AASB 138 & AI Model Weights: Capitalising LLM Fine-Tuning Costs in Australia

The rapid evolution of Artificial Intelligence, particularly Large Language Models (LLMs), presents both unprecedented opportunities and significant accounting complexities for Australian businesses. While the operational benefits of AI tools are widely discussed, the financial statement implications of internally developing proprietary AI models, specifically the fine-tuning of LLMs and the creation of unique AI model weights, remain a critical yet often overlooked area. This article moves beyond simply 'using AI' to delve into the intricate balance sheet treatment of internal AI development under Australian Accounting Standards. Graham Chee, FCPA, GRCP, principal of Local Knowledge, writes from a practice that pairs FCPA-grade compliance with Goldman Sachs, BNP Investment Management and Merrill Lynch institutional experience on the highly technical accounting treatment of R&D versus asset creation for custom-trained AI models. Here, we will dissect the application of AASB 138 Intangible Assets, providing a robust framework for Australian entities grappling with the capitalisation of LLM fine-tuning costs. Readers will gain a clear understanding of the distinctions between research and development, the stringent criteria for asset recognition, and the ongoing accounting considerations for these cutting-edge digital assets.

For AI model weights resulting from LLM fine-tuning to be recognised as an intangible asset under AASB 138, an entity must satisfy all six specific criteria of the 'development phase'. These criteria are stringent and require significant evidentiary support. First, the technical feasibility of completing the intangible asset must be demonstrable. For AI, this means proving the custom model can perform its intended function reliably. Second, the entity must have the intention to complete the intangible asset and use or sell it. This requires clear management plans and strategic objectives. Third, the ability to use or sell the intangible asset must be established, considering market demand or internal operational needs. Fourth, the intangible asset must be able to generate probable future economic benefits, such as revenue from new products/services or cost savings. Fifth, adequate technical and financial resources to complete the development and to use or sell the intangible asset must be available. Finally, and crucially for AI, the expenditure attributable to the intangible asset during its development must be reliably measurable [AASB 138, para 57]. This last point necessitates meticulous tracking of computational resources, developer salaries, data acquisition costs, and other direct expenses associated with the fine-tuning process. Without clear evidence for each of these points, capitalisation is not permissible, and all costs must be expensed.

Once AI model weights are recognised as an intangible asset, their initial measurement is at cost. This cost includes all directly attributable expenditures incurred from the date the capitalisation criteria are first met. This can encompass costs such as employee benefits for developers and data scientists, computational resources (e.g., GPU time, cloud services), data acquisition and labelling, and fees for technical consultants directly involved in the fine-tuning process. Subsequent to initial recognition, an entity must choose either the cost model or the revaluation model for that class of intangible assets [AASB 138, para 72]. The revaluation model is rarely applied to internally generated intangible assets like AI model weights due to the absence of an active market for such unique assets. Therefore, the cost model is almost universally adopted. Under the cost model, the asset is carried at its cost less any accumulated amortisation and impairment losses. The useful life of AI model weights needs to be assessed. Given the rapid pace of technological change in AI, a finite useful life is highly probable. Amortisation should be recognised on a systematic basis over the asset's useful life, reflecting the pattern in which the asset's future economic benefits are consumed [AASB 138, para 97]. Regular impairment reviews are also crucial, especially given the dynamic nature of AI, to ensure the carrying amount does not exceed the asset's recoverable amount [AASB 138, para 104].

Achieving AASB 138 compliance for AI model weights hinges on meticulous documentation. This process should be embedded into the AI development workflow. Here's a numbered process for practical application:

1. Project Initiation & Scope Definition: Clearly define the project's objectives, expected outputs (e.g., specific LLM fine-tuning outcomes), and anticipated future economic benefits. Document the problem being solved and the proposed AI solution.
2. Phase Delineation & Milestone Tracking: Establish clear internal milestones that mark the transition from research to development. Document technical feasibility assessments, prototypes, and proof-of-concept demonstrations. Maintain detailed project plans and progress reports.
3. Expenditure Tracking: Implement granular tracking of all costs directly attributable to the AI development. This includes timesheets for personnel, invoices for cloud computing resources, data acquisition costs, and external consulting fees. Categorise expenses by project and activity.
4. Resource Allocation & Management Intent: Document the allocation of financial and technical resources to the project. Maintain records of management's intention to complete, use, or sell the developed AI model, including market analyses or internal business cases.
5. Technical Feasibility Evidence: Keep detailed logs of technical tests, performance metrics, and validation results that demonstrate the AI model's ability to achieve its intended function. This is critical for proving the 'technical feasibility' criterion.
6. Regular Review & Impairment Assessment: Conduct periodic reviews of the project's progress against the AASB 138 criteria. Document any changes in expectations regarding future economic benefits or technical viability, performing impairment assessments as required [AASB 138, para 104].

This structured approach provides the audit trail necessary to support capitalisation claims and withstand scrutiny.