Meta's $6.6B research tax credits questioned amid massive AI data center buildout

Meta's rapidly expanding AI infrastructure programme is drawing congressional scrutiny over whether expenses associated with the design, construction and operation of highly specialised data centers should qualify for federal research tax credits.

The company reported approximately $6.6 billion in federal research credits between 2023 and 2025, increasing from around $700 million in 2023 to about $2 billion in 2024 and nearly $4 billion in 2025. It is not publicly known how much of that total, if any, relates specifically to data center development, which is now one of the main questions lawmakers are asking the company to clarify.

The issue centres on the increasingly blurred boundary between experimental engineering and conventional capital construction. Research credits are generally intended for activities involving technological uncertainty and a process of experimentation. AI data centers, however, are becoming significantly more complex than conventional facilities as rapidly increasing computing densities drive new approaches to cooling, electrical distribution, power generation and mechanical integration.

Meta argues that some of its AI-optimised infrastructure involves first-of-kind engineering that requires extensive real-world testing and cannot always be resolved through simulation alone. Lawmakers are questioning where that experimentation ends and the construction of permanent buildings, electrical systems, cooling infrastructure and other long-life assets begins.

The scale of Meta's current construction programme makes that distinction particularly significant. Its Hyperion AI campus in Richland Parish, Louisiana, has developed into an approximately $27 billion project incorporating around 4 million sq ft of buildings together with extensive power, cooling and connectivity infrastructure. The campus could eventually require as much as 5 GW of electricity and is expected to support more than 5,000 workers during peak construction.

Meta is also developing a more than $10 billion, 1 GW campus in Lebanon, Indiana, while another 1 GW development in El Paso is associated with approximately $14 billion of investment in buildings and supporting infrastructure. Across the wider programme, Meta expects AI infrastructure expenditure in 2026 to reach as much as $145 billion.

The tax question therefore has implications well beyond individual data center buildings. Modern AI campuses increasingly resemble integrated infrastructure systems, combining computing halls with substations, high-capacity electrical networks, advanced cooling systems, backup generation, telecommunications infrastructure and, in some cases, dedicated energy facilities.

Congressional inquiries are now seeking more detailed information on which AI data center expenditures Meta considers eligible, how those costs were classified and whether professional legal or accounting opinions supported that interpretation. The company has been given until October 31 to provide additional information.

No determination has been made that Meta's claims were improper, and the amount of research credits associated specifically with data center construction remains unknown. The scrutiny nevertheless raises a broader question likely to become increasingly relevant as AI infrastructure evolves: when a facility incorporates technologies and systems that have never previously been built at that scale, how should experimental engineering be separated from the permanent infrastructure required to make it work?