Meta's AI data center tax credit: $3.9 billion in 2025 for calling its servers 'pilot models'

Meta's AI data center tax credit is a quiet subsidy of the superintelligence race, and it is bigger than most of Meta's own announcements about it. The New York Times reported on September 30, 2026 that Meta has described some of its AI data centers to the Internal Revenue Service as "pilot models," which lets it claim a research and experimentation tax credit created in 1981. According to the Times's figures, as reported by The Next Web, the credits cut Meta's tax bill by $3.9 billion in 2025, up from $2 billion in 2024 and $700 million in 2023.

That is a tax strategy, not a scandal in itself. The question is whether a hyperscale server farm full of Nvidia chips is an experiment. The IRS has not publicly answered it, and Meta's own filing says the answer is uncertain.

What the credit is for

The research and experimentation credit rewards companies for qualifying research costs. As the Times framed it, the credit rebates supplies used in qualifying experiments, not routine business operations. The reported strategy treats the data centers as pilot models and the chips as experimental material.

The credit's original sponsor is unconvinced. James Shannon, the credit's original 1981 sponsor, is quoted as saying: "This has gone way, way beyond what anybody could have imagined." Meta's reply, from spokesman Andy Stone, is that "Like other companies that invest at this scale, we use the tax incentives Congress established decades ago."

Both can be true. The credit is lawful to claim, and its scope is the part lawyers argue about.

What Meta's own filing says

We opened Meta's annual report for 2025 to check the numbers independently. The filing does not give the $3.9 billion figure, so that remains the Times's reporting. What it does confirm is the risk.

As of December 31, 2025, Meta had net uncertain tax positions of $11.23 billion, which the company says were "predominantly accrued for uncertainties with our research tax credits and transfer pricing with our foreign subsidiaries." Gross unrecognized tax benefits stood at $16.45 billion at the end of 2025, against $15.13 billion a year earlier and $11.67 billion at the end of 2023. The company adds that the ultimate settlement "will depend upon resolution of tax audits, litigation, or events that would otherwise change the assessment."

A note on a figure you may see elsewhere. Several news summaries say Meta's tax reserve rose 45%, from $12.9 billion to $18.74 billion. We searched the filing and the only $18.74 billion we found is a deferred income taxes line in the cash flow discussion, so we could not match that claim and we do not repeat it. The $11.23 billion and $16.45 billion figures come straight from the document.

Why a tax story belongs on a superintelligence desk

The same annual report says Meta is "working to develop the next generation of AI models and advance our vision to build superintelligence, which we define as AI that surpasses human intelligence," and adds: "Although it is inherently difficult to predict when superintelligence may be achieved, we are investing now." Those investments are enormous. The Decoder, summarizing the Times, puts Meta's research and development spending at $200 billion over five years and describes a Hyperion cluster targeting 5 gigawatts.

A tax credit on that spending works like a discount. When the cost of a data center is partly refunded through the tax code, the effective price of compute falls. Compute is the physical race: the labs that can afford the largest clusters set the pace. Meta's discount is paid by the public through lower federal revenue. That is a legitimate policy question for the superintelligence era, because the same public is being asked to trust the labs on safety.

We covered a related fight over who pays for the AI build-out in how flexible data centers could free 100 gigawatts. The two stories rhyme: the cost of the race is being pushed onto grids and ledgers that nobody voted on.

What could happen next

There are three plausible paths, and the evidence does not yet favor one.

• The IRS accepts it. Meta keeps the credits, other hyperscalers copy the structure, and the subsidy grows with capital spending.
• The IRS challenges it. Meta's own reserves signal that it sees this as possible. A loss could claw back billions and force a restatement of positions.
• Congress narrows it. The 1981 sponsor's comment shows the political appetite for tightening the definition. Any fix would ripple to every company claiming similar credits.

The Decoder adds that EY approved the strategy and is now marketing the approach to other companies. If that is right, this is not a single-company quirk but a template.

What we cannot tell you

We did not see the IRS correspondence, Meta's tax returns or the Times article itself, which we did not open. The $3.9 billion, $2 billion and $700 million figures, the pilot model description and the quotes from Stone and Shannon all come through secondary coverage of the Times report. We have no sign the IRS has opened a dispute. Treat anything beyond those caveats as unknown.

Our take

We would not call this evasion. We would call it a large, lawful, untested bet that data centers are experiments, made at a scale that now matters to the federal budget. The useful signal is Meta's own filing: its accountants think the risk is real enough to book billions against it. What we would watch is the next 10-Q for any change in uncertain positions, any sign of an IRS challenge and whether lawmakers move to define what counts as a pilot model.