Meta's $6.6 Billion R&D Tax Credit: The Pilot Model Problem

The short version

It is also the largest test yet of the data center R&D tax credit.

‍

At Strike, we run R&D credit studies for a living, and we defend them in audit. Our view: Meta's data centers almost certainly contain real qualified research. Meta's claim, as reported, is not that. Treating commercially available, depreciable production hardware as a consumable research supply because the building it sits in is labeled an experiment inverts the statute, ignores the supply rule's express exclusion for depreciable property, and runs straight into the cases the IRS has already won on this exact theory. If the facts are what the reporting says they are, the position depends on the IRS not looking closely, a risk Meta's own 10-K acknowledges by tying its reserve to whether its documentation will hold up, and the IRS is now looking.

The scale makes it everyone's problem. By the senators' arithmetic against the Joint Committee on Taxation's $32 billion estimate, Meta alone would account for more than 10% of all federal research credits claimed in 2025. When one taxpayer stretches a credit that far, Congress does not write a rule for that taxpayer. It tightens the rule for everyone, including the machine shop, the software startup and the family farm the credit was built for in 1981.

Below is what the public record actually shows, what it conspicuously does not, how the four-part test treats a GPU, how big the exposure is, who signed off, and why it matters to companies that will never spend $145 billion on anything.

What the record shows, and what it does not

Everything public about Meta's data center credit comes from three places: Meta's SEC filings, the New York Times investigation of September 30, 2026, and the congressional letters that followed it. Six facts, each with its source.

• Meta 10-K, FY2025. Research credits of about $700 million (2023), $2.0 billion (2024) and $3.912 billion (2025), worth 4.6 points off the 2025 effective rate. Year-end unrecognized tax benefits of $16.45 billion gross, of which $11.25 billion would affect the effective tax rate if recognized, "primarily" research credits and transfer pricing, with the reserve tied to whether Meta's documentation "corroborating the nature of our activities" will be sufficient.
• Meta 10-Q, June 30, 2026. Gross unrecognized tax benefits rose to $18.74 billion.
• New York Times, Sept. 30, 2026. Four people familiar with Meta's operations say that from late 2024 Meta tagged chips bound for AI data centers, classified those facilities as pilot models, and treated the chips as research supplies. Finance staff questioned whether it would survive IRS review, outside tax attorneys were consulted, and Meta's auditor advised on the credit and has pitched the approach to other companies. Meta declined to say what made the data centers experimental.
• Senate Finance Democrats, Oct. 13, 2026. Ten questions due October 31: whether Meta claimed the data centers as pilot models, how much of each year's credit was chips and computing equipment, and whether its auditor or any other firm opined on eligibility, with copies of any opinions.
• House bipartisan letter to Treasury, Oct. 7, 2026. Twenty-eight members ask Treasury to confirm that ordinary data center construction and deployment costs do not qualify, and that scale or configuration alone does not make a facility experimental.
• Meta statement (Andy Stone). AI-optimized data centers are "inherently new" and require testing "that cannot be simulated." Meta cites $200 billion of R&D spend over five years, $57 billion in 2025.

Now the more important list. Nothing in the public record shows any of the following, and the Senate letter is drafted precisely to force these answers:

1. How much of the $6.6 billion is data center hardware. Meta discloses one credit number. It does not break out chips, servers, networking, cooling, construction or power from the wages and contract research it has always claimed.
2. Any delineation between research and production use. The only separation reported is a procurement tag: chip goes to an AI site versus a conventional site. That is a destination label, not an activity analysis. It says nothing about whether a given GPU ran a controlled experiment or served Llama inference to two billion users on day one.
3. The business components. Section 41 requires research to be aimed at a specific product, process, technique or software. Is the business component the building, the rack, the cooling loop, the power architecture, the cluster topology, or the model being trained? Each answer produces a different, and much smaller, set of eligible costs.
4. The uncertainty and the alternatives evaluated. A process of experimentation means hypotheses, alternatives, and testing that could fail. Meta's public defense is "testing and experimentation that cannot be simulated." That describes commissioning. Every data center on earth is commissioned.
5. How a depreciated asset became a supply. Section 41(b)(2)(C) excludes property of a character subject to depreciation from supplies by its terms. Meta's financial statements assign its servers a useful life of roughly 5.5 years. A chip that serves production for five years is of a depreciable character whatever the return calls it. It is not a consumed supply.
6. Who prepared the study and on what standard. Meta's own reserve language concedes the documentation may be insufficient. That is not the language of a taxpayer with a contemporaneous, project-level substantiation file.

