Meta has told the US Internal Revenue Service that its AI data centres should be treated as experiments rather than long-lived assets, The New York Times reports.

The classification matters because of how the IRS taxes research and development spending. Under rules that tightened after 2022, companies must amortise R&D expenses over several years instead of deducting them immediately. How Meta characterises its massive server farms and the custom silicon and GPU fleets inside them directly affects how much tax it pays, and when.

The dispute puts a number on the scale of Meta's infrastructure buildout. The company has committed tens of billions of dollars to AI-capable data centres, including facilities across the United States, and has said capital expenditures will keep climbing as it trains larger models. Treating that spending as experimental R&D rather than depreciable property would allow Meta to account for it on far more favourable terms.

The IRS sees it differently. According to the NYT report, the agency is challenging Meta's position, arguing that data centres are functional infrastructure — buildings, power systems, racks of servers — built to serve users for years, not open-ended research projects with uncertain outcomes. The outcome of the fight could reshape how every hyperscaler accounts for AI infrastructure, from Microsoft and Google to Amazon.

For the semiconductor supply chain, the stakes are considerable. Hyperscaler capital expenditure is the single largest demand signal for advanced AI accelerators, high-bandwidth memory and networking silicon. If the IRS forces a stricter accounting treatment, it could compress the after-tax return on data centre investment and slow the pace of orders flowing to chip suppliers. If Meta prevails, it would effectively subsidise the AI buildout through the tax code and encourage rivals to reclassify their own infrastructure spending the same way.

Meta is not the first company to test the boundary between infrastructure and experiment. The NYT notes that the R&D amortisation rules introduced by the 2017 Tax Cuts and Jobs Act, which took effect for expenses incurred after 2021, have already produced a wave of disputes between large technology firms and the tax authority. AI spending has now become the largest and most contested frontier of that conflict.

The case also lands at an awkward moment for Meta politically. The company has cast itself as an American AI champion, promising domestic investment and jobs, while simultaneously seeking a tax treatment that reduces what it owes the US government on that same spending. Lawmakers scrutinising both Big Tech taxation and AI infrastructure policy are likely to watch the dispute closely.

Neither Meta nor the IRS has publicly quantified the exact tax difference at stake, and the proceedings remain confidential. A resolution could come through settlement, administrative litigation in tax court, or legislative reform of the R&D amortisation rules, an option industry lobbyists have pushed for years.

Whichever way the ruling goes, it will set a precedent for how the hundreds of billions of dollars earmarked for AI data centres over the next decade are taxed — and by extension, how quickly that money turns into orders for the chipmakers supplying them.