AI Infrastructure and the R&D Tax Credit: When Does a Data Center Become a Pilot Model?
A recent New York Times investigation into Meta’s use of the Research and Development (R&D) Tax Credit raises a question that extends well beyond one company—or even the technology industry:
Can a multibillion-dollar piece of infrastructure itself be part of qualified research?
According to the September 2026 reporting, Meta has treated certain artificial intelligence data centers as experimental “pilot models” for tax purposes and has reportedly included certain equipment used in those facilities, including expensive AI chips, as supplies in calculating its research credit.
The dollar amounts have attracted attention. Meta’s public filings report $3.9 billion of R&D tax credits in 2025, compared with approximately $2 billion in 2024 and $700 million in 2023.
Those disclosures, however, do not tell us how much of those credits relate to AI data centers, nor do they provide the underlying project documentation necessary to determine whether Meta’s particular position satisfies Section 41.
For taxpayers, that may not be the most interesting question anyway.
The broader question is:
When a company builds expensive tangible property in order to resolve technological uncertainty, where is the line between a pilot model used in research and depreciable property used in the taxpayer’s business?
What Is a Pilot Model?
The concept of a pilot model is not new.
Treasury Regulation §1.174-2 defines a pilot model as a representation or model of a product produced to evaluate and resolve uncertainty concerning the product during its development or improvement. Importantly, a pilot model can be fully functional.
The regulations provide several examples.
A manufacturer can build a working model of a custom machine to resolve uncertainty regarding its design. An aircraft manufacturer can build an experimental aircraft to evaluate whether a new design will work. Multiple models can be constructed when necessary to test a product under different conditions.
The regulations also make an important point: the ultimate success, failure, sale or use of the product does not determine whether the underlying activity constituted research.
A fully functioning asset can therefore have been experimental during its development.
That principle becomes particularly interesting when applied to today’s large-scale technology projects.
A Data Center Is Not a QRE Simply Because It Is Innovative
The fact that a company builds something technologically advanced does not make its cost a qualified research expenditure.
Section 41 contains much narrower rules regarding the expenditures that can enter the credit calculation.
For supplies, Section 41 generally permits tangible property used in the conduct of qualified research but specifically excludes land, improvements to land and property of a character subject to depreciation.
That distinction is critical.
A taxpayer cannot simply characterize an entire facility as experimental and conclude that the cost of the facility is a QRE.
The IRS’s own research credit audit guidance instructs examiners to scrutinize prototype expenditures to determine whether a prototype—or components contained within it—is property of a character subject to depreciation.
Calling something a “pilot model” or “prototype” is not enough.
The underlying facts matter.
The More Interesting Question May Be the Components
This is where the reported Meta position becomes particularly interesting.
The New York Times reporting indicates that Meta’s position involves not simply calling an entire data center a QRE, but reportedly treating certain computer chips and other equipment incorporated into experimental AI data centers as research supplies.
That raises a more nuanced question:
Can components incorporated into a much larger experimental system qualify even when the overall system ultimately becomes an operating asset?
The regulations provide some precedent for thinking about research at the component level.
For example, Treasury Regulation §1.174-2 describes a manufacturer developing a new component for an existing machine. Costs associated with producing the experimental component to resolve design uncertainty may constitute research expenditures even though the machine into which it is ultimately incorporated is used in the taxpayer’s business.
Another example involves the development of a new compressor blade for an aircraft engine. The experimental blade can be treated separately from the aircraft engine on which it is tested.
This does not answer whether an AI chip incorporated into a data center qualifies as a Section 41 supply. Section 174 and Section 41 are not identical, and Section 41 imposes its own limitations on qualified research expenses.
But it demonstrates why the analysis cannot necessarily stop at the facility level.
Where Does Experimentation End and Production Begin?
Perhaps the most important issue is determining when technological uncertainty was resolved.
Research expenditures generally relate to activities undertaken to discover information necessary to eliminate uncertainty regarding the capability, method or appropriate design of a product or process.
Once that uncertainty has been eliminated and the taxpayer begins producing or operating the resulting asset in the ordinary course of its business, the analysis changes.
For a traditional manufacturer, the distinction can sometimes be relatively visible.
