R&D Tax Credits for AI and Machine Learning Companies

The four-part test under IRC Section 41 requires a permitted purpose (a new or improved business component), technical uncertainty, a process of experimentation, and reliance on principles of engineering or computer science. Machine learning development often meets all four almost by definition. Training a model involves hypothesis formation, iterative testing, and measurable evaluation against benchmarks. The outcome is genuinely uncertain. You do not know in advance whether a particular architecture, training dataset, or hyperparameter configuration will produce acceptable performance. This is not routine software development where the path to a working solution is well understood.

Designing and testing novel model architectures or adapting existing architectures for new problem domains qualifies. Developing custom training pipelines, including data preprocessing, augmentation, and feature engineering, involves technical uncertainty. Building evaluation frameworks and benchmark suites to measure model performance against defined thresholds is qualifying work. Developing inference optimization techniques for production deployment, such as model quantization, pruning, or distillation, is eligible. Creating synthetic data generation systems to address training data gaps also qualifies. Fine-tuning foundation models for domain-specific tasks, where performance targets require experimentation with prompting strategies, LoRA configurations, or training data curation, is increasingly common and typically qualifies.

Routine data labeling and annotation, unless you are building novel annotation tools, does not meet the technical uncertainty bar. Using off-the-shelf ML APIs like calling a pre-trained model through an API without modification is not R&D. Standard data cleaning and ETL pipeline work using established tools and methods rarely qualifies. Model monitoring and maintenance in production, where you are not developing new capabilities, is generally excluded. The distinction comes down to whether the work involves genuine experimentation or follows a known playbook.

Amounts paid for the right to use computers in qualified research, such as cloud compute for training runs, hyperparameter sweeps and development-environment inference testing, can be qualified research expenses under §41(b)(2)(A)(iii). Production inference costs generally are not. The key is that the compute must be used in the process of experimentation. For example, a $50,000 GPU bill for training runs during model development qualifies. The same cluster running inference for paying customers does not. Keep your cloud bills separated by environment, or tag resources in AWS, GCP, or Azure so you can isolate development and training costs from production workloads.

A seed-stage AI company in its first year of research with $1,132,500 of QREs would claim a federal credit of about $67,950 under the alternative simplified credit. This is a hypothetical example.

• Wages: four US-based ML engineers paid $200,000 each spend about 85% of their time on model development and evaluation. Because that is 80% or more, the substantially all rule lets all $800,000 count.
• Cloud compute: $420,000 of GPU spend, of which $300,000 is training runs, sweeps and evaluation in development accounts. The $120,000 of production inference is excluded. QREs: $300,000.
• Contract research: a US firm paid $50,000 to build a custom annotation tool. At 65%, $32,500 counts.

Total QREs: $1,132,500. With no QREs in the prior three years, the credit is 6%: $67,950. If the company elects the reduced credit under §280C(c)(2) to keep its full §174A deduction, the credit is 79% of that, about $53,681.

If the company has gross receipts under $5 million and none before the last five years, it can apply the credit against employer payroll taxes in the quarters after it files, which turns the credit into cash well before the company has taxable income.

Rented compute and purchased hardware are treated differently. Amounts paid to a cloud provider for the right to use computers in qualified research can be QREs. GPUs or servers you buy are depreciable property, and §41(b)(2)(C) excludes property of a character subject to depreciation from qualified supplies, so their cost never enters the credit calculation.

Purchased hardware has its own benefit. Under the One Big Beautiful Bill Act, 100% bonus depreciation is permanent for qualifying property acquired after January 19, 2025, so a cluster bought in 2026 can generally be deducted in the year it is placed in service. For a pre-revenue company that deduction mostly adds to a net operating loss, while rented compute used in research can add to a credit that may be usable against payroll taxes now. That difference is worth modeling before a large hardware purchase.

Data work qualifies only when it is part of the experimentation, and contractor costs count at 65% at most. Paying a vendor to label images under a fixed rubric is routine data collection and usually does not qualify. Paying a US engineering firm to design a new active-learning pipeline that decides which samples to label may qualify, because the method itself is uncertain.

Location matters as much as the task. Many labeling and annotation teams work outside the United States, and research performed abroad is excluded from the credit entirely. Contract terms matter too: the company must bear the financial risk if the work fails and keep substantial rights to the results. A fixed-fee contract that pays only on success, or one that leaves the vendor owning the model, can disqualify the payments.

ML teams already produce most of the evidence the IRS wants; the job is to keep it. Useful records include:

1. Experiment tracking logs (from tools such as MLflow or Weights and Biases) showing hyperparameters, datasets and results for each run.
2. Evaluation reports that state the target metric at the start and the result of each approach.
3. Design documents or research notes explaining why a baseline was not good enough.
4. Cloud bills or tags that separate training and evaluation accounts from production.
5. A quarterly list of business components (models or systems) with the engineers who worked on each and their approximate time.

Keep the failed experiments. A record showing three approaches that missed the target before a fourth worked is the clearest proof of a process of experimentation.

AI software built only for your own back-office use faces a stricter test. Under Treasury Regulation §1.41-4(c)(6), internal-use software must also be innovative, involve significant economic risk, and not be commercially available, on top of the four-part test.

A model you sell or offer to customers is not internal-use software. A model your operations team uses to route support tickets internally is, and it qualifies only if it clears that high threshold of innovation. Most companies building such tools on top of commercial models will find the third prong hard to meet, so treat those projects conservatively.

Since the One Big Beautiful Bill Act, domestic research costs paid in tax years beginning after December 31, 2024, including model development and software work, are deductible immediately under §174A. Research done outside the United States is still capitalized and amortized over 15 years under §174, and it is excluded from the §41 credit.

Many AI startups use researchers or contractors abroad. Track where each person does the work, because that location decides both the deduction timing and whether their cost can count toward the credit. This is general information as of October 2026, not tax advice for your specific facts.