FAQ

The questions a FinOps lead actually types, answered in posting shape with the paragraph that governs each answer. For your exact instrument, run it through the classifier.

Fourteen questions

Can we capitalize LLM training costs?

Yes, when the run develops buyer-owned internal-use software and, under ASU 2025-06, it is probable the project completes and no significant development uncertainty remains. Exploratory or novel runs whose feasibility is unproven are expensed as research even with the budget committed.

Is a reserved compute commitment CapEx or OpEx?

Neither at signing. A signed commitment is not consumption; the amount paid ahead is a prepaid asset that unwinds to expense, cost of revenue or a capitalized asset as capacity is drawn down, under ASU 2018-15.

How are prepaid GPU or token credits treated on the balance sheet?

As a prepaid asset at purchase, released to the P&L as the credits are consumed, to cost of revenue when they serve a paid product. Unused credits that expire are written off to expense.

Does ASU 2018-15 apply to AI inference spend?

It applies to the arrangement. Metered access to a hosted model is a service contract, so the fee is expensed as received. ASU 2018-15 only lets you capitalize distinct implementation costs of integrating the hosted service.

How does US GAAP differ from IFRS on the same token spend?

US GAAP uses the ASC 350-40 internal-use model gated by ASU 2025-06; IFRS uses the IAS 38 research-versus-development split with six recognition criteria. Cloud configuration diverges most: ASU 2018-15 permits capitalizing implementation, the IFRIC March 2021 decision generally expenses it.

Is on-demand inference ever capitalizable?

No. Running the model is post-implementation activity, expensed as incurred, or booked to cost of revenue when it is a direct input to a paid product.

When is inference cost of revenue rather than operating expense?

When the token spend is a direct input to a product a customer pays for. The cash cost is the same, but COGS classification sizes AI gross margin, which is what investors read for unit economics.

Can I capitalize a fine-tuning run?

When it produces a buyer-owned artifact used internally and the ASU 2025-06 gate is cleared, yes. If the fine-tuning is embedded in a product you sell, ASC 985-20 applies instead, with capitalization starting only at technological feasibility.

What is the probable-to-complete threshold?

The ASU 2025-06 condition that capitalization can begin only once management has committed funding and it is probable the project completes and the software is used. Approving a budget does not satisfy it on its own.

How much of an AI build is capitalizable?

As a planning heuristic, often around half, with the rest in preliminary evaluation and post-implementation running. Treat it as a defensible range derived from measurable development-window activity, not a fixed percentage you can post without support.

How do I set a useful life for a model that ages in 18 months?

Reflect the short life honestly rather than stretching it. Amortize over the period the asset genuinely delivers benefit, cap it by the commitment term, and review it annually, adjusting prospectively.

Is a committed-use discount a gain or an asset?

Neither. The commitment drives the posting: prepaid amounts are an asset, an unpaid minimum is an executory contract expensed as consumed, and the discount simply lowers each period's charge. An unavoidable unused minimum can create a loss.

How do I split one commitment across lines?

Allocate on measurable, evidenced usage, capacity hours or token volumes by activity, and document the basis at the time. The capitalized portion amortizes, the prepaid unwinds, the expensed portion hits the period.

Does capitalizing token spend flatter EBITDA?

It moves cost below the EBITDA line into amortization, which flatters near-term operating margin. That is exactly why aggressive capitalization is an earnings-quality red flag; the defensible posting reflects the economics, not the metric.