Axis 1 / cost taxonomy
The AI cost taxonomy
Head note
The AI spend stack breaks into ten cost lines, from compute and data through to the trained model weights and the cloud that serves them. Each line has a default treatment and a governing standard under both IFRS and US GAAP. This taxonomy is the entry point: find your spend, then follow it to the treatment and the paragraph that decides it.
The stack at a glance
The default treatments below are the common position, not a verdict for your facts. Most lines are conditional on the phase or stage the spend falls in. Open a line for the framework-by-framework routing and a labelled illustrative example.
Training compute
GPU hours to pre-train or train
Pre-training vs fine-tuning
the phase boundary in one page
Data acquisition
buying or licensing training data
Data labelling
annotation and the labelling pipeline
RLHF and alignment
reward models and human feedback
Prompt and eval engineering
prompts and the evaluation harness
Model weights
the recognised intangible asset
Cloud and hosting
the ASU 2018-15 CCA test
Third-party AI APIs
metered external inference
MLOps and data pipeline
orchestration and pipeline build
Sources of record
- 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
- S3Applying ASC 985-20 and ASC 350-40 to software costs, Crowe (US GAAP). https://www.crowe.com/insights/how-to-apply-asc-985-20-asc-350-40-software-costs