Meta Claimed $3.9 Billion in Tax Credits by Classifying AI Data Centers as Experiments

Meta quietly applied a 1981 federal research tax credit — written to reward scientists running laboratory experiments — to its commercial AI infrastructure by reclassifying server farms in Louisiana and Ohio as "pilot models" and Nvidia GPU purchases as "experimental materials," claiming approximately $3.9 billion in research credits for fiscal year 2025, according to an investigation by The New York Times. The filing revealed that Meta's own auditor, Ernst & Young, both approved the interpretation and is actively marketing the identical strategy to other large AI chip buyers, raising the prospect of an industrywide tax recharacterization that could cost the federal government tens of billions of dollars a year.

Meta's research credit trajectory tells a story of sudden acceleration. The company claimed roughly $700 million in Section 41 research credits for fiscal 2023, approximately $2 billion for 2024, and $3.9 billion for 2025 — a near-sixfold increase over two years, according to its annual securities filing. The Joint Committee on Taxation estimated that all U.S. companies combined claimed $32.1 billion in Section 41 credits in 2025, meaning Meta's single claim represented approximately 12 percent of the entire nationwide program.

How a 1981 Laboratory Law Became a Data-Center Deduction

IRC Section 41, enacted under the Economic Recovery Tax Act of 1981, provides a tax credit for "increasing research activities." To qualify, expenses must clear a four-part test: the spending must constitute a Section 174 expenditure, the research must be technological in nature, it must seek to discover information useful in developing or improving a business component, and substantially all activities must constitute a "process of experimentation."

The pivot that Meta's strategy exploits lies in a definition buried in the Section 174 regulations: a "pilot model" is defined as "any representation or model of a product" built to resolve technological uncertainty. By characterizing AI data centers as pilot models testing whether AI infrastructure can operate at massive scale, Meta argued that the construction and equipment costs — including the Nvidia GPUs inside them — qualify as research expenditures. The GPUs were separately labeled "experimental materials," a designation that permits consumable supplies used in qualified research to be deducted rather than depreciated.

James Shannon, the congressman who introduced the original 1981 legislation, reviewed the approach and told the Times it had gone "way, way beyond what anybody could have imagined." Andre Shevchuk, a tax expert at the advisory firm BPM, described the interpretation as "kind of wild and out there," but acknowledged it might hold if "you had a data center that you're building out to cure cancer" — not one running commercial social-media recommendation systems.

The most direct legal obstacle is the commercial production exclusion embedded in Section 41(d)(4)(A), which explicitly bars credits for research that begins "after the beginning of commercial production of a business component." Meta's data centers serve Instagram Reels ranking algorithms, WhatsApp messaging infrastructure, and Meta AI, all of which are commercially deployed products generating revenue. The company's own public statements undercut the experimental framing. In January 2025, Mark Zuckerberg described the same AI infrastructure as essential to "drive core products." By July 2025, he was describing it as the foundation for "hundreds of billions [of dollars] into compute for superintelligence." Neither characterization describes a laboratory experiment resolving technological uncertainty under the four-part test.

EY Designed the Strategy, Audited the Filing, and Is Now Selling It

The deeper concern raised by the investigation is not Meta's claim alone but the structural position of Ernst & Young in creating it. EY serves as Meta's external auditor, responsible for issuing an opinion on whether Meta's financial statements fairly present the company's position — including the adequacy of its tax reserves. EY also helped design the pilot-model strategy and is now marketing the same approach to other large purchasers of Nvidia hardware.

This dual role creates a conflict that tax experts have highlighted: an auditor cannot objectively evaluate the legal sustainability of a tax position it designed and for which it has a commercial interest in popularizing. EY declined to comment, according to the Times.

Meta's 10-Q for the period ending June 2026 disclosed a gross unrecognized tax benefit balance of $18.74 billion — an accounting reserve covering all uncertain tax positions across the company, including foreign transfer pricing disputes as well as domestic research credits. The figure does not represent a specific repayment obligation for the data-center strategy, and cannot be attributed to any single credit claim. But its magnitude signals that Meta's tax department has identified substantial exposure requiring reservation, and that the IRS is actively examining the company's federal returns for tax years 2020 through 2023.

Meta's Strategy Sits Inside a Broader Political Moment for AI Tax Policy

The Section 41 issue arrives alongside separate Congressional scrutiny. On September 28, Senator Elizabeth Warren, joined by Senators Tina Smith and Jeff Merkley, sent letters to the chief executives of Meta, Amazon, Alphabet, and Microsoft raising concerns about how AI companies are using deductions made available under the One Big Beautiful Bill Act of 2025, which created new bonus depreciation provisions for AI data center capital expenditures. The OBBBA provisions operate through a distinct mechanism from Section 41 research credits, but both represent federal policy choices about subsidizing AI infrastructure — and both are now under political pressure simultaneously.

Meta's Hyperion campus in Richland Parish, Louisiana, anchors the scale of what is at stake. The facility represents a planned investment exceeding $50 billion at five gigawatts of capacity, making it one of the largest planned private infrastructure projects in U.S. history. It is also receiving approximately $3.3 billion in separate state and local tax incentives, meaning the same facility benefits from multiple layers of public subsidy simultaneously.

What the IRS Has Not Yet Done — and What It Could

The IRS has not formally challenged Meta's pilot-model characterization. Under standard audit timelines, the agency is currently examining the 2020–2023 tax years; the 2025 filing is years away from formal examination. The absence of a challenge is not a legal endorsement: Section 41 claims routinely receive scrutiny years after filing, and the commercial production exclusion has been used by the IRS to deny credits in cases where research activities were inseparable from production-scale operations.

If the IRS does challenge and prevail on the commercial production exclusion, Meta would owe taxes on the disallowed credits plus interest — a material exposure given the $3.9 billion claimed in 2025 alone. More consequential for the broader market, a successful IRS challenge before EY's pitch reaches other hyperscalers would effectively close the strategy. If the approach instead spreads to Google, Amazon, Microsoft, and xAI before any regulatory or legislative response, the cumulative annual cost to federal revenues could reach multiples of Meta's individual claim.

The 1981 law that Shannon introduced was designed to make laboratory science cheaper. Whether a five-gigawatt commercial AI supercluster qualifies as a laboratory is a question now sitting with the Internal Revenue Service — and, increasingly, with Congress.