The New York Times Reports Meta Declares AI Data Centers as Experiments to Claim Research and Experiment Tax Credits
The New York Times reported that Meta classified its AI data centers as large-scale experiments that may fail when filing with the IRS, categorizing them as pilot models to claim research and experiment tax credits.
This credit comes from federal legislation enacted in 1981, originally intended to subsidize materials consumed in experiments, not mature commercial operations. Starting at the end of 2024, Meta will separate accounting for chips sent to AI data centers from those sent to regular data centers. The former will be reported as experimental materials, primarily graphics processors procured from Nvidia.
The amount of the credit will increase from $700 million in 2023 to $2 billion in 2024, and then to $3.9 billion in 2025. According to securities filings, Meta has become the largest beneficiary of this credit among publicly traded companies. The company claims to investors that its AI investments are yielding results, while stating to tax authorities that the facilities are still experimental.
The company's own accounting has flagged this treatment as a legal risk. Securities documents warn that the IRS may seek to recover tens of billions of dollars, with the uncertainty of research tax credits being a major exposure. Provisions for uncertain tax positions have risen by 45% to $18.74 billion. The IRS has previously opposed counting commercialized, purchasable equipment and technology towards this credit.
The auditing firm Ernst & Young has approved this treatment and participated in its establishment, now promoting the same method to other companies purchasing AI chips. James Shannon, the author of the 1981 legislation, told The New York Times that the original intent of the law was for labor, knowledge, and information, and that Meta's usage has far exceeded anyone's imagination at the time. Meta defends that its R&D spending over the past five years has been about $200 billion.
In market mechanics, the seller is Nvidia's chips, the buyer is Meta's capital expenditures, and the subsidy comes from federal tax credits. This is not a transaction event-driven scenario, but rather accounting classification-driven. Cash flow shifts from tax liabilities to capital expenditures for data centers. Beneficiaries are platforms and chip suppliers that can classify graphics processors as experimental materials, while other taxpayers within the same tax base are pressured, and shareholders may have to pay additional taxes if pursued for recovery.
Source: Public Information
ABAB AI Insight
The research and experiment credit was established by Congress in 1981 for laboratory materials, not for data centers. Meta's previous credit amounts were in the hundreds of millions, with 2023 being $700 million. The turning point appears at the end of 2024 with the separation of accounting: chips for regular data centers continue to depreciate as assets, while chips for AI data centers are reported as experimental consumption. The same supplier and model have their tax identities split by destination.
The money spent hasn't decreased; it's just the timing of tax payments that has shifted. Capital expenditures in 2025 are about $72 billion, with the Louisiana Hyperion project alone targeting over $5 billion and aiming for 5 gigawatts. The $3.9 billion credit effectively subsidizes part of the federal tax bill directly into the data center budget. Ernst & Young both audits and designs this structure, then markets it to other buyers, indicating a path from one company to an industry template.
This is similar to pharmaceutical companies including phase III clinical trial materials in R&D credits, but drugs cannot be sold before approval. Meta's advertising and recommendation systems are already utilizing these clusters. It is more akin to energy companies writing commercial drilling as exploration expenses. The industry position is in capacity expansion, not in laboratory validation. Approximately 28 data centers are operational across the U.S., and Zuckerberg stated in the earnings call that the investment has already been recouped.
Structurally, this represents a transfer of pricing power. The economic cost of chips consists of the purchase price plus tax burden, and Meta uses the pilot model to shift the tax burden off its books. The reason this occurs is that the 1981 statute prices materials based on "whether it is experimental," not on "whether it has been used for commercial reasoning." Once classification power falls into the hands of taxpayers, commercial clusters can be written as experiments, and the federal tax base effectively pays part of the down payment for graphics processors.
ABAB News · Cognitive Law
- The same machine, just changing the subject is a subsidy.
- Claims production externally, claims experiment to the tax bureau.
- Whoever holds the classification power, the tax base subsidizes their capacity.