Accounting for AI Data and Consumption Cost

Key Takeaways

• AI token capitalization requires granular consumption record-keeping to comply with the capitalization requirements. As standalone assets or as direct cost of the AIM, they support data costs that costs depend on source and purpose: internally generated data is generally expensed, while purchased data with alternative future use can be capitalized as an intangible asset under ASC 350-30.
• Preprocessing software and LLM’s/Foundation models require a judgement about whether they function as stand alone assets or as direct cost of the AIM model they support.

Artificial Intelligence (AI) is changing how organizations operate, from IT and software development to employee productivity and customer experience. It is common for most software vendors to bill for consumption cost instead of having fixed pricing for the service, which tends to be a significant cost for most entities.

Whether an entity builds, buys, or blends AI technology, the core accounting considerations remain the same. Data is a very important input for most AI models, yet the accounting can be significantly different depending on the nature of the data as well as its use. its use.

How Do You Account for Usage of AI Tokens and Credits

Software vendors are increasingly shifting to software- as- a- service (SaaS) pricing models with consumption based pricing, such as AI tokens, instead of per user or per seat pricing, thereby introducing more variable AI cost driven by unpredictability, A customer entering into a SaaS arrangement accounts for the implementation cost in accordance with ASC 350-40 and ASU 2018-15, while the ongoing subscriptions to the services and usage fees are typically expensed as incurred.

ASC 350-40-30-1 allows for the direct cost incurred during the software development phase to be capitalized, therefore, allowing AI tokens used for activities such as, training the model; cloud compute for model interface or coding/functionality creation with the use of AI, to be capitalized, which can result in a significantly improved adjusted EBITDA, as well as reflecting the entity’s capital investment in intangible AI assets more accurately. The standard requires a clear link between the direct cost incurred and the software asset created, which entails more than a general spread of the token cost amongst the various AI uses.

The participation of the infrastructure, data, security management, software development, and accounting departments during innovation planning is highly recommended, as the AI token usage needs to be tracked in granular detail, for example, the assignment of AI tokens to the various projects, tasks, and resources within the entity.

This means that every token-related expense can be clearly associated with the underlying activity it supports/is used for. Hands-on project management is essential for software projects to identify and track token consumption by project phase, as well as the nature of activity, to allow capitalization of the qualifying consumption cost. Software projects under ASC 985-20 allow capitalization of direct and indirect development costs once the technological feasibility of the project is established.

How Do You Account for AI Training and Output Data?

Data Used for Training a Single AI Model

The purpose of the AI software drives the accounting treatment.

• Under ASC 730-10, costs for data used to train R&D AI software, such as a pilot program, are expensed as incurred.
• For AI software to be marketed or sold, ASC 985-20, internally generated data cost is always expensed, while data purchased to train the foundation model/LLM is generally treated as R&D until technological feasibility is reached, unless the data has an alternative future use. Alternative use requires both expected economic benefit from another use and no dependence on further development after acquisition. After technological feasibility is reached, purchased data costs are capitalized as direct production costs until the software is available for general release. Later purchased training data costs are expensed as maintenance unless they relate to upgrades that meet capitalization criteria.
• For internal-use AI software, ASC 350-40 requires purchased training data costs to be expensed until the preliminary project stage is complete, management has committed funding, and completion and intended use are probable, while internally generated data costs are always expensed. Once those criteria are met, data purchase costs are capitalized during development until the AI project is substantially complete and ready for use. Post-completion purchased training data costs are expensed as maintenance unless they support upgrades that meet capitalization criteria.

Data Used by a Single AI Model to Produce Output

US GAAP does not provide specific guidance, so ASC 350-30 is generally applied by analogy. Internally generated data costs are expensed as incurred. Purchased data is generally capitalized as an intangible asset when acquired individually or with a group of assets in a transaction other than a business combination, an acquisition by a not-for-profit entity, or a joint venture upon formation may meet asset recognition criteria in FASB Concepts Statement No. 5, Recognition and Measurement in Financial Statements of Business Enterprises, even though they do not meet either the contractual-legal criterion or the separability criterion (for example, specially-trained employees).  

