Training compute
Training compute is the GPU and accelerator time consumed to train or fine-tune a model. Its treatment follows the activity it funds: compute spent on research or pre-training is expensed; compute consumed to develop a qualifying, feasible asset can be capitalised from the point the recognition tests are met; committed but unconsumed compute is a prepayment.
Routing by framework
Building or pre-training a model from scratch is typically research phase: the entity cannot yet demonstrate technical feasibility or probable future economic benefits, so IAS 38 requires the spend to be expensed.
Exploratory pre-training generally carries significant development uncertainty, so the probable-to-complete threshold is not met and costs are expensed as incurred.
The compute cost does not carry its own treatment. It inherits the treatment of the activity it funds, which is why the same GPU invoice can be expensed in one month and capitalised in the next as a build crosses the phase boundaryIAS 38§54.
Prepaid compute commitments
Reserved-capacity and committed-spend cloud contracts are common. Where the entity has paid or is obliged to pay for compute it has not yet consumed, the unconsumed portion is a prepayment, released to expense or into a qualifying asset as it is used IAS 38§70.
An entity commits to 10,000 GPU hours for a fine-tuning build and prepays the provider. In the reporting period it consumes 6,000 hours: 2,000 on architecture experiments (research, expensed) and 4,000 developing the qualifying model after the recognition criteria were met (capitalised). The remaining 4,000 unconsumed hours are carried as a prepayment. All figures are illustrative.
- 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