Absence of disclosure is not proof of a bad claim. But a company confident in a $3.9 billion position does not usually describe it to investors as a question of whether its paperwork will hold up.

What is a pilot model under the R&D tax credit?

The research credit has three cost buckets: wages for people doing, supervising or supporting qualified research; contract research; and supplies. Meta's wage claim is old news and, at $57 billion of annual R&D expense, enormous on its own. The new money is in supplies, and the supply rule is where this claim lives or dies.

The statute. Section 41(b)(2)(C) defines supplies as tangible property other than (i) land and improvements to land, and (ii) "property of a character subject to the allowance for depreciation." The IRS audit technique guide says the same thing in plainer words: supplies must be non-depreciable tangible property used directly in qualified research. A server rack with a 5.5-year useful life is the textbook example of what the exclusion covers. Congress wrote it that way on purpose. The credit was meant to subsidize what gets used up in an experiment, not the capital assets a business would buy anyway.

Meta will say that book and tax treatment routinely differ, and they do. To claim a chip as a supply, Meta had to expense it under Section 174A on the return rather than depreciate it for tax. That is not the problem. The problem is that Section 41(b)(2)(C)(ii) does not ask how the taxpayer treated the property; it asks whether the property is of a character subject to the allowance for depreciation, tested in the taxpayer's hands. In TG Missouri, the molds qualified because the taxpayer held them for sale to the customer, so they were never of a depreciable character in its hands. A GPU Meta installs and runs for five years is of a depreciable character however it is booked. Treasury Regulation 1.174-2(b)(1) makes the same point from the other direction: Section 174 does not reach expenditures for the acquisition of depreciable property, only the depreciation allowance on such property to the extent it is used in research.

So the test is not what Meta entered on its books or its return. It is what the asset is. Multi-year production hardware does not become a consumed supply because it was the first batch installed in a new building.

The exception. Treasury Regulation 1.174-2(a)(4) defines a pilot model as any representation or model of a product that is produced to evaluate and resolve uncertainty concerning the product during its development or improvement. The 2014 regulations confirmed that costs of producing a pilot model can be research expenditures even if the model is later sold or placed in service. The classic case is TG Missouri (2009), where production molds built to resolve design uncertainty, then sold to the customer, counted as supplies because the taxpayer never depreciated them. Trinity Industries (2010) allowed first-in-class ship hulls on similar reasoning.

The limits. Three cases draw the fence Meta has to stay inside:

• Union Carbide (2d Cir. 2012). When a company runs experiments inside its production process, only the incremental supply cost attributable to the research qualifies. Raw materials it would have consumed anyway making product for sale are not research supplies, even if an experiment was running on the line. A GPU that trains or serves Meta's commercial models is the production line.
• Betz (Tax Court 2023). A manufacturer called every custom oxidizer it built a pilot model because each was uniquely designed. The court said no: the taxpayer must show that its purpose in building the unit was to evaluate and resolve uncertainty, and the absence of early-stage testing showed the units were final products, not models. The court also made clear that testimony does not substitute for contemporaneous documentation of how the model was used to resolve uncertainty.
• Little Sandy Coal (Tax Court 2021, aff'd 7th Cir. 2023). A shipbuilder claimed first-in-class vessels as pilot models. The Tax Court and the Seventh Circuit (62 F.4th 287) did not hold that a first-in-class unit can never be a pilot model. They held that the label does not carry the whole unit through the substantially-all test: the taxpayer must prove experimentation activity by activity, and it could not.

Put those together and the doctrine is narrow. A pilot model is a representation built to answer a question. Once the question is answered, the next unit is production. A facility that is simultaneously a "pilot" for tax purposes and the backbone of the company's revenue for investor purposes is asking the Tax Court to accept that the question is never answered, at every site, on every chip, year after year.

Do data centers qualify for the R&D tax credit? A GPU against the four-part test

Qualified research under Section 41(d) must pass four tests at the business component level. Here is how a rack of commercially purchased GPUs inside a Meta AI data center fares on each, judged on what is public.

