Meta Told the IRS Its AI Data Centers Are Experimental, Shaving $6 Billion Off Its Tax Bill

Meta has claimed billions of dollars in federal research tax credits over the past two years by classifying its 28 AI data centers as experimental "pilot models," projects that could conceivably "fail," rather than as standard business operations, a categorization the credit does not permit, according to reporting by The New York Times drawing on four people familiar with Meta's operations and a review of securities filings. The savings escalated sharply: $700 million in 2023, $2 billion in 2024 and $3.9 billion in 2025, nearly $6 billion in total over two years. Meta also labels the Nvidia chips destined for AI training facilities as experimental "supplies," while chips for regular cloud and storage data centers are treated differently, a distinction that lets the pricey hardware qualify for the credit. The company is the single biggest claimant of this credit, exceeding 10% of the entire federal research credit pot, which the Joint Committee on Taxation projects will cost the US government $32.1 billion in fiscal 2025.

What happened

The scheme, as described in the Times reporting, rests on a definitional maneuver. The research tax credit dates to 1981 and was designed to reward genuine experimentation; classifying commercial GPU clusters as experimental was called "kind of wild" by an accounting professor who once advised at EY, per the Times. Meta's own accountants appear to share doubts about the position's durability: the company's reserve for uncertain tax positions jumped 45% over two years, from $12.9 billion to $18.74 billion, with research tax credits listed as the first uncertainty. "Meta is claiming billions of dollars in tax benefits that its own accountants are telling investors are at risk of being overturned by the I.R.S.," said Lisa De Simone, a former EY tax adviser, in the Times report.

Meta's auditor EY approved the strategy and, per the Times, helped design it, and is now pitching the same playbook to other firms buying AI hardware. Meta spokesperson Andy Stone defended the practice: "Meta is one of the largest investors in research and development in the United States... Like other companies that invest at this scale, we use the tax incentives Congress established decades ago to encourage this type of domestic investment." The company separately reports more than $57 billion in R&D spending.

Political pressure is building alongside the accounting questions. On Sept. 28, Sen. Elizabeth Warren led Democrats in letters to the CEOs of Meta, Amazon, Alphabet and Microsoft demanding details on AI and data-center tax deductions and lobbying records ahead of the 2025 One Big Beautiful Bill Act. "Rather than regulate AI to protect American families or address the nation's affordability crisis, the Trump administration and its Republican allies in Congress have prioritized handing out colossal tax breaks to giant corporations," the letter to Zuckerberg stated, as quoted in the findings carried by WebProNews.

Why it matters

The scale of the claim is the story. Meta's federal tax expense fell from $9.6 billion in 2024 to $2.8 billion in 2025 while pretax income rose $15.3 billion over the same period. Capital expenditures hit $72 billion, "the vast majority" tied to data centers and AI, while free cash flow dropped to $784 million, roughly an $8 billion decrease year over year. A credit intended to spur domestic innovation is now, on these numbers, offsetting a substantial share of the tax bill of the largest AI builder in the country.

The exposure is not theoretical. Meta already faces IRS scrutiny over $16 billion in taxes and penalties on profits allegedly routed to the Cayman Islands, and fights the agency in tax court over a prior research-credit dispute seeking to claw back $355 million tied to Zuckerberg stock options. The 45% jump in the uncertain-tax-position reserve signals that Meta's own finance team assigns meaningful probability to the IRS overturning the research-credit claims. Beyond Meta, the Times notes the credit's original sponsor, then-Rep. James Shannon, said its current use has "gone way, way beyond what anybody could have imagined," and six firms (Microsoft, Alphabet, Amazon, Meta, JPMorgan Chase and Nvidia) now account for $83 billion in federal tax breaks, roughly 40% of all corporate breaks. Investor Michael Burry separately said this week he expects a downturn in AI data-center spending, adding a market dimension to the fiscal one.

How it is being covered

The story is carried by 33 sources, with a left/centre/right split of roughly 60/20/20, and 11 of 15 rated outlets sit in the high or very high factuality bands, per ground.news. Coverage skews left of centre, driven by the Times origin and the Warren letters, but the blindspot flag is clear, meaning no significant partisan gap in what is being omitted. The framing across outlets converges on the same core facts: the pilot-model classification, the $6 billion trajectory and the rising reserve for uncertain tax positions.

What to watch

Watch for an IRS response, since the agency has both an active tax-court dispute with Meta over a prior research credit and a fresh $16 billion Cayman routing case, making the research-credit claims a third front. Watch the uncertain-tax-position reserve in Meta's next filings: another jump would signal the company's own accountants see rising risk. Watch Congress, where the Warren-led letters to four tech CEOs could produce hearings or credit-restriction language in the One Big Beautiful Bill Act negotiations. Watch EY, which per the Times is pitching the same structure to other AI hardware buyers, meaning the exposure may not stay confined to one company. Finally, watch Meta's Louisiana Hyperion campus, where state exemptions total roughly $3.3 billion over 20 years and the campus valuation was raised from about $10 billion to more than $50 billion with 5 GW of capacity, up to 20% of state power.

By the numbers

Source snapshot