Published: October 5, 2026 · Last updated: October 5, 2026
TL;DR: Meta told the IRS its newest AI data centers count as “experimental” facilities that might fail, a classification that lets it depreciate the buildings much faster. The New York Times reports that move helped cut Meta’s expected 2025 federal tax bill from about $9.6 billion to roughly $2.8 billion.
Calling a multi-billion-dollar data center an experiment sounds odd until you see what it does to a tax bill.
According to the New York Times, Meta has been labeling a chunk of its newest AI data center buildout as experimental equipment rather than standard long-life property. Tax law treats the two very differently. Standard commercial buildings get depreciated over decades. Equipment classified as experimental, built for a use that might not pan out, can be written off far faster, sometimes within a handful of years. Faster depreciation means bigger deductions now, which means a smaller tax bill now.
The numbers reported are not small. The Times’ reporting, summarized across several outlets including AI Weekly and Crypto Briefing, puts Meta’s expected 2025 federal tax liability at around $9.6 billion before the experimental classification and about $2.8 billion after it, a swing of roughly $6.8 billion. Other outlets have cited figures in the $3.9 billion to $6 billion range depending on which tax year and which facilities are included in the comparison, but every version of the story agrees on the mechanism and the direction: a reclassification, not a change in spending, is driving most of the savings.
Meta’s argument, as described in the reporting, is not unreasonable on its face. AI data centers built for frontier model training genuinely do carry more uncertainty than a typical server farm. Hardware requirements shift fast, chip generations turn over quickly, and a facility designed around today’s training runs could be obsolete for that purpose well before a standard commercial building would be. Calling that uncertain is not pure fiction.
What is drawing scrutiny is the scale and the timing. Meta is in the middle of the largest capital spending program in its history, with AI infrastructure commitments running into the hundreds of billions of dollars over the next several years. A tax strategy that shaves billions off near-term liabilities while that spending ramps up changes the math on how much of the AI boom is actually being funded by shareholders versus by foregone tax revenue.
Meta has not denied using the experimental classification. It has pointed to the uncertainty inherent in AI hardware cycles as justification, consistent with how the Times characterized the company’s position. Tax specialists quoted in the broader coverage note that the IRS has latitude to challenge aggressive depreciation schedules after the fact, but that doing so against a company with Meta’s legal resources is neither quick nor guaranteed.
Other hyperscalers building AI infrastructure at similar scale are widely expected to explore the same approach, if they have not already. Tax code written before anyone imagined data centers this expensive is proving flexible in ways that work heavily in the builders’ favor.
Related: Amazon Is Investigating the Employees Who Testified Against Its Data Centers and A Robotics Startup Backed by Nvidia and Bill Gates Is Reportedly Raising $700 Million.
Bottom Line: Nobody is accusing Meta of breaking the law here, and that is sort of the point. When the tax code lets a company this size cut billions off its bill by filing the right paperwork, the AI spending boom is being subsidized by all of us whether we like it or not.
Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest
