Meta classified AI data centers as experiments to cut tax bill

By Splitfeed EditorialLast updated Sep 30, 2026, 17:40 UTC

On September 30, 2026, the New York Times published an investigation finding that Meta has been classifying its large-scale AI data centers as "pilot models" and Nvidia chips as experimental materials in order to claim a federal research and development tax credit that was enacted in 1981. The strategy generated $3.9 billion in federal tax savings for Meta in 2025 alone. According to the report, Meta's own accountants have internally flagged the classification as legally risky, suggesting the company is aware the approach sits on uncertain legal ground.

The core tension the investigation surfaces is a contradiction between two different framings Meta has applied to the same physical infrastructure. In January 2025, CEO Mark Zuckerberg publicly described these data centers as assets that would drive Meta's core products and business, language that signals mature commercial deployment rather than early-stage experimentation. That framing sits uncomfortably alongside the experimental classification Meta uses when filing for the R&D credit, which was originally designed by Congress to subsidize genuinely uncertain, early-stage research. Whether large-scale commercial AI infrastructure qualifies under that original intent is the legal and policy question at the center of the story, and the gap between Zuckerberg's public statements and the tax filings gives investigators and regulators a concrete point of comparison.

None of the three camps have published reactions to this story yet. The pro-AI camp would typically argue that AI development involves genuine experimentation and that using available legal incentives is standard corporate practice that should not be penalized. The anti-AI camp would typically treat this as evidence that large AI companies shift costs onto the public while accruing private benefits, reinforcing arguments for tighter regulation and greater corporate accountability. The middle ground camp would typically call for clearer statutory definitions of what qualifies as research and development in the context of modern AI infrastructure, framing the issue as a policy gap rather than straightforward wrongdoing.

The argument is unlikely to be settled without either a formal IRS audit or legal challenge targeting Meta's specific classification method, or new congressional guidance clarifying how the 1981 credit applies to AI infrastructure at commercial scale. Any IRS ruling or court filing on the matter would be the clearest signal of where this goes next.