Meta's AI Data Center Tax Credit Saved $3.9B in 2025

Meta is cutting its US tax bill by billions of dollars with an unusual label. According to the New York Times, the company classifies its AI data centers as "pilot models" and the Nvidia chips inside them as experimental materials. That lets it claim a federal research tax credit meant to reward research and development.

The amounts are large, and they are rising. Meta saved $3.9 billion through the credit in 2025. That is up from $2 billion in 2024 and $700 million in 2023. The Times reports that no other publicly traded company benefits more from this credit.

Experiments at multi-gigawatt scale

The "pilot model" framing is hard to match with how Meta describes these facilities elsewhere.

In July 2025, Mark Zuckerberg said Meta would "invest hundreds of billions of dollars into compute to build superintelligence." The plan rests on several clusters rated at multiple gigawatts:

• Prometheus, the first, is already partly running.
• Hyperion, the second, is meant to grow to 5 GW over several years.

Zuckerberg also wrote last summer: "We have the capital from our business to do this."

By June 2026, Meta was laying out its compute infrastructure in public. It named partnerships with Nvidia, AMD, AWS, Arm and Broadcom, along with its own custom MTIA chips. That reads like a long-term industrial buildout, not a lab trial. In January 2025, Zuckerberg himself said these data centers would "drive our core products and business."

That is the core tension. For investors, the facilities are the backbone of the business. For the tax authorities, they are experiments.

A 1981 law stretched to its limit

The credit comes from a 1981 law. James Shannon, the former congressman who introduced it, told the NYT that it was meant for "people power, knowledge, information." In his view, Meta's use has "gone way, way beyond what anybody could have imagined."

Meta disagrees. It points to $200 billion in R&D spending over the past five years as evidence that its claims are justified.

Meta's own filings flag the risk

The clearest sign of doubt comes from Meta's own filings. In documents submitted to the SEC, the US securities regulator, the company warns that the tax savings could be challenged.

It has also set money aside. Meta's reserves for uncertain tax positions rose 45 percent to $18.74 billion. Companies build these reserves when they think a tax authority might reject part of what they claimed.

Even so, the strategy may pay off if it fails. If the IRS, the US federal tax agency, later recovers the money, Meta will still have had the use of that capital in the meantime. According to the report, that capital helped lift the company's stock price. In other words, the upside comes early and any penalty comes late.

The auditor built the model and is selling it

EY, Meta's auditor, signed off on the approach. The Times reports that EY also helped Meta design the tax credit setup in the first place. It is now pitching the same approach to other companies that want to offset their AI chip purchases.

That puts the audit firm on both sides: it helps build the structure, approves it in the books, and then sells it on.

The Bigger Picture

This story matters beyond Meta's balance sheet. It suggests that the AI infrastructure race is being partly funded by tax rules written decades before GPUs existed. If one company can treat hundreds of billions in chip and data center spending as research, the effective cost of building AI capacity drops. It drops most for the firms that are already spending the most.

That also affects competition. Capital intensity is already one of the biggest barriers in frontier AI, as shown by the sums labs like OpenAI are raising ahead of an IPO. A tax benefit that grows with hardware spending tends to favor the largest buyers. Smaller players with thinner margins get far less out of it.

The EY angle may be the most important detail. If the same structure is being marketed to other companies, this could turn from a single aggressive filing into an industry pattern. It is worth watching whether other large AI spenders start reporting similar credits, and whether their own filings carry similar risk warnings.

The regulatory side is still open. Meta's $18.74 billion reserve shows the company itself does not treat the outcome as settled. Washington is already scrutinising AI labs on other fronts, so the next signals to watch are any IRS challenge, pressure from lawmakers to tighten the definition of research, or changes in how auditors treat these claims. None of that is guaranteed. For now, the gap between what Meta tells investors and what it tells the tax authorities is on the public record.