Meta Turns AI Data Centers Into a $3.9 Billion Credit

Meta booked a $3.9 billion research tax credit in 2025 after labeling its AI data centers as “pilot models,” people familiar with the company’s operations said. The claim, which began in late 2024, made Meta the largest user of that credit among publicly traded companies.

The same move is already being shopped beyond Menlo Park. Meta’s auditor, EY, signed off on the plan and has since pitched other firms on using the research credit to offset chips bought for AI training, those people said. That is how a 1981 wage-heavy incentive becomes, in practice, a rebate on Nvidia hardware.

$700 Million Became $3.9 Billion in Two Years

Meta’s annual report for 2025 records research and development tax credits of $3.9 billion, equal to 4.6% of pretax income. The credit was $2 billion in 2024 and $700 million in 2023, the year before the data-center strategy. That is more than a fivefold rise in two years.

A review of securities filings found no other public company took more. Apple, Amazon, Alphabet, and Microsoft each still report research credits above $1 billion a year. Amazon’s 2025 research credit was $2.4 billion, down from $2.6 billion in 2024. None of those peers has flagged the research credit as a risk in investor reports, or said they apply it to AI campuses.

THE RESEARCH CREDIT, 2023 TO 2025

Meta’s $3.9 billion is more than a tenth of the $32.1 billion the Joint Committee on Taxation projected the credit would cost the Treasury in 2025. Congressional Research Service analysts Jane G. Gravelle and Mark P. Keightley, writing in February, put the five-year hit at $188.9 billion from 2025 to 2029, second among corporate tax breaks.

On current federal tax, Meta’s 2025 bill was thin even before anyone argues about chips. Domestic pretax income was $79.644 billion. Current federal tax expense was $2.82 billion, or 3.5%. Cash paid to the U.S. Treasury, federal only, was $4.118 billion. The book credit of $3.9 billion sits in the same neighborhood as the cash that actually went to Washington.

Meta’s FY2025 Form 10-K says unrecognized tax benefits “were primarily accrued for the uncertainties with our research tax credits and transfer pricing with our foreign subsidiaries.” Gross unrecognized benefits were $16.45 billion at year-end 2025. By June 2026 that stock of tax bets had risen to $18.74 billion, up 45% from $12.9 billion two years earlier, when the data-center claim was new.

Federal Law Leaves Server Chips off the Supplies List

The credit lives in Section 41 of the tax code. Qualified costs are mostly wages for people doing research, a slice of contract research, computer rentals, and supplies. The statute is blunt on that last bucket: supplies exclude property subject to depreciation. Land is out. So is anything with a depreciable character.

A data-center GPU is bought to last. Meta’s property and equipment, net, stood at $225.724 billion on June 30, 2026, up from $176.4 billion at the end of 2025. Those chips sit on the balance sheet as assets. A supply, for Section 41, is something consumed in the experiment, not a server that will be written down over years.

WHY THE CHIP CLAIM SITS ON THIN ICE

• The depreciation bar: Section 41 says supplies cannot be property of a character subject to depreciation, which is how working accelerators are booked.
• The production cutoff: Research after commercial production of the business component does not qualify, and Meta’s models already rank ads and run consumer products.
• Proven hardware: The IRS has rejected claims on “proven and commercially available equipment and technology,” people briefed on the reviews said.
• The four-part test: Each business component needs technological uncertainty, a process of experimentation, and a qualified purpose, applied piece by piece, not to a whole campus.

Treasury rules define a pilot model as a representation of a product built to resolve uncertainty during development. Layouts, cooling, and the way thousands of chips talk to each other can still be uncertain. The chips themselves are not a mystery. They are the product Nvidia sells, and they already run production traffic.

Lawyers Meta consulted pointed to a 2021 Tax Court fight in which an Indiana shipbuilder lost the credit on new vessel types. Building something new was not enough. The taxpayer had to isolate the uncertain parts and show scientific tests that resolved them. Jeffrey Moeller of Ivins, Phillips and Barker, one of the lawyers people said Meta called, has argued that even a proven product can count if it is a supply used to settle a larger uncertainty. He would not discuss Meta. Alex Sadler, a former Justice Department tax lawyer at Morgan Lewis who also advised, told a Virginia conference that pilot models are “an area of controversy” and that the IRS “doesn’t like” companies dressing commercial production as research. “What if we have a $10 billion data center that does cool stuff that hasn’t really been done?” he asked. “Is all the cost a research expenditure?”

Why Meta Labels Working Campuses as Pilot Models

The idea surfaced in the summer of 2024, as Meta broke ground on tens of thousands of acres of data centers, four people with knowledge of the operations said. One of the largest bills on any AI hall is the chip order. An employee overseeing the build-out proposed tapping the research credit. After months of legal memos, the company began labeling chips bound for AI halls differently from chips bound for ordinary data centers.

Meta runs 28 data centers nationwide. Some of that fleet still serves classic social traffic. The AI slice is what the company now treats, on the return, as a giant experiment that could fail. Finance staff questioned whether the IRS would accept it. The agency has challenged basic supplies at other firms. Meta still took the position.

Characterizing its A.I. data centers as experimental is kind of wild and out there.

Andre Shevchuck, partner, BPM

Andy Stone, a Meta spokesman, said the company is “one of the largest investors in research and development in the United States.” Over the last five years, he said, Meta invested $200 billion in R&D, including $57 billion in the last year, “advancing frontier research, building new technology and supporting American jobs.” He said Meta uses “the tax incentives Congress established decades ago to encourage this type of domestic investment.” The company declined to say what made the campuses experimental, or why tens of billions of dollars of chips qualified.

Stone also called unrecognized tax benefits “simply a mandated accounting measure of uncertainty,” a snapshot of unresolved issues of many kinds. The 10-K still lists research credits first among those issues.

