Do Cloud, GPU and LLM API Costs Qualify for the R&D Tax Credit?

Quick answer: Cloud computing costs may qualify for the federal R&D tax credit when they are used in qualified research and meet the applicable expense requirements. Section 41 includes certain amounts paid to another person for the right to use computers in conducting qualified research, subject to specific regulatory conditions. Routine production hosting and general IT expenses generally do not qualify. The tax treatment of LLM API fees requires a separate analysis of the services purchased and how they are used. Careful documentation and a supportable allocation method are essential. 

For many software and AI companies, cloud bills now rival payroll. Training a model can burn through thousands of GPU hours. Development teams spin up environments every day. So it’s no surprise that founders and CFOs ask whether those costs can boost their R&D tax credit.

In many cases they can. But the rules require a clear line between research use and everything else.

The Legal Basis: Computer Use Costs

Qualified research expenses (QREs) include wages, supplies and contract research. Within the in-house research category, Section 41 also includes amounts paid or incurred to another person for the right to use computers in the conduct of qualified research.

Cloud computing fits naturally into that language. When you pay AWS, Microsoft Azure or Google Cloud for compute and storage used to develop and test new software, you’re paying for the right to use computers. If that use is part of qualified research, the cost can count.

See how QREs fit together on our research credit page.

What Cloud Costs Commonly Qualify

  • Development environments where engineers build and test new features or products
  • Testing and staging environments used to evaluate performance, scalability or reliability
  • Model training and experimentation, including GPU instances used to train and evaluate AI models
  • Simulation workloads used to test designs before building them
  • Data processing for experiments, when it is part of resolving technical uncertainty

What Cloud Costs Usually Don't Qualify

  • Production hosting that serves paying customers
  • General IT infrastructure, such as email, file storage and internal business apps
  • Backups and disaster recovery for production systems
  • Monitoring of live systems after commercial release
  • Costs for projects that don’t meet the four-part test

The key question is simple: was this computing used to resolve technical uncertainty through experimentation, or to run the business?

GPU Costs: Rented vs. Owned

How a company obtains GPU computing resources can affect the tax treatment of the associated expenses.

  • Rented GPU compute: Payments to a cloud provider or GPU marketplace may qualify under the computer-use expense category when the arrangement satisfies the applicable regulatory conditions and the resources are used in qualified research.
  • Purchased GPUs and servers: The acquisition cost of depreciable equipment generally does not qualify as a QRE under the supplies category. Depreciation deductions or other tax treatment may be available under separate rules.

The distinction is not simply whether a company rents or owns hardware. Eligibility depends on the applicable expense category, the nature of the arrangement, and the use of the resources. Renting GPU capacity does not automatically make the expense eligible.

LLM API Fees: The Gray Area

Many software and AI teams pay per token to access large language models through APIs. Whether these fees qualify as QREs requires a separate analysis.

Section 41 includes certain payments for the right to use computers in conducting qualified research. However, this provision does not automatically establish that every LLM API charge falls within that category. The nature of the service, the contractual arrangement and the underlying activity all matter.

The IRS sources reviewed for this article do not specifically resolve the treatment of LLM API fees. Accordingly, these costs should not be treated as eligible solely because the API is used during development or experimentation.

  • Potentially stronger case for review: API usage connected to systematic experimentation with alternative model configurations, retrieval methods or technical approaches to resolve uncertainty in developing a new or improved product.
  • Potentially weaker case for review: API usage that powers routine customer-facing features, supports ordinary business tasks or performs work without a qualifying research purpose.

Even when the underlying activity meets the research requirements, the API charge must independently fit an eligible expense category. Companies should retain usage logs, document the purpose of the work and obtain a supportable tax analysis before including these fees in a claim.

How to Allocate Cloud Costs

Cloud invoices rarely separate research from production. Here’s a practical approach:

  1. Tag resources. Use cloud tags or separate accounts for development, testing, training and production.
  2. Map accounts to projects. Link each environment to a specific business component.
  3. Use cost reports. AWS Cost Explorer, Azure Cost Management and Google Cloud Billing reports can break down spending by tag, project or account.
  4. Apply a reasonable method where tags aren’t available, such as usage metrics or engineering estimates backed by data.
  5. Keep the evidence. Save billing exports and tagging policies with your credit study.

Companies with clean tagging often capture more cloud QREs with less audit risk.

A Worked Example

A SaaS company spends $600,000 a year on cloud services. Its tagging shows:

  • Production: $380,000
  • Development and staging: $140,000
  • Model training and experimentation: $80,000

If the development and training work relates to qualifying business components, up to $220,000 of cloud spend may be included as QREs, subject to review. The amount of research credit available depends on the applicable credit calculation method, the company’s QREs and other tax factors.

State Considerations

Many states start with federal QREs but add their own rules, including requirements that research be performed in the state. Where cloud servers are located versus where engineers work can raise questions. Check our states served page for state-specific opportunities.

Documentation Checklist

  • Cloud billing exports by account, tag or project
  • Tagging policies and account structure
  • Project lists linking environments to business components
  • Engineering records showing experimentation
  • API usage logs and purposes for LLM costs

Why Cloud Costs Get Overlooked

Many companies never include cloud costs in their R&D credit because the bills sit with IT or finance, not engineering. The invoices are large and hard to read, and nobody owns the job of splitting them. As a result, companies often claim engineering wages but leave the computing those engineers relied on out of the calculation entirely.

Fixing this usually starts with a short conversation between engineering and finance. Which accounts or projects are used for development? Which are used for training models? Which serve customers? Once that map exists, the cost data often follows quickly.

Common Mistakes With Cloud QREs

  • Claiming the entire cloud bill. Production and business systems must be excluded.
  • Using a flat percentage with no support. Allocations need a reasonable, documented basis.
  • Ignoring shared environments. When development and production share resources, you need a defensible split.
  • Forgetting state sourcing rules. Some states care where research is performed.

The Bottom Line on Cloud Costs

Cloud computing is one of the fastest-growing costs in software and AI development, and it’s one of the most commonly missed parts of the R&D credit. The opportunity is real, but so is the need for a clean, documented split between research and production. Companies that tag resources, map environments to projects and document the purpose of API usage can often add meaningful value to their credit while keeping the claim defensible if the IRS asks questions.

Key Takeaways

  • Cloud computing used for qualified research can count as a QRE.
  • Production hosting and general IT generally don’t qualify.
  • Rented GPU time may qualify; purchased GPUs usually don’t.
  • LLM API fees are a gray area that needs careful documentation.

Frequently Asked Questions

Do AWS costs qualify for the R&D tax credit?

AWS costs can qualify when they’re used for qualified research, such as development, testing or model training. Production hosting and general IT costs usually don’t.

Do GPU costs count as qualified research expenses?

Rented GPU time used for qualified research may count as computer use costs. Purchased GPUs are depreciable property and generally aren’t QREs.

Can LLM API fees be included in the R&D tax credit?

Possibly. There’s no specific IRS guidance, so API costs used for experimentation during development need careful documentation and a reasonable position.

How do I separate research cloud costs from production costs?

Tag cloud resources or use separate accounts for development, testing, training and production, then map them to specific projects.

Does production hosting qualify for the R&D credit?

Generally no. Hosting that serves customers after commercial release isn’t part of qualified research.

Want to capture your cloud QREs? Get a free estimate from TaxDrone.AI.