Meta's tax bill dropped 71% while it spent $72 billion on AI

Why it matters: normally, a company builds an expensive facility and depreciates its cost gradually, spreading the tax deduction out over the years that facility is expected to be useful. Research and experimental expenditures work differently. They can often be deducted immediately, in full, in the year the money is spent.

By classifying its AI data centers as experimental, essentially telling the IRS these are facilities that might not actually work as intended, Meta can front-load billions in deductions rather than spreading them across many future tax years.

• 2024 federal tax expense: $9.6 billion.
• 2025 federal tax expense: $2.8 billion.
• 2025 capital spending: approximately $72 billion, mostly data centers and AI infrastructure.
• Change: a $6.8 billion drop in taxes owed, even as spending surged.

Background: the specific mechanism is the One Big Beautiful Bill Act, signed into law in 2025. It reinstated immediate expensing for domestic research and experimental costs, alongside permanent 100% bonus depreciation, which lets companies deduct the full cost of qualifying equipment in the year it's put into service rather than over time.

In real life it's a bit like being allowed to deduct the entire cost of a new car the year you buy it, instead of spreading that deduction out over the years you actually drive it. The total deduction might end up similar either way, but getting it all at once is worth a lot more right now.

Background: the federal "experimental" classification isn't Meta's only tax advantage tied to AI infrastructure. In Louisiana, the company's Hyperion data center project, a planned 5-gigawatt supercluster that has grown from an initial $10 billion estimate to more than $50 billion, received its own major state incentive.

Louisiana's governor signed a 20-year sales tax exemption covering state and local sales taxes on data center equipment, including the GPUs used to train and run AI models, for qualifying facilities built before 2029. Meta still pays a 1% local sales tax on its purchases there, but the broader exemption represents a significant additional layer of tax relief stacked directly on top of the federal experimental-expensing strategy.

Why it matters: that combination, aggressive federal deductions plus generous state exemptions, means the effective tax burden on this kind of AI infrastructure spending can end up dramatically lower than the sticker-price investment numbers alone suggest. Local business contracts and infrastructure spending tied to the project offer some direct community benefit in exchange, but they don't offset the scale of the tax relief involved.

The catch: no, and that's central to why this has become a bigger story than one company's tax filing. Senator Elizabeth Warren, along with Senators Tina Smith and Jeff Merkley, sent formal letters on September 28, 2026, to the CEOs of Meta, Amazon, Alphabet, and Microsoft, all citing similar patterns.

By the numbers: the letters note Microsoft's federal income tax expense shrank by more than $11 billion between fiscal 2025 and fiscal 2026. Amazon's federal tax bill was nearly $8 billion lower in fiscal 2025 than the prior year. Alphabet's combined federal and state tax expense dropped more than $7 billion over the same stretch.

Background: immediate expensing for research costs isn't a brand-new idea. Section 174 of the tax code, originally enacted in 1954, has allowed companies to deduct research expenditures in the year they occurred for most of the past 70 years.

That changed under the 2017 Tax Cuts and Jobs Act, which required research and experimental expenses paid after 2021 to instead be capitalized and amortized, spread out over five years for domestic research and 15 years for research conducted abroad. That shift quietly raised effective tax bills for research-heavy companies for several years without most of the public noticing.

Why it matters: the One Big Beautiful Bill Act, signed in July 2025, reversed that requirement entirely, creating a new provision, Section 174A, that permanently restored immediate deduction for domestic research costs. It even let companies retroactively accelerate unamortized research costs from 2022 through 2024 into their first tax year after 2024.

That retroactive piece is part of why some companies' 2025 tax bills dropped so sharply all at once. It wasn't just current-year AI spending, but years of previously deferred deductions landing at the same time.

Who's affected: the letters request that each company explain precisely which deductions it claimed connected to AI and data center investments, and disclose any lobbying it conducted before the tax law passed. All four companies were given until October 12, 2026, to respond.

The senators specifically flagged permanent 100% bonus depreciation and retroactive research-and-development expensing as the two provisions doing the heavy lifting here. Their letters also noted that overall corporate tax receipts have fallen 25% even as corporate profits have kept climbing, framing this as a broader pattern rather than an isolated concern about any single company.

What's next: no company has publicly responded to Warren's letter as of this reporting, and Meta hasn't issued a public statement addressing its specific tax classification either. That silence leaves the underlying legal question genuinely open: this appears to follow the letter of a law Congress itself passed in 2025, but whether that outcome matches what lawmakers actually intended is now exactly what's being scrutinized.

Who's affected: every one of these companies has publicly justified massive AI spending partly by pointing to future productivity gains and competitive necessity. Simultaneously reducing near-term tax obligations through provisions tied to that same spending adds a financial dimension to the AI buildout story that goes well beyond just the technology itself.

• This doesn't change any AI product you use directly. It's a corporate tax and policy story playing out behind the scenes of the AI boom.
• It's a useful window into how new tax law gets used in practice, sometimes in ways that weren't obviously anticipated when the legislation passed.
• Watch for the October 12 deadline. How, or whether, these companies respond to Warren's letters will be a real signal of how this story develops next.
• This is playing out across the whole industry, not just at one company. Four of the largest AI spenders all show similar tax reductions during the same buildout period.