Ernst & Young Is Pitching the Same Chip Play

EY had to sign off. The firm was among the advisers Meta consulted when it built the position, the people said, and it has since taken the chip-and-credit story to other companies buying training hardware. If that sales pitch works, Meta is not a one-off. It is the reference account for a productized tax position.

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.

Lisa De Simone, accounting professor, University of Texas

De Simone previously worked as a tax adviser at EY. Shawn Marchant, who runs credits and incentives at Tanner and spent more than a decade on the research incentive at EY, said he would be “skeptical” of claiming the credit for all the computer chips in all the data centers. Shevchuck offered a hypothetical that might pass: a data center built out “to cure cancer.” Ranking Instagram ads is a longer walk.

Other large tech companies already collect more than $1 billion a year from the same statute, almost entirely on salaries for researchers and engineers, the traditional base of the credit. They have not told investors the credit is a live audit risk. They have not said they re-badge AI halls as pilots. Meta did both, which is why its unrecognized-tax-benefit line moved with the strategy.

The louder problem is not Meta’s 2025 return. It is a template. If the IRS lets a production GPU fleet count as a consumed supply, every buyer of training chips has a map. The credit then stops paying for uncertain lab work and starts paying for scale.

Tax Court Still Has the Zuckerberg Options Case

This is not Meta’s first stretch of Section 41. In 2013 the company treated $4.1 billion of stock options Mark Zuckerberg exercised as research wages, on the theory that he helped invent software such as News Feed. That slice supported about $355 million of a broader research-credit claim of more than $618 million for 2012 and 2013. The IRS disallowed the options piece. The fight is still in U.S. Tax Court.

A 1995 Sun Microsystems case had allowed employee stock-option income to count as research wages, which is the hook Meta is still hanging on. The agency has not folded. Aggressive research-credit positions at large companies often end in a settlement that still leaves the taxpayer ahead, because the cash was cheap in the meantime and the stock had the use of it.

FROM THE 1981 CREDIT TO THE 2026 FILINGS

1. 1981: The Economic Recovery Tax Act creates the research credit, first set at 25% and written as a temporary spur after Japan’s tech rise.
2. 2013: Meta claims Zuckerberg’s $4.1 billion option payout as research pay; the IRS later seeks to recover $355 million.
3. 2015: The PATH Act makes the credit permanent in its current form.
4. Summer 2024: A Meta build-out employee proposes using the credit on AI data-center chips; labeling of AI-bound chips starts later that year.
5. 2025: The research credit reaches $3.9 billion. Joint Committee staff put the government-wide cost at $32.1 billion.
6. September 30, 2026: People familiar with the operations describe the pilot-model treatment; Meta’s unrecognized tax benefits stand at $18.74 billion.

James Shannon, the former Massachusetts representative who sponsored the 1981 credit, told interviewers it was aimed at “people power, knowledge, information,” not “making things.” CRS notes the policy case: social returns to research often run two to four times the private return, so a subsidy can be rational. The same report estimates the effective average credit on a marginal project at 8.2%, far below the 20% regular rate or the 14% alternative simplified rate, because of the way the base is built. Shannon’s objection is about object, not rate. A campus full of working accelerators was not the picture in 1981.

A $31.08 Billion Quarter Left $784 Million of Free Cash

The credit is landing on a company that is already spending faster than it is keeping cash. In the quarter ended June 30, 2026, Meta’s free cash flow was $784 million, down 91% from $8.549 billion a year earlier. Operating cash flow was $31.86 billion. Capital spending, including principal payments on finance leases, was $31.08 billion, about 98% of that operating cash.

Q2 2026 CASH AGAINST THE BUILD

• Free cash flow: $784 million, versus $8.549 billion in the second quarter of 2025.
• Capital expenditures: $31.08 billion, with full-year guidance narrowed to $130 billion to $145 billion.
• Cash and marketable securities: $90.26 billion as of June 30, 2026.
• Long-term debt: $83.66 billion, up from $58.744 billion at year-end 2025.

Revenue in the quarter was $60.801 billion, up 28%. Research and development expense was $21.656 billion. The effective tax rate for the quarter was 16%, and the company said it expects 15% to 17% for the rest of 2026. Net income was $15.848 billion. The ads engine is throwing off cash. The server bill is taking it back.

AI is accelerating our core business today, powering our next generation of products, and opening the door to entirely new enterprise opportunities. The results are already showing, and I’m optimistic about the potential ahead.

Mark Zuckerberg, founder and CEO, Meta Q2 2026 earnings release

That is the investor line. The tax line still needs the campuses to look unfinished. Both can be filed in the same year. Only one of them has to survive an exam.

The Copycat Path Turns a Research Credit Into GPU Policy

A 1980s statute was written for labs that might fail. The clusters now in service train a model one day and serve commercial inference the next, on the same floor. That blur is real, and it is also the opening. If training and serving cannot be pulled apart in an audit, the production fleet inherits the pilot label. The credit then follows the purchase order, not the experiment.

This has gone way, way beyond what anybody could have imagined.

James Shannon, former U.S. representative, sponsor of the 1981 research credit

Meta already has a second, larger Tax Court docket over profits the IRS says were shifted through the Cayman Islands, a claim the agency has put near $16 billion in tax and penalties. Research credits are the smaller fight on paper and the one other AI buyers can copy without a Caribbean structure. EY does not need the Cayman case to sell a chip memo.

If examiners unwind the supplies theory, Meta owes tax, interest, and a dent in a credit it has already shown to investors. If they do not, the next 10-K cycle will fill with quieter versions of the same claim, and Section 41 will have become industrial policy for accelerators without Congress rewriting a line. The return is already on file. The copycats are being briefed.