US GAAP Considerations for AI Infrastructure
- Published
- Aug 18, 2026
- By
- Alwyn Kruger
- Topics
- Share
Key Takeaways
- The accounting team should collaborate with project managers and developers, because AI-related costs are easily missed when accounting works in isolation.
- Middleware and supporting software each require a judgment about whether they function as standalone assets or as direct costs of the AI model they support.
- Hardware generally follows ASC 360 if purchased, ASC 842 if leased, and ASC 350-40 if accessed as a service.
- Data storage and data center arrangements turn on whether the contract conveys a software license or provides access to a vendor's service.
AI brings a wide range of costs, each with its own accounting implications. At its simplest, an AI model needs tools to process data, storage to retain it, and an ecosystem that allows the model to function, improve, and scale. Most GenAI models rely heavily on data and require significant infrastructure investment.
Setting up that ecosystem takes planning and coordination across multiple projects and departments.
That complexity makes AI-related costs easy to miss if accounting works in isolation. Accounting teams should stay close to project managers and software developers to understand related software projects and account for them accurately.
To do that, accounting should review AI project budgets project by project, determine whether any projects should be combined for accounting purposes, identify which costs meet the entity’s capitalization thresholds, and keep an open dialogue with the development team about cost types, overruns, and related considerations.
Many AI systems rely on foundation models, including large language models (LLMs), but not all do.
Traditional computer vision, robotics, optimization, scheduling, and rule-based AI may not use foundation models, although some are moving in that direction. Foundation models are pre-trained on large data sets and can be fine-tuned for specific tasks. LLMs are foundation models trained primarily on text.
How Do You Account for Supporting Software or Middleware?
Internal orchestration layers that connect components of the same application are not capable of reuse and are not separate assets.
Internal-use AI Software Models
Entities should not automatically default to accounting for software that supports the AI model as a direct cost of the AI software model in development. Judgment is to be applied to determine whether the middleware or supporting software can function as an independent economic utility and support multiple software programs. If so, it is likely to be recognized as stand-alone software (depending on whether it meets the conditions for capitalization on a stand-alone basis under ASC 350-40).
Entities need to determine the intent of the supporting software and or middleware, because if the AI model and supporting software and or middleware display the following characteristics:
- Go live at the same time
- Future updates to the supporting software and or middleware is only to be conducted if updates are made to the main AI software
- The group of software is retired at the same time, the supporting software and or middleware is most likely to be treated as a direct cost of the AI software model it supports, instead of being capitalized on an individual basis, with associated cost designation being determined by the project phase of the AI software model it supports.
AI Software That Will be Marketed or Sold
The cost associated with supporting software and/or middleware of an AI software model to be marketed or sold is to be accounted for as follows:
Supporting Software
The related accounting treatment is customer-right focused; judgment needs to be applied to determine the following (which do not represent a complete list of attributes): The supporting software is a separately identifiable product, module, or component that is sold, licensed, or marketed independently, or has distinct technological feasibility, revenue streams, or customer rights. If most of these factors are relevant, the associated cost will likely be capitalized as a separate software asset when the supporting software project has reached technological feasibility as per ASC 985-20.
Middleware
Judgment is to be applied to determine if the middleware is to be treated as a separate stand-alone asset. Attributes to consider include, but are not limited to, whether the future customer will be able to independently license the software, whether it has separate pricing and contractual rights, and whether it can be bundled or sold separately. If most of the characters are present, it is likely to be stand-alone software. For it to be capitalized as such, the software should be separately identifiable, have its own development roadmap with independent version upgrades, and achieve technological feasibility on its own merit before costs are capitalized, in accordance with ASC 985-20.
For both the supporting software and middleware, if they are not stand-alone assets, cost is to be capitalized when the supported AI Software model reaches technological feasibility and ceases once the supported software is ready for general release.
How Do You Account for Hardware and Storage?
Hardware Layers
This includes “compute” to train and run models, for example, GPUs, TPUs and CPUs; “memory” to house models and data, for example HBM and RAM; “storage” to store datasets, for example, NVMe and clouds storage; “network” to connect systems, for example, InfiniBand and Ethernet; and “cooling/power” to support operations, for example, liquid cooling and power systems.
Data Storage as a Service
Data Storage as a Service is a virtualized, internet-based service where data is stored on remote servers managed by a provider. It does not become capitalized because it supports capitalizable software. Entities need to determine whether the arrangement provides access to a vendor’s services or transfers a software license. The implementation cost for service contracts is capitalizable under ASC 350-40, is disclosed as a prepaid or other asset on the balance sheet, and is amortized as an operating expense. The fees are to be expensed as incurred. If the agreement with the vendor transfers a software license for self-storage solutions or similar software, it is to be capitalized under ASC 350-40.
Data Centers
Entities use data center space to manage their IT needs and support their operations. Data centers are essential a physical facility or a few physical facilities for storing, processing, and distributing large amounts of data, needed for various business applications. It provides the necessary infrastructure for cloud storage and computing, ensuring entities can manage their data efficiently and securely. Data center space is either leased, acquired, or accessed as a service, with the latter being the most common. ASC 360 guidance applies when a facility or a condominium-style unit is purchased. At the same time, ASC 842 governs accounting if the entity controls the right to use the identified space, and ASC 350-40 applies to service agreements entered into.
Hardware
Similar to data center space, hardware, for example, NAS which are specialized servers that house storage devises, a CPU, network interfaces, and sometimes their own operating system, can be.purchased, leased or accessed as a service, therefore, it is essential to assess the right to control the use of the identified asset in the vendor agreement to determine if it is to be accounted for as an asset acquisition, lease or service agreement, which are guided by ASC 360, ASC 842 and ASC 350-40, respectively.
Allocation of Cost to AI Software
Rarely, data-related infrastructure costs, lease costs, depreciation of compute servers, and laaS related to compute capacity are capitalized as an indirect cost for AI Software developed to be marketed or sold, because it is typically not dedicated to a particular software project.
AI cost accounting requires early coordination between accounting, project management, and development teams. Data, preprocessing software, middleware, hardware, infrastructure, and data center arrangements can each fall under different guidance, including ASC 350-30, ASC 350-40, ASC 985-20, ASC 730-10, ASC 360, and ASC 842. The earlier teams identify cost types and project phases, the easier it is to avoid missed or misclassified costs.
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.
What's on Your Mind?
Alwyn Kruger
Alwyn Kruger is a Director of Technical Accounting with 25 years of experience advising on complex accounting matters, financial reporting, accounting standards, and transaction-related issues.
Start a conversation with Alwyn
