only once development-phase criteria are met; under US GAAP, ASC 350-40350-40-25 permits it only for qualifying application-development costs, replaced by a probable-to-complete test under ASU 2025-06 for annual periods beginning after 15 December 2027. Most experimentation and pre-training is expensed.
Page updated 16 November 2026· Sources: IAS 38 (IASB, IFRS Foundation); ASC 350-40 and ASU 2025-06 (FASB)
Recognise the spend as an intangible asset and amortise it over its useful life.
Charge the spend to the P&L in the period it is incurred.
Pick a cost line to route it to capitalise, expense or prepaid under IFRS.
Capitalise or expense: the AI cost lines
Cost taxonomy →The default treatment of each AI cost line under each framework, the gate that decides a conditional line, and the paragraphs that govern it. Conditional means capitalise only from the point the gate is met; everything before it is expensed.
| Cost line | IFRS default | US GAAP default | Deciding gate (IFRS / US GAAP) | Governing paragraphs |
|---|---|---|---|---|
| Pre-training compute | Expense | Expense | Research phase versus development phase / Significant development uncertainty | IAS 38 §54; IAS 38 §68; ASC 350-40-25; ASU 2025-06 probable-to-complete |
| Fine-tuning compute | Conditional | Conditional | All six development-phase recognition criteria met / Probable to complete and no significant development uncertainty | IAS 38 §57; ASC 350-40-25; ASU 2025-06 probable-to-complete |
| Data acquisition and licensing | Conditional | Conditional | Does the data meet the asset recognition definition / Prepaid consumption versus development-phase input | IAS 38 §25; IAS 38 §68; ASC 350-40-25 |
| Data labelling and annotation | Conditional | Conditional | Research phase versus development phase / Application-development stage versus preliminary or post-implementation | IAS 38 §57; IAS 38 §54; ASC 350-40-25 |
| RLHF and alignment | Expense | Expense | Research phase versus development phase / Significant development uncertainty | IAS 38 §54; ASU 2025-06 significant uncertainty; ASC 350-40-25 |
| Prompt and eval engineering | Conditional | Conditional | Durable asset versus period cost / Internal-use software versus operating cost | IAS 38 §57; IAS 38 §68; ASC 350-40-25 |
| Trained model weights (the asset) | Capitalise | Capitalise | Directly attributable development-phase costs only / Qualifying application-development costs only | IAS 38 §57; IAS 38 §66; ASC 350-40-25 |
| Cloud and hosting implementation | Expense | Conditional | Service contract versus controlled intangible / Cloud computing arrangement implementation costs | IAS 38 §68; ASU 2018-15 350-40-15; ASC 350-40-25 |
| Inference and serving | Expense | Expense | Operating consumption / Post-implementation and operation | IAS 38 §68; ASC 350-40-25 |
| Third-party AI APIs | Expense | Expense | Service contract / Service contract with possible implementation costs | IAS 38 §68; ASU 2018-15 350-40-15 |
| MLOps and data pipeline | Conditional | Conditional | Development-phase build versus operating cost / Application-development build versus operation | IAS 38 §57; IAS 38 §68; ASC 350-40-25 |
The five governing standards
Source-of-record index →Intangible assets, including internally generated intangibles
Internal-use software
Software to be sold, leased or marketed
Cloud computing arrangements that are service contracts
Targeted improvements to internal-use software (ASC 350-40)
Enter from a cost line
Priya's route. Find your GPU, data, labelling or hosting spend and see the default treatment and governing standard.
AI cost taxonomy→Enter from a standard
The auditor's route. Read IAS 38 and ASC 350-40 laid out clause by clause, with the cost lines each one governs.
Standards index→Enter from a treatment
The FP&A route. See what capitalise, expense and prepaid do to the numbers, and which standards mandate each.
Treatments→Questions a controller types
Can AI spend go on the balance sheet instead of the P&L?
Some of it, under conditions. Under IFRS, IAS 38 permits capitalisation only for development-phase costs incurred once all six recognition criteria are met. Under US GAAP, ASC 350-40 permits it for qualifying internal-use software costs once completion is probable, now framed by ASU 2025-06. 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 application-development threshold under ASC 350-40, so costs from that point may be capitalised if the recognition tests are met and no significant development uncertainty remains.
What did ASU 2025-06 change?
ASU 2025-06 replaces the project-stage model in ASC 350-40 with a probable-to-complete recognition threshold and adds a significant-development-uncertainty condition that defers capitalisation of novel or unproven software. It is effective for annual periods beginning after 15 December 2027, with early adoption permitted.