Meta Calls Its AI Data Centers A Science Experiment

Meta wants you to believe its AI data centers are science experiments. The New York Times reporting that anchors this week's episode of What Just Happened suggests otherwise, and the numbers involved are large enough to make even seasoned tax watchers do a double take.

Since 2023, Meta has ramped up its use of the federal Research and Experimentation tax credit, a 1981 law designed to encourage genuine innovation, the kind of work that gave us breakthroughs in medicine and technology. In 2023, Meta claimed $700 million under this credit. In 2024, that number hit $2 billion. By 2025, it had nearly doubled again to $3.9 billion, more than Apple and Alphabet combined, and roughly 15 times what Merck claimed for actual pharmaceutical research.

The Pilot Model Trick

Here is the sleight of hand. Since 2024, Meta has reportedly classified its AI data centers as "pilot models," essentially test versions, and then claimed the credit on the Nvidia chips running inside them. The problem, as Neeta lays out, is that the IRS has historically been strict about this credit: off-the-shelf supplies and commercial products generally do not qualify as "qualified research." A chip Meta buys from Nvidia to run its existing AI products is a far cry from a lab experiment.

Mark Zuckerberg himself undercut the "research" framing when he told public markets that AI is accelerating Meta's core business today. That is a statement about commercial deployment, not scientific discovery. Meanwhile, Meta has guided for $130 billion to $145 billion in capital expenditure this year, most of it AI-related, suggesting these data centers are very much business infrastructure rather than speculative R&D.

Why This Should Worry Taxpayers, Not Just Shareholders

The episode's sharper point is about scale and contagion. The R&D credit cost the federal government $32.1 billion in fiscal year 2025, and Meta alone is capturing about 10 percent of that pool. Layer on top of that more than $3.3 billion in state and local tax breaks from Louisiana for Meta's Hyperion data center, and you start to see a company that has become remarkably skilled at redirecting public money toward infrastructure that, as Neeta puts it, nobody in these neighborhoods actually asked for.

What should alarm observers even more is the detail about EY, Meta's auditor. According to the reporting, some of Meta's own accountants pushed back internally on this use of the credit, yet EY appears to have liked the strategy enough to start sharing it with other clients. That turns a single company's aggressive accounting into a potential template, one that could spread the same R&D mischaracterization across the tech sector and beyond.

What Comes Next

This is not Meta's first rodeo with the IRS. The company is already in tax court over a separate maneuver involving $355 million in credits tied to more than $4 billion of Zuckerberg's stock options dating back to 2012 and 2013. How that older case resolves may be a preview of how aggressively the IRS is willing to challenge Meta's much larger R&D claims now.

Meta, for its part, is unbothered. It has told the public it plans to spend more than $200 billion on research and development over five years. Whether that spending looks anything like the research the 1981 law intended, or simply funds more AI data centers dressed up as pilot projects, is the question regulators and taxpayers should be asking. As Neeta notes, the IRS tends to be precise about its interpretations of tax law. This may be the moment it needs to prove that.