The Take — Meta's tax credit is a subsidy nobody voted for

The US is having its loudest argument yet about who pays for AI, and it is having it about the wrong bill. Congress voted 417 to 3 to make data centers pay for their own power. Meanwhile, on the other channel — the tax code — a single company moved $6.7 billion of its federal tax bill into a line item the public never debated. I think that is the more consequential number of the two, and the more fragile one.

Start with what the reporting establishes, because the facts are checkable and they check. The New York Times reported this week that Meta claimed a research tax credit on its AI data centers by classifying them as experimental "pilot models," on the argument that the Nvidia chips it buys belong to the experiment rather than to running a business. By the paper's count, the credit has climbed from $700 million in 2023 to $2 billion in 2024 to $3.9 billion in 2025. Meta's own fiscal 2025 filing lists current federal tax of $2.82 billion, against $9.57 billion a year earlier. We laid the mechanics out in this morning's coverage — Meta cut its federal tax bill by calling AI data centers experiments — and the part worth dwelling on is not the size of the swing but where it sits in Meta's own words.

The company's filing states that its unrecognized tax benefits are "predominantly accrued for uncertainties with our research tax credits." That is a company telling its shareholders, in a document it signs under penalty of securities law, that the position is exposed. Andre Shevchuck, a partner at BPM who works on research credits, told the Times the approach was "kind of wild and out there," which is roughly the professional equivalent of a raised eyebrow. So the honest description of the $6.7 billion is not a discount anyone granted Meta. It is a contested position that Meta is taking and reserving against, priced today by the company's own assessment of its odds.

Here is the mechanism, and it is the whole argument. The credit — a 1980s invention meant to push private money into genuinely uncertain development — is not self-executing. It is claimed on a return, defended only if the return is examined, and the examination rate for a company of this size is a very small number. When a position is disclosed to shareholders as uncertain and taken anyway, the expected outcome is not a coin flip. It is a claim that gets made every year and challenged rarely. Whatever the legal merits turn out to be, the operational subsidy arrives on schedule, and the legal reckoning may not arrive at all.

That is what makes this different from a rate fight. Data-center power costs are visible, local, and attributable — somebody's utility bill goes up, they call a legislator, and the House votes 417 to 3. The tax channel has no such feedback: no ratepayer to notify, no hearing, no line on a bill. And the visible channel is already correcting itself. A neocloud announced a 375-megawatt Oklahoma site this month with no incentives at all, paying full market value for the land, under a state ratepayer protection law rather than a federal one. Set that next to a federal credit measured in billions on compute, and the asymmetry stops being theoretical: the part of the public contribution that is visible is shrinking, and the part that is invisible scales with capex.

The reason to care about a single company's footnote is that the classification is portable. Nothing in the argument is specific to Meta. Every lab and hyperscaler buys GPUs for training runs whose outcome nobody can predict in advance, and "supplies used in a process of experimentation" is a description that fits a training cluster about as well as it fits a chemistry bench. If the largest capital buildout in the history of American corporate investment can be routed through the research credit, then the credit stops being an innovation policy and becomes the default financing channel for AI infrastructure — and the answer to "how much should the public contribute" gets set by filed positions rather than by any vote.

The strongest case for the other side

Give it a fair hearing, because the steelman is real and it is not "tax avoidance is fine."

The credit exists to subsidise exactly the kind of work frontier labs do, and its central test — a four-part inquiry into whether the activity is technological in nature, aimed at eliminating uncertainty, and conducted by a process of experimentation — is genuinely contested when the subject is a model-training run whose capability no one can predict. Congress wrote a credit broad enough to reach this work and never narrowed it; a company maximising a benefit the statute appears to offer is doing what the tax system asks of it, and the "pilot model" framing is aggressive but not invented. Meta reserves the exposure, which is the responsible version of taking a contested position. If the argument is that the IRS is outgunned, that is a question about appