Meta saved $3.9bn in 2025 by calling its AI data centers pilot models

Meta is using a federal research tax credit to cut its tax bill by billions of dollars, according to a New York Times report summarised by The Decoder. The method: classify its AI data centers as "pilot models" and Nvidia chips as experimental materials.
The figures are growing fast. Meta saved $3.9 billion in 2025, up from $2 billion the year before and $700 million in 2023. That makes it the biggest beneficiary of this credit among all publicly traded companies, the Times reports.
The Decoder argues that the "pilot model" label is hard to square with what Meta tells investors and the public. In July 2025, Zuckerberg announced plans to "invest hundreds of billions of dollars into compute to build superintelligence," anchored by several multi-gigawatt clusters. The first, Prometheus, is already partly online. The second, Hyperion, is supposed to scale to 5 GW over several years. "We have the capital from our business to do this," Zuckerberg wrote last summer. By June 2026, Meta was openly detailing its compute infrastructure, including partnerships with Nvidia, AMD, AWS, Arm and Broadcom, alongside its own custom MTIA chips. Zuckerberg also said in January 2025 that these data centers would "drive our core products and business."
The credit dates back to a 1981 law. James Shannon, the congressman who introduced it, told the NYT it was meant for "people power, knowledge, information," and that Meta's use has "gone way, way beyond what anybody could have imagined."
Meta defends the practice by pointing to $200 billion spent on R&D over the past five years. But The Decoder notes that Meta's own accountants see the strategy as legally risky. In SEC filings, the company warns the savings could be challenged, and its reserves for uncertain tax positions jumped 45 percent to $18.74 billion. The Decoder's own assessment is that even if the IRS claws back the money, Meta likely still comes out ahead, because the capital was put to work in the meantime, boosting its stock price.
EY, Meta's auditor, approved the strategy. The firm also helped Meta set up the scheme in the first place and is now pitching the same approach to other companies looking to offset their AI chip purchases, the Times reports.
Key facts
- Meta saved $3.9 billion in 2025 through a federal research tax credit, up from $2 billion the year before and $700 million in 2023.
- To qualify, it classifies AI data centers as "pilot models" and Nvidia chips as experimental materials; the NYT calls Meta the credit's biggest beneficiary among publicly traded companies.
- Meta defends the practice by citing $200 billion of R&D spending over five years, while its SEC filings warn the savings could be challenged and reserves for uncertain tax positions rose 45 percent to $18.74 billion.
- The credit comes from a 1981 law; its sponsor, James Shannon, says Meta's use has "gone way, way beyond what anybody could have imagined."
- EY, Meta's auditor, approved the strategy, helped set it up, and is now pitching it to other companies buying AI chips.
Why it matters
The sums are large and rising: $700 million in 2023, $2 billion the year before 2025, and $3.9 billion in 2025. They come from a 1981 research credit applied to the data centers and chips that underpin Meta's AI build-out. The Decoder contrasts the "pilot model" label with Meta's public statements about multi-gigawatt clusters such as Prometheus and Hyperion, and with Zuckerberg's own words that the data centers will "drive our core products and business." The credit's sponsor says the use has gone far beyond what was imagined.
Who it affects
Meta is the direct beneficiary, and the NYT names it the biggest user of the credit among all publicly traded companies. Other companies may be next: EY, which helped Meta set up the scheme, is pitching the same approach to firms looking to offset their AI chip purchases. The article does not name which other companies EY is pitching the approach to. Taxpayers are affected in the sense that the credit reduces what Meta owes under the federal tax code.
How to use it
This is a reporting story rather than a product or tool, so there is nothing for readers to adopt. The practical point for finance and tax teams is that the approach is being marketed: per the Times, EY is pitching it to other companies offsetting AI chip purchases. Meta's own SEC filings warn that the savings could be challenged, so the story is a reference point for how companies are treating AI infrastructure spending for tax purposes.
How solid is it
The figures and the EY details are the New York Times' reporting, relayed by The Decoder, not original findings. The Decoder adds commentary of its own: that the label is hard to square with Meta's public statements and that Meta likely comes out ahead even if the IRS claws back the money. The article does not give the publication date of the NYT report. The figure for 2024 is given only as "the year before" 2025. Meta's response to the specific "pilot model" label is not given beyond pointing to $200 billion of R&D spending.
Risks and caveats
The article does not say when or whether the IRS will challenge the credit. It does not say whether Meta's claims have been ruled unlawful; only that Meta itself warns they could be challenged. The 45 percent rise in reserves to $18.74 billion is reported alongside this risk, but the article does not say the reserve is solely tied to this credit. The Decoder's view that Meta would still come out ahead even after a clawback is its own hedged assessment ("likely"), not a finding.
“gone way, way beyond what anybody could have imagined”
— James Shannon, the congressman who introduced the tax credit, on Meta's use of it, to the New York Times