OpenAI’s Call for Modernized R&D Credit Rule Makes Perfect Sense

It would be easy to say OpenAI’s request to the Treasury Department last week to modernize a computer-related tax credit rule is just another tech company asking for an even friendlier credit, but the artificial intelligence provider has a point.

The research and development credit is still using antiquated frameworks to determine credit eligibility for things that modern AI research may require access to. Absent an overhaul, the regulations may exclude an entire class of technology research from R&D tax benefits.

To correct the error, the Treasury doesn’t have to turn the R&D credit into a blank check for AI firms. However, it should preserve the rule’s basic anti-abuse principle while recognizing that a company’s reserved or exclusive use of computing capacity in a provider-owned, off-premises data center isn’t the same as owning and operating the infrastructure.

The Treasury essentially is trying to stop a company from getting a credit for renting its own computer back to itself with some ginned-up paperwork. Section 41 of the tax code allows an R&D credit for certain amounts paid to another person for the right to use computers in qualified research. Treasury regulations narrow that by requiring the computer to be owned and operated by someone else, located off the taxpayer’s premises, and not be a computer of which the taxpayer is the “primary user.”

If a taxpayer was a machine’s primary user, the Treasury might reasonably suspect the arrangement was closer to ownership or dedicated internal equipment being moved off-site than true third-party computer access. But modern compute doesn’t fit neatly into that frame.

AI companies and other research-intensive businesses often don’t rent one identifiable machine. They pay cloud or data center providers for access to computing power inside provider-owned facilities, where the “computer” being accessed could mean anything from a GPU (a specialized computer component used to perform many calculations at once, especially useful in AI training) to a rack, a virtual machine, a cloud region, or some other abstraction. That is where the rule starts to wobble under its own specificity.

If a company reserves or exclusively uses provider-owned compute capacity, is it the primary user or the customer of a service designed to provide designated capacity tailored to user need?

That distinction matters, because modern research may require reserved capacity for security and performance purposes. The Treasury shouldn’t treat those ordinary commercial choices as evidence that a taxpayer bought the hotel simply because it booked the conference room for a week.

Is the taxpayer really buying third-party service, or has it effectively acquired, operated, or taken control of the infrastructure? That is the actual modern anti-abuse question.

Photographer: Andrey Rudakov/Bloomberg via Getty Images

The primary user test was built for a world that no longer reflects modern research computing. When the relevant “computer” is discrete, physical, and easy to identify, ensuring other people were also using the computer made sense. The test falls apart when the taxpayer isn’t paying for time on a given machine, but for capacity inside a provider-owned cloud or data center.

The basic regulatory question is more nuanced than the Treasury’s framework assumes. What is the “computer” for purposes of ascertaining identifiable “users?” The answer affects whether modern research costs qualify for a credit Congress designed to encourage research.

AI training and other compute-heavy tasks may need dedicated capacity for security and performance purposes. Excluding them from favorable tax treatment places AI as a second-class citizen in an economy seemingly reorienting itself around AI as the technology of the future.

The current rule risks confusing the form of a modern service with the substance of ownership because of outdated regulations. If dedicated use alone makes the taxpayer the disqualifying primary user, then the Treasury hasn’t prevented abuse so much as punished companies for buying the type of reliable compute capacity modern research requires. A rule intended to disqualify fake outsourcing becomes one that penalizes real outsourcing at best or creates significant market distortions at worst.

This doesn’t mean every compute contract should qualify for the R&D credit. If the taxpayer effectively controls the infrastructure or bears the burden of ownership, the Treasury should be skeptical. But it’s wrong to view exclusive use alone as synonymous with ownership.

The rule should focus more on substance: Is the taxpayer really buying third-party service, or has it effectively acquired, operated, or taken control of the infrastructure? That is the actual modern anti-abuse question.

The Treasury could also add a practical safe harbor. Where the provider owns the equipment, operates and maintains it, keeps it off the taxpayer’s premises, and bears the incidents of ownership, the taxpayer should be presumed to be buying a service.

That presumption of service purchase, rather than ownership, should still be rebuttable. For example, if the taxpayer bears the risk of the hardware’s decline in value or is obviously routing its own equipment through a thin-infrastructure provider for the label, the Treasury has every reason to treat the arrangement as ownership and deny the credit. Commercial availability of similar capacity to unrelated customers is good evidence the taxpayer is buying a service, but it shouldn’t be the complete analysis. The test is who owns the thing in substance.

Modernizing the rule should make the R&D credit both easier to administer and harder to game. While R&D policy doesn’t need to become an AI subsidy machine, it also shouldn’t wander through a data center asking where all the mainframes went.

Andrew Leahey is an assistant professor of law at Drexel Kline School of Law, where he teaches classes on tax, technology, and regulation. Follow him on Mastodon at @andrew@esq.social.

Read More Technically Speaking