Meta’s AI construction program has acquired another competitive dimension: how much of its infrastructure can count as research for federal tax purposes. A September 30 investigation reported that the company claimed about $2 billion in benefits for 2024 and $3.9 billion for 2025 by treating certain new AI data centers, specialized chips and related equipment as experimental “pilot models.”
Nearly $6 billion is consequential, but it is neither a new government grant nor a settled entitlement. The reported treatment applies an existing research credit to a disputed category of costs. The public record described in the reporting contains no final IRS rejection or judicial finding that Meta committed illegal tax evasion.
A campus can support research and commercial production
A pilot model is a representation used to evaluate a new or improved product or process. Meta’s reported interpretation treats certain AI campuses as integrated experimental systems: chips, networking, cooling and software operate together while engineers investigate new models and infrastructure methods. It distinguishes them from conventional facilities serving established storage, advertising and social-network operations.
Meta spokesperson Andy Stone defended the company as a major American research investor using incentives Congress created for domestic investment. That defense does not settle which expenses qualify. A research-heavy business can purchase advanced equipment without every cost automatically becoming a qualified research expense.
The difficult case is mixed use. A cluster may train an uncertain new model today and serve a commercial recommendation system tomorrow. It can remain valuable long after one experiment ends. In a shared computing environment, workloads move between machines rather than assigning a permanent research purpose to each chip.
Critics argue that Meta needs this capacity to compete and would build much of it without a credit. That is a policy argument about whether the incentive creates additional investment, not automatic statutory disqualification. Company wealth and commercial necessity cannot replace examining the actual activities and expenses.
The evidence has to follow the experiment
The IRS research-credit instructions describe four requirements: technological research, uncertainty, improvement of a business component and a process of experimentation. Developing AI architectures, training approaches or hardware configurations can involve real uncertainty. Simply attaching the word experimental to a functioning campus is not the same evidence.
Pilot-model claims need a connection to relevant business components and supporting documentation. The practical questions concern what uncertainty engineers investigated, which experiments addressed it and how claimed costs relate to that work. Production use on the same site makes allocation important rather than making all research impossible.
A narrowly adverse outcome could concern insufficient records or poor cost allocation while leaving some pilot-model theory intact. A broader challenge to the classification would reach further. The IRS could accept portions, reject others or settle; a dispute could also lead to litigation. None of these possibilities establishes an already issued decision.
The distinction matters beyond semantics. If physically similar campuses receive different treatment mainly because companies label them differently, accounting choices could influence competition without corresponding differences in experimentation. Administrable rules would need to distinguish documented experimental use from ordinary commercial capacity.
A tax reserve is not an IRS verdict
Meta’s 2025 annual tax disclosure reported $16.45 billion in gross unrecognized tax benefits at year-end and $11.23 billion in net uncertain tax positions within long-term income taxes. The filing linked major uncertainties to research credits and transfer pricing involving foreign subsidiaries. Neither balance can be assigned entirely to AI data centers.
These measures also are not interchangeable with the reported $5.9 billion of two-year benefits. They describe broader accounting exposures. Changes in a future reserve could reflect settlements, reassessment, recognition of previously uncertain benefits or other disputes—not necessarily a result in this particular case.
The filing says a benefit is recognized only when the tax position is more likely than not to survive examination. That is management’s accounting assessment, not an IRS approval. Conversely, identifying uncertainty is not an admission of misconduct. Meta can defend an interpretation while making provisions for an adverse outcome.
Ernst & Young’s audit of the financial statements addresses accounting and reserves; it does not approve tax eligibility on the agency’s behalf. Accounting professor and former EY adviser Lisa De Simone highlighted this tension in coverage. Reports that similar strategies might be relevant to other clients suggest potential wider adoption, not proof that every competitor already claims them.
Federal credits are only one layer of support
State and local incentives operate under different objectives and rules. The state tax landscape includes different treatment of equipment purchases, electricity, buildings and tangible personal property. Research credits target innovation; local incentives generally seek investment, construction, jobs or regional development. Combining them requires tracing distinct benefits rather than counting one program as another.
Meta’s Richland Parish campus in Louisiana illustrates the overlap. An AP review of infrastructure costs described confidentiality limiting visibility into its Entergy agreement, three planned gas plants totaling 2,262 megawatts and Meta funding roughly half their construction costs over 15 years. Customers would still help pay for a transmission line serving the site.
That reporting raises questions about public visibility, long-lived obligations and who bears costs if demand changes. It does not decide whether Meta’s federal research claim meets the legal tests. Sales-tax treatment, utility arrangements and a federal credit are separate mechanisms even when they support one campus.
The Institute on Taxation and Economic Policy’s criticism focuses on subsidizing investment Meta already intended to make. A contrasting principle is neutral treatment of genuine qualified research regardless of the taxpayer’s size or profitability. The factual dividing line remains use, documentation and allocation—not simply whether a company is popular or wealthy.
The outcome could change how AI capacity is financed
If Meta retains the benefits, more conservative competitors could face a tax disadvantage, encouraging Amazon, Microsoft, Google, Oracle or others to examine comparable approaches. That is a possible response, not evidence of industry-wide adoption. A clear agency challenge, disclosed settlement or public court record would be stronger evidence of the actual boundary than another accusation.
Ownership adds a structural issue. Vertically integrated companies own products, models and campuses, while smaller laboratories often rent computing. That does not mean rented capacity cannot qualify: treatment depends on contract terms, expense categories and documentation. Comparing the routes requires more than contrasting a campus owner with a cloud customer.
Congress could clarify how shared AI infrastructure, depreciable equipment and pilot models interact. Rules that are too restrictive could penalize genuine hardware experiments; vague rules could subsidize ordinary production investment under a perpetual research label. Rapidly changing models make a workable allocation method more useful than assuming everything on an AI site has one purpose.
The next evidence is specific: which costs any IRS action challenges, how Meta explains subsequent reserve changes and whether competitors disclose similar treatment. A recordkeeping dispute, a classification ruling and a finding of fraud would be materially different outcomes. For now, the reported nearly $6 billion is an aggressive research-credit claim with meaningful uncertainty—not money unquestionably retained and not proven unlawful evasion.