Meta classified AI data centers as experiments to cut taxes

By Splitfeed EditorialLast updated Sep 30, 2026, 19:10 UTC

On September 30, 2026, the New York Times published an investigation finding that Meta has been classifying its AI data centers as "pilot models" and its Nvidia chips as experimental materials in order to claim a federal research and experimentation tax credit originally enacted in 1981. The strategy saved Meta $3.9 billion in federal taxes in 2025 alone. Meta's own accountants have flagged the approach as legally risky, and internal language from Mark Zuckerberg complicates the company's position: in January 2025, he publicly described these same data centers as infrastructure that would "drive our core products and business," a characterization that sits awkwardly alongside the experimental framing used to justify the tax savings.

The 1981 research credit was designed to encourage speculative, exploratory work rather than the buildout of large-scale commercial infrastructure. Whether Meta's AI data centers qualify under that original intent is the central legal and policy question here. The $3.9 billion figure represents a significant share of what would otherwise be federal tax revenue, and the fact that Meta's own accountants view the classification as legally risky suggests the company is aware it is pushing against the boundaries of the credit's intended scope. Zuckerberg's January 2025 remarks create a specific evidentiary tension that regulators or courts could scrutinize closely.

The Anti-AI camp has responded with sharp criticism. The Roosevelt Institute argued that Meta is exploiting the credit by labeling "proven chips" as research materials to obtain a nearly $4 billion tax break. Accounts including Adam Rogers and Leah McElrath framed the practice as a form of stealth grift, with McElrath quoting the core mechanism directly: Meta claims its data centers are a giant experiment that could fail, specifically to access a credit meant for genuine research. The Daily Edge drew a pointed comparison to legally aggressive tax behavior more broadly. The Pro-AI camp has not yet published reactions; typically, that camp would argue that large-scale AI infrastructure investment carries genuine technical risk and that using available tax incentives is standard corporate practice, not misconduct. The Middle Ground camp has also not reacted; it would typically call for updated tax rules that reflect modern AI investment realities without endorsing what critics call abuse.

The argument is likely to move toward formal ground if the IRS audits Meta's filings or if Congress holds hearings on whether the 1981 credit needs to be amended. Any IRS ruling or litigation outcome would be the clearest test of whether Meta's classification survives legal scrutiny.