Meta Saved $3.9 Billion in 2025 by Calling AI Data Centers Experiments
Meta classified AI data centers as "pilot models" and Nvidia chips as experimental materials, saving $3.9 billion in federal taxes in 2025 under a 1981 research credit.
Updated
Why it matters
- Meta saved $3.9 billion in US federal taxes in 2025 by classifying AI data centers as 'pilot models' and Nvidia chips as experimental materials.
- The tax credit Meta uses dates back to 1981 and was intended to encourage industrial research; Meta's own accountants see the strategy as legally risky.
- In January 2025, Zuckerberg said the same data centers would 'drive our core products and business.'
Meta saved $3.9 billion in US federal taxes in 2025 by classifying its AI data centers as "pilot models" and its Nvidia chips as experimental materials, the New York Times reports.
The classification lets Meta claim a tax credit that Congress created in 1981 to encourage industrial research and experimentation. By arguing that its massive AI infrastructure buildout qualifies as experimental work, Meta reduces its federal tax bill by billions of dollars each year. The $3.9 billion figure covers 2025 alone.
The strategy carries legal risk. Even Meta's own accountants consider the approach legally questionable, according to the New York Times. The company is betting that the classification will hold up under scrutiny, but the position sits in contested territory between genuine research expenditure and ordinary capital investment in production infrastructure.
There is a tension at the center of Meta's argument. In January 2025, Mark Zuckerberg said the same data centers, now framed as experimental for tax purposes, would "drive our core products and business." That statement describes production infrastructure serving Meta's existing commercial operations, not speculative research. Zuckerberg made the remark months before the 2025 tax filings that claim the experimental classification.
Why this matters
The stakes extend beyond Meta's balance sheet. AI companies are pouring unprecedented capital into data centers — facilities packed with Nvidia GPUs that cost billions to build and run. How that spending is categorized for tax purposes determines how much of the bill flows back to the US Treasury. If Meta's approach stands, it establishes a template other AI infrastructure builders can copy, potentially diverting tens of billions in federal revenue as the data center boom accelerates.
The 1981 credit was designed for companies taking genuine research risks — testing new processes, prototyping unproven technology. Applying it to hyperscale data centers that power revenue-generating products tests the boundary of what Congress intended. A successful challenge from the Internal Revenue Service could force Meta to repay the saved taxes, with penalties.
Meta has not hidden its infrastructure ambitions. The company has publicly committed to enormous AI capital expenditure, and the data centers at issue are central to that plan. The gap between Meta's public framing — these facilities power our core business — and its tax framing — these facilities are experiments — is the crux of the legal exposure.
The numbers
-
$3.9 billion: Meta's estimated federal tax savings in 2025 from the research credit classification, per the New York Times.
-
1981: The year Congress enacted the tax credit Meta now relies on, intended to spur industrial research and experimentation.
-
January 2025: When Zuckerberg said the data centers would "drive our core products and business."
What comes next
The IRS could audit the classification and challenge the credit claims. Meta's own accountants already flag the strategy as risky, which means an internal paper trail acknowledges the exposure. If regulators or lawmakers move to close the loophole — either through enforcement or legislation narrowing the 1981 credit's scope — Meta and other AI builders racing to construct data centers would face a materially higher tax burden on the single largest cost line in their AI strategies.
Original: nytimes.com
More from James Calloway
Show full bio
News editor covering industry trends and analytics at AI In Context.
164 articles
Related articles
- Goldman Sachs Sees $1.2 Trillion Big Tech AI Spending Wave by 2027
- Meta launches enterprise business to monetize its $600 billion AI bet
- OpenAI Tells Energy Dept: Fast Permits and Federal Backing for AI Supercomputer Hubs
- Meta Launches Enterprise AI Platform, Poaches MongoDB CEO to Run It
- Meta Connect 2026: Zuckerberg Takes the Stage Amid Smart Glasses Scrutiny