Meta's AI data centers test how far the federal research tax credit can stretch
Meta treats its AI data centers as research "pilot models" to claim federal tax credits, the New York Times reported, in a year its credits totaled $3.912 billion. Meta has not said how much of that came from the data centers, and tax practitioners doubt that whole facilities of chips can pass the IRS test for research supplies.
The Board Room · Leadership desk
What happened
- Meta reportedly began the practice in 2024, labeling chips bound for AI facilities differently for tax purposes from chips sent to its standard data centers.
- A March 2024 Meta engineering post described two 24,576-GPU clusters on different network designs, built to compare how each supported AI training at larger scale.
- The IRS research-credit audit guide says supplies generally must be nondepreciable tangible property used directly in qualified research.
- Meta spokesman Andy Stone defended the company's use of incentives Congress enacted for domestic investment, citing $57 billion of research spending in 2025.
Why it matters
- Who's exposed exposure If the IRS challenges the strategy and the chips fail the supplies test, Meta stands to keep less of a benefit it has already booked.
- Who has to decide decision Companies claiming AI hardware as research supplies have to tie specific equipment to a specific technical uncertainty, since a pilot-model label alone does not meet the criterion.
- What it sets up precedent An IRS determination on Meta's treatment would give other companies booking AI infrastructure against the credit a reading to measure their own claims against, after those credits are already claimed.
Next story · Product · Oct 4
Trump's AI accord puts three of its four layers of control inside the companies it coversConfirmed16 publishers
Meta went ahead in 2024 despite doubts inside its own finance department. Staff there questioned whether the treatment would hold up under scrutiny, and the company took advice from several law firms before proceeding, according to the Times investigation as summarized by implicator.ai [12]. That account comes from interviews with people familiar with Meta's operations [5]. Meta's filing does not break out how much of the $3.912 billion credit came from data centers, and the reporting does not establish which equipment qualified or how Meta treats it on its books [3][11].
Tax practitioners who work on the credit make the skeptic's case. Andre Shevchuck, a BPM partner who specializes in the credit, called characterizing AI data centers as experimental "kind of wild and out there" [9]. Shawn Marchant, who runs Tanner's credit and incentives practice, said he would be "skeptical" of claiming the credit for every chip in every data center [8]. Meta's position has a better defence in Jeffrey Moeller's reading. The Ivins, Phillips and Barker tax lawyer said commercially available, proven products could qualify as research supplies if they are needed to resolve technical uncertainty in a project. He declined to discuss Meta [13].
Moeller's test fits the experiment Meta has described more easily than the scope the Times reported. The Nvidia chips are proven technology. The uncertain part is the server-rack layouts and the ways of connecting thousands of chips for training [18]. The two clusters in Meta's 2024 engineering post total 49,152 GPUs [19]. The reported labeling covered chips bound for AI facilities in general [4]. Marchant's objection applies most to that second, larger group. Accounting is a second hurdle. If the chips are depreciated, the audit guide's definition of a supply excludes them [7].
The Times reporting is about Meta and no other company. Any other company counting heavy AI spending toward the credit faces the same criterion. Claimed supplies must be used to resolve technical uncertainty in qualified research, and calling a facility a pilot model does not establish that on its own [6]. A finance chief who books AI hardware as research supplies this quarter faces the question Meta's finance staff raised in 2024 [12].
For Meta, the trade-off is a credit booked now against an uncertainty that stays open. Meta's June 30, 2026 quarterly filing lists $18.74 billion in gross unrecognized tax benefits across research credits and foreign transfer pricing [16]. That is about 4.8 times the 2025 research credit, though it covers a broader set of positions [20]. Meta spokesman Andy Stone said the unrecognized tax benefits figure is "simply a mandated accounting measure of uncertainty" [17]. Whether the chip purchases qualify could affect how much of the claimed benefit Meta keeps if the IRS challenges the strategy [10].
What to watch
- Whether the IRS opens a challenge to Meta's treatment of AI chips as research supplies.
- Whether a later Meta filing isolates the data-center strategy within its unrecognized tax benefits or moves the $18.74 billion figure.
- Whether other large AI spenders disclose a similar classification of data-center chips as research supplies.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+15
- Incentives65
- Confidence58
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Meta classifies its AI data centers as "pilot models" to claim federal research tax credits, according to a September 30 New York Times investigation.
- [2]
The approach treats chips bought from suppliers including Nvidia as supplies used in experiments.
- [3]
Meta's 2025 annual filing records $3.912 billion in total research and development tax credits for the year ended December 31, 2025, and does not break out credits from data centers or Nvidia chips.
- [4]
Meta reportedly began using the approach in 2024, labeling chips bound for AI facilities differently for tax purposes from those sent to standard data centers.
- [5]
The account of Meta's internal classification rests on interviews with people familiar with its operations; public filings do not disclose the pilot-model designation.
- [6]
The relevant criterion is whether claimed supplies are used to resolve technical uncertainty in qualified research; calling a facility a pilot model does not by itself establish that its equipment meets that criterion.
- [7]
The IRS research-credit audit guide says supplies generally must be nondepreciable tangible property used directly in qualified research; property subject to depreciation falls outside that definition.
- [8]
Shawn Marchant, who runs Tanner's credit and incentives practice, said he would be "skeptical" of claiming the credit for every chip in every data center.
- [9]
Andre Shevchuck, a BPM partner specializing in the credit, called characterizing AI data centers as experimental "kind of wild and out there."
- [10]
Whether those purchases qualify could affect how much of the claimed benefit Meta could retain if the IRS were to challenge the strategy.
- [11]
Meta has not explained which equipment qualified for the tax credit, and the available account does not establish Meta's exact equipment treatment or an IRS determination on the strategy.
- [12]
Meta sought advice from several law firms before proceeding, amid questions within its finance department about whether the treatment would withstand scrutiny.
- [13]
Tax lawyer Jeffrey Moeller, of Ivins, Phillips and Barker, said commercially available, proven products could qualify as research supplies if needed to resolve technical uncertainty in a project. He declined to discuss Meta specifically.
- [14]
In a March 12, 2024 engineering post, Meta described two clusters of 24,576 graphics processors each, connected by different types of networks, and used the two designs to assess how well each network could support AI training and how it would perform in larger clusters.
- [15]
Meta spokesman Andy Stone defended the company's use of incentives Congress enacted to encourage domestic investment and said Meta had invested $200 billion in research and development over the preceding five years, including $57 billion in 2025.
- [16]
Meta's June 30, 2026 quarterly filing lists $18.74 billion in gross unrecognized tax benefits across research credits and foreign transfer pricing, without isolating the data-center strategy.
- [17]
Stone described unrecognized tax benefits as "simply a mandated accounting measure of uncertainty," covering many unresolved issues.
- [18]
Meta is testing server-rack layouts and ways to connect thousands of chips for AI training; the Nvidia chips themselves are proven technology.
- [19]
The two clusters described in Meta's March 2024 engineering post total 49,152 GPUs.
- [20]
Meta's $18.74 billion in gross unrecognized tax benefits is about 4.8 times its $3.912 billion 2025 research credit.
Sources
2 independent publishers whose own reporting we read for this story.
- finance.yahoo.comMeta reportedly calls its AI data centers ‘pilot models’ — its research tax credits jumped from $700M to $3.9B
2 articles · September 30, 2026
- implicator.aiMeta Claims Research Tax Credits on AI Data Centers
1 article · September 30, 2026
Topics and entities
Topics
- AI Data Centers
- Uncertain tax positions
- Federal research tax credit