Such transactions are commonly bargained exchange transactions conducted at arm's length, which provide reliable evidence of the existence and fair value of those assets. Thus, those assets shall be recognized as intangible assets.  

Even though the data acquired has stand-alone identifiable contract rights and is initially measured at cost, including transaction costs, it only has a single use benefit (aka to be used to produce outcome), therefore not capitalizable under ASC 350-30 or ASC 985-20, but to be expensed. 

Acquired Data to be Used by Multiple AI Software Models

Capitalize purchased data with alternative future use under ASC 350-30 when it has identifiable contractual rights, can be used on its own, supports multiple applications, and provides stand-alone future economic benefits. In that case, the asset is closer to an intangible asset than software development. Data used for multiple activities, which can be any combination of the following activities, is capitalized:

• Training AI software models, including R&D, internal-use, and models to be marketed or sold
• Producing AI model output
• Maintaining AI models
• Amortize the intangible asset over its useful life and allocate the cost to the related projects:
• Amortization allocated to R&D projects is expensed under ASC 730-10.
• Amortization allocated to internal-use AI software training and allocated to an “AI model to produce outputs” are expensed because ASC 350-40 does not allow capitalization of indirect costs.
• Amortization allocated to AI models to be marketed or sold is expensed until technological feasibility is reached. After that point, it is capitalized as an indirect cost of the software in development until the software is available for general release. Only incremental production costs directly tied to development activities are capitalizable; general overhead is not.
•  Amortization tied to ongoing AI output or model maintenance is expensed under ASC 350-40.

Acquired Access Rights to Data for a Limited Time Period

Expense data access fees over the service period, regardless of project phase. Implementation costs for the hosting service’s software infrastructure are deferred as a prepaid asset and recognized over the contract period.

How Are Data Preprocessing Costs Accounted For?

Manual Preprocessing of Data

Involves byte pair encoding (BPE), data migration, dataset cleaning (removing low-quality, duplicate, or toxic data; adding synthetic data), identifying, ranking, Pll masking, and index filtering of data components. These types of costs are to be expensed as incurred.

Development of Pre-Processing Software (PPS)

The initial step is to determine if the software will be used by an internal-use AI model, an AI model to be marketed or sold, or a combination thereof:

• PPS developed for internal-use AI model(s): This involves judgment about whether the software can function independently, supports multiple applications, has a distinct useful life from the application it supports, and can be upgraded or replaced independently. Furthermore, the capitalization requirements under ASC 350-40 to qualify as stand-alone software must be met. The direct cost of a stand-alone PPS software project is to be capitalized once the project reaches the development phase and is halted when the PPS software is substantially complete and ready for intended use. On the other hand, if it is determined to be embedded into internal-use AI software, the accounting treatment of the direct cost is directed by the project phase of the AI software it supports.
• PPS is marketed or sold as a separate software product: The cost incurred is to be expensed as R&D until the PPS project reaches technological feasibility. Once reached, the cost is to be capitalized until the product is ready for general release.
• PPS to support a specific product to be marketed or sold is to be capitalized as part of the main product being developed; therefore, cost is to be expensed until the AI software model that’s supported obtains technological feasibility and is halted once the underlying software is ready for general release.
• PPS only exists to enable other products to be developed, marketed, or sold: Expense PPS development cost as R&D following ASC 730, until and unless portions meet ASC 985-20 criteria to be capitalized as part of a specific product after the capitalization requirements of the underlying product are met.

Acquired PPS

The accounting treatment depends on the purpose of the software and the timing of the cost. The same rationale as discussed for PPS that is internally developed is to be applied.

How EisnerAmper Can Help

Accounting for AI data and infrastructure costs requires judgment at every layer, from training data and preprocessing software to middleware, hardware, and data center arrangements. Missed or misclassified costs can affect reported earnings, balance sheet strength, and audit readiness.

EisnerAmper's Technical Accounting Advisory team works alongside accounting, project management, and development teams to evaluate which standards apply across the AI ecosystem, determine when costs should be capitalized or expensed, and prepare for evolving reporting requirements. To discuss how these considerations apply to your organization, contact our team using the form below.