A company builds and tests several prototypes, selects the successful design and then begins production.
For an AI data center—or another highly complex piece of infrastructure—the distinction may be less obvious.
A company could potentially be testing questions involving computing architecture, chip configurations, cooling systems, networking design, energy efficiency, model-training performance or the ability to operate a system at unprecedented scale.
At the same time, that infrastructure may also be processing real workloads and providing economic value to the business.
The fact that an asset ultimately works does not establish that uncertainty never existed.
But neither does continuing to optimize an operating asset mean that all of its costs remain research expenditures indefinitely.
The taxpayer must be able to identify what uncertainty existed, what was being tested, which costs related to that experimentation, and when the uncertainty was resolved.
This Is Not Just a Technology-Company Issue
The Meta story is particularly visible because of the size of the investments involved, but the underlying issue is familiar across many industries.
Consider a manufacturer constructing a first-of-kind automated production cell.
An automotive company developing a new battery manufacturing process.
A semiconductor manufacturer constructing experimental fabrication equipment.
A chemical company building a pilot processing system.
An aerospace company constructing a fully functional experimental aircraft.
Or an energy company developing a new system intended to operate under conditions that have not previously been demonstrated.
In each case, the finished property may eventually become useful to the taxpayer.
That does not necessarily tell us whether qualified research occurred during its development.
The relevant analysis is more granular: What technological uncertainty existed? What property was actually produced or consumed to resolve that uncertainty? Which expenditures fall within the categories permitted by Section 41? And which costs relate instead to depreciable property, production or ordinary business operations?
The Dollar Amount Does Not Determine Qualification
The scale of today’s AI infrastructure makes these questions more visible.
A traditional prototype might cost hundreds of thousands or several million dollars. An experimental AI computing environment can potentially involve billions of dollars of equipment.
But Section 41 does not contain a rule that an expenditure stops being related to research because the project is large.
Nor does a large investment become qualified research merely because the underlying technology is cutting-edge.
The same statutory framework must be applied.
That makes documentation and cost identification particularly important for taxpayers pursuing large experimental infrastructure projects.
A defensible position should establish the connection between the technological uncertainty, the experimentation undertaken to resolve it, and the specific expenditures included in the credit calculation.
What Should Taxpayers Take From the Meta Discussion?
The lesson for taxpayers should not be that AI data centers qualify for the R&D Tax Credit.
Nor should the lesson be that they cannot.
We do not have Meta’s tax returns, research credit study, technical documentation or complete facts. Meta’s public financial statements disclose substantial research credits and substantial unrecognized tax benefits related in part to uncertainties surrounding research credits, but they do not provide enough information to evaluate this particular position.
Instead, the Meta story provides a timely reason for companies to reconsider how they analyze experimental property.
For taxpayers developing expensive prototypes, pilot plants, automated production systems, specialized computing environments or other first-of-kind infrastructure, several questions deserve attention:
What technological uncertainty existed when the property was constructed?
Was the property—or a component of it—actually produced to evaluate and resolve that uncertainty?
What experimentation occurred?
Which specific components were consumed or used in that research?
Is any claimed property of a character subject to depreciation?
When was the uncertainty resolved and ordinary production or business use initiated?
Can the taxpayer connect the expenditures being claimed to the qualified research activities?
These are not new questions.
What is new is the scale and complexity of the assets to which companies may increasingly need to apply them.
Old Rules, New Infrastructure
The R&D Tax Credit was created long before artificial intelligence companies began constructing massive computing environments containing thousands of highly specialized processors.
Yet the underlying tax question remains remarkably familiar.
A taxpayer develops something it has not built before. Technological uncertainty exists. Property is constructed and tested to determine whether a particular design or method will work. Some experiments succeed. Others fail. Eventually, the uncertainty is resolved and the resulting technology moves toward commercial or operational use.
AI has not created that fact pattern.
It has simply created a very large—and very visible—version of it.
The challenge for taxpayers is therefore not to determine whether an AI data center qualifies for the R&D Tax Credit.
The better question is:
What research occurred within the development of that data center, what property was used to conduct that research, and which of those expenditures actually meet the requirements of Section 41?
That distinction may ultimately matter far more than whether the asset is called a data center, a prototype—or a pilot model.