‍

The honest version of this analysis is not "zero." It is "shrink back." Regulation 1.41-4(b)(2) lets a taxpayer apply the tests to the most significant subset of a product that satisfies them. A new liquid-cooling loop, a novel 800-volt DC power distribution, a first-of-its-kind cluster interconnect, a custom accelerator Meta designed itself, a tented rapid-deployment structure, a rack configuration that had never been run at scale: each of those can be a legitimate business component with wages, contract research and, in some cases, incremental supplies attached. Meta's engineers have published on several of them.

Meta's 10-K says the company has designed and built its own data centers and key portions of its technical infrastructure, and the public record agrees: the next-generation design that led Meta to tear down and restart Hyperion in 2022, and the 140 kW liquid-cooled rack it contributed to the Open Compute Project in 2024, were produced by Meta's in-house engineering teams. That is the research. It is paid for in salaries, and it is claimable as wages. It does not change what a purchased accelerator is.

What shrink-back cannot do is turn tens of billions of dollars of off-the-shelf accelerators into supplies because they were the first batch installed in a new building. The GPU is not the thing being designed. It is a purchased input that functions exactly as its manufacturer specified. The uncertainty, if any, lives in the system around it, and the credit follows the uncertainty, not the invoice total.

Data center R&D tax credits: where research ends and production begins

Every credible research credit study for a company that builds and operates its own infrastructure answers one question first: where does the experiment end and the operation begin? Meta has not answered it publicly, and the shape of its disclosures suggests it may not have answered it the way an examiner will require.

What a defensible delineation looks like. A taxpayer claiming hardware inside a pilot facility should be able to produce, for each claimed site or cluster: the design hypothesis being tested (for example, whether a given liquid-cooling density holds thermal targets at full load); the alternatives evaluated and rejected; the test protocol and the dates it ran; the results and the design changes they drove; the date the facility or cluster was released to production; and a cost cutoff at that date. Costs before release are candidates for the credit. Costs after release are operations. Chips consumed in destructive or non-recoverable testing are supplies. Chips that go on to serve customers are capital assets and are not, whatever the building was called.

What the reporting describes instead. A tag applied at procurement that routes a chip to an AI site rather than a conventional site, and a facility-level label applied after the fact. Neither of those is a cutoff. Neither identifies a hypothesis. Neither distinguishes the first 1,000 GPUs used to validate a cluster design from the next 100,000 installed because the design worked.

The arithmetic that makes the point. Meta's regular credit rate is 20% of qualified expenses over a base amount; its alternative simplified credit is 14% of expenses over half the prior three-year average. Taking the increase from the 2023 credit to the 2025 credit, roughly $3.2 billion, and dividing by those marginal rates implies somewhere between $16 billion and $23 billion of incremental qualified expenses in 2025. Meta's 2025 capital expenditures were $72.2 billion. Wages did not grow by $16 billion in two years. The only line item large enough is hardware. Either Meta treated a very large share of its AI hardware purchases as research supplies, or the public numbers do not add up.

Why the deduction side matters too. Under the One Big Beautiful Bill Act, domestic research costs are fully deductible under new Section 174A beginning in 2025, and bonus depreciation is back at 100%. In 2025 the deduction outcome is the same either way, so classifying a chip as a research supply costs Meta nothing on the deduction and earns a credit worth roughly 14 cents on the dollar before the Section 280C reduction. In 2024, when Section 174 still required five-year amortization, calling a chip a research cost would have been worse than depreciating it. That the strategy reportedly began in late 2024 and nearly doubled in 2025 is consistent with a tax function optimizing a rule change, not with a laboratory discovering it had been doing experiments all along.

None of this requires Meta to be wrong about the engineering. Its AI campuses are at the frontier of power density, cooling and networking. The point is that frontier engineering is a wage story and a design story. It is not a story in which the most expensive commercially available component in the world stops being a capital asset.

How much of Meta's $6.6 billion is at risk?

Larger than the headline, and growing faster than the headline.

‍

Four things make this worse for Meta than a normal large-dollar dispute.

1. The statute is open on every year that matters. The 2024 return was filed in late 2025 and stays open until late 2028. The 2025 return was due this month. Meta is under continuous IRS examination; there is no waiting out the clock.
2. The reserve is explicitly about documentation. Meta's 10-K ties its uncertain position to whether its records corroborating the nature of its activities will be sufficient. A legal-theory dispute can be argued by lawyers. A documentation dispute is decided by what exists in the file on the day the information document request arrives. Betz was lost on exactly that.
3. It is a repeating position. Every quarter Meta books more hardware under the same theory, the exposure compounds. If the IRS or Treasury rejects the theory in 2027, the question is not $4 billion. It is $4 billion plus everything claimed in the interim.
4. The penalty defense may be thinner than usual. Reasonable cause under Section 6664 ordinarily rests on reliance on a competent advisor's opinion. Meta has not said it has one, the Senate has asked for copies, and the Times reported that the advisor was also the auditor. Reliance on an advisor with a stake in the outcome is a weaker defense than reliance on an independent one.

For scale: a $5 billion adjustment with penalty and three years of interest approaches $7 billion. Meta earned roughly $60 billion in 2025 and can pay it. The risk is not solvency. It is that a position carried in the effective tax rate for three years reverses in a single quarter, and that the correction lands at the same time Treasury writes guidance that forecloses the strategy for the years ahead.

Did Meta's auditor sign off on the R&D credit?

Three different parties had to say yes to this credit, and the record on each is uncomfortable.

The auditor. Ernst & Young has audited Meta since 2007, per the 2026 proxy. According to the Times, EY also advised Meta on the credit and has since pitched the chip-as-supply approach to other AI companies. Meta's 2026 proxy shows EY billed $8.4 million in tax fees for 2025, of which only $1.8 million was tax advisory (2024: $1.6 million; 2023: $0.8 million). Read that again. According to the Times, the firm that audits Meta's reserve for uncertain tax positions, which covers this position among others, also advised on the position, and per the 2026 proxy its tax advisory fee would not cover the chips in one rack.

That combination raises a question the Senate letter asks directly and a question shareholders can fairly ask. PCAOB Rule 3522 bars an auditor from marketing, planning or opining in favor of a transaction it recommended if a significant purpose is tax avoidance, unless the treatment is at least more likely than not to be allowed, and Rule 3524 requires the audit committee to pre-approve tax services after the auditor describes their scope and fee structure. Meta's reserve does not answer the Rule 3522 question either way: under ASC 740, a company can expect a position to be sustained and still reserve part of the benefit because of how much of it it expects to keep. The public record does not answer the obvious questions: was the credit work booked as compliance, did another firm do the study, and what did the audit committee review when it pre-approved EY's tax services? Meta has not said. Questions 4 and 5 of the Senate letter, which ask whether EY or any other firm opined and demand copies, are aimed at exactly this gap.

The audit committee. Meta's audit committee, made up of independent directors, pre-approved every EY service in 2024 and 2025, according to the proxy. Pre-approval is routine. If, as the Times reported, the auditor advised on a tax position it then audits, and that management itself flags in the 10-K as potentially undocumented, pre-approving that work is a governance judgment shareholders can fairly question.

The company's own finance staff. The Times reported that some of Meta's finance personnel questioned whether the treatment would pass IRS review, and that Meta consulted outside tax attorneys. The 10-K then disclosed the position as uncertain. In other words, the people closest to the numbers expressed doubt, the company went ahead, and the doubt was written into the financial statements. Read together, the Times' account of finance staff raising doubts and the 10-K's reserve language read like a company that weighed the audit risk on $6.6 billion and decided to proceed.

Who actually prepared the study. Unknown. Meta has never named its research credit preparer in any filing, and the proxy fee table cannot be reconciled to a study of this size. The possibilities are that EY's credit work sits inside the $6.6 million compliance line, that a second firm or a law firm built the claim, or that it was built in-house with EY sign-off. Each has a different implication for penalty protection and privilege, and Meta's October 31 response is the first time the public may learn which.

The Zuckerberg precedent

This is not Meta's first maximal reading of Section 41, and the earlier one tells you how the company litigates.

For 2012 and 2013, Facebook claimed research credits of $618 million. About $354.7 million of that rested on treating $4.1 billion of stock option income Mark Zuckerberg recognized on exercise as wages for qualified research. The IRS disallowed it. The dispute has been in Tax Court for years; both sides filed cross-motions for partial summary judgment in May 2025 and the court has not ruled.

The legal question is narrow and revealing. Qualified wages must be paid for qualified services: engaging in, directly supervising or directly supporting research. Meta argues the court should look at what Zuckerberg was doing in 2005, when the options were granted, because he was writing code then. The IRS points to the option agreement itself, which says the grant was not consideration for services rendered before the vesting commencement date, and argues the relevant period is 2008 to 2010, when he was running a company that grew from several hundred employees to more than two thousand.

Whether Meta wins that motion is almost beside the point. What the case shows is a tax department willing to classify a chief executive's option payout as a research wage, carry the position for more than a decade, and litigate it to summary judgment. That is the same department now classifying GPUs as consumable supplies. The Senate letter pairs the two deliberately, and it is right to: a pattern of positions at the outer edge of the statute is itself evidence an examiner weighs when deciding how hard to push on the next one.

R&D tax credit audit risk: what Meta's claim means for every other claimant

The research credit survived forty-five years of budget fights because it is small-business policy that happens to be used by large companies. The 1981 Congress worried about Japan and about domestic manufacturers underinvesting in engineering. The 2015 Congress made it permanent and added the payroll tax offset so that startups with no income tax could use it. The 2025 Congress restored full expensing under Section 174A. At every step, the political case was the machine shop, the software firm, the medical device startup, the ag producer running field trials.

Meta's position threatens that case in three ways.

It invites a legislative fix that will not be surgical. When one taxpayer claims more than 10% of a $32 billion incentive on a theory that converts capital expenditure into research, the response is not a Meta-only rule. The House letter already asks Treasury to confirm that "scale or configuration" does not make a facility experimental. The next step is a statutory carve-out for data centers, then for computing equipment generally, then a tightening of the supply definition that sweeps in the test materials a contract manufacturer actually consumes. Every tightening lands hardest on the companies without a tax department to argue the edge cases.

It changes examiner behavior now. IRS Large Business and International runs the research credit as a campaign issue. A multi-billion-dollar pilot model claim at the largest claimant in the country will produce internal guidance, training and templates. Those templates get used on the $400,000 claim at a regional manufacturer. Expect more information document requests on supplies, more insistence on a documented production cutoff, and more skepticism of any claim where a capital asset shows up on Form 6765.

It makes documentation the whole game. Meta's reserve language is the tell. The company did not say its theory was uncertain. It said its documentation might not corroborate its activities. That is the failure mode we see in audit more than any other: a plausible theory, an aggressive number, and a file that cannot connect the two at the project level. Betz lost $500,000 of credits with a file that rested on testimony. Meta is risking $5 billion on what the Times describes as a procurement tag.

Here is the standard we apply, and the one we think every claimant should expect to meet from now on:

1. Identify each business component before counting a dollar. Name the product, process or design being improved. A building is rarely the right answer; the system inside it usually is.
2. State the uncertainty at the start, in writing, with the alternatives considered. If you cannot say what might have failed, it was not an experiment.
3. Fix the date the component was released to production. Costs after that date are operations.
4. Claim supplies only for property that is not depreciated and that was consumed in the experiment. If it was depreciated for tax, or went into production use, it is not a supply. Under Union Carbide, if you would have bought it anyway, only the incremental cost qualifies.
5. Build the file from contemporaneous records, dated when created, with sources cited. Nothing backdated, nothing reconstructed from memory, nothing that depends on an engineer's recollection in a deposition five years later.

That is not a conservative posture. It is the posture that lets a client take every dollar the statute allows and keep it. The alternative is what Meta is now defending in front of the Senate.

‍

‍

A $6 billion credit and a $100 billion question

What follows is my read of the valuation math, not a forecast. Time and events will tell.

The instinct on Wall Street will be to treat this as a tax footnote: $6.6 billion of credits against a company that generates more than $100 billion of operating cash flow, with most of the dispute already parked in a reserve. That instinct is right about the past and wrong about the future. The back years are a cash event. The forward years are a valuation event, and the gap between the two is roughly a factor of fifteen.

The back years are small. Under the accounting rules, Meta only recognized the portion of the credit it judged more likely than not to survive. The rest sits inside $18.74 billion of gross unrecognized tax benefits that already accrue interest and penalties. If the IRS disallows the 2024 and 2025 data center claims, the income statement absorbs only the unreserved slice, and it absorbs it once. Markets look through one-time tax charges. Meta took a $15.9 billion charge in the third quarter of 2025 and the stock's reaction that day had nothing to do with it.

The forward years are not. The credit was worth 4.6 points of Meta's 2025 effective tax rate. If the data center portion is three to four points today and five to seven points on $130 billion to $145 billion of 2026 capital spending, foreclosing it permanently cuts recurring net income by 4% to 8%. At Meta's current earnings multiple, a 4% to 8% haircut to recurring earnings is $75 billion to $150 billion of market value at a constant multiple. Two things soften the arithmetic. Meta expects to pay the corporate alternative minimum tax from 2025, which already defers some of the credit's cash value into carryforwards the market may not fully count. And losing the credit restores the portion of the Section 174A deduction that Section 280C had reduced, clawing back about a fifth of the loss unless Meta already elected the reduced credit. Net of both, four to six points of recurring earnings is the realistic exposure.

The multiple is where it could get expensive. This credit quietly subsidized roughly five and a half cents of every dollar Meta spent on property and equipment in 2025. Investors have shown exactly one consistent reaction to Meta's AI buildout: they punish capital spending that outruns the visible return. Remove the subsidy at the same moment the company guides to another record capex year, and the question stops being "what is the tax rate" and becomes "what is the real cost of this buildout." That question compresses the multiple. A 6% earnings cut can become a 10% to 12% repricing when it arrives wrapped in a capex guide, and 10% of Meta is roughly $190 billion.

Three paths, in order of likelihood as I read it. First, the IRS examines, Meta settles the back years for a fraction of the claim and keeps a narrowed version of the position going forward, limited to genuinely experimental subcomponents. This is the modal outcome and it costs a few points of earnings, not a tenth of the company. Second, Treasury issues guidance that computing infrastructure deployed for commercial workloads is not a pilot model, which ends the strategy for every hyperscaler at once and triggers the forward repricing above. Third, nothing happens for years, the position is litigated in Tax Court in 2029, and the exposure compounds with every quarter of chips booked under the same theory. The third path is the one investors should least want and the one Meta's disclosures suggest it is prepared to take.

What would change my mind: a Senate response that describes project-level test protocols and production cutoffs, or a 10-Q that quantifies the data center portion and shows it is a minority of the credit. Either would make this a documentation dispute over a modest number, which is what Meta's communications team would like everyone to believe it already is. The absence of either would tell its own story.

What to watch

‍

If Meta's October 31 response describes a project-level file with hypotheses, test protocols, production cutoffs and a supply population limited to consumed property, this article will have been too harsh and we will say so. If it describes a facility-level label and a procurement tag, the IRS has its road map and so does everyone else who claims this credit.

Official Sources

Disclosure: Jonathan Cardella is the founder and Chief Executive Officer of Strike Tax Advisory, which has recovered more than $400 million in R&D tax credits for more than 1,600 companies and defends every study it delivers. This article is written in that capacity and reflects the author's opinion based on public filings, court records and published reporting as of October 20, 2026; facts attributed to news reports have not been independently verified. It is not tax advice or legal advice for any taxpayer, and it is not investment advice, a recommendation to buy or sell any security, or a prediction of Meta's stock price or tax outcome. The author does not hold a position in Meta securities. Strike Tax Advisory has no engagement with Meta Platforms, Inc. Readers should consult their own advisors before acting on anything here.

Work With Strike Tax Advisory

Strike Tax Advisory is an R&D tax credit specialist firm helping innovative businesses across manufacturing, agriculture, engineering, software, and other technical industries claim federal and state research credits.

For a company that builds, tests or operates its own technical infrastructure, our team will:

• Identify each business component and the uncertainty behind it before counting a dollar
• Fix the date each component was released to production and separate research costs from operations
• Limit supplies to non-depreciable property consumed in the research
• Build the study from contemporaneous, project-level records and prepare the Form 6765 support with your CPA
• Stand behind the claim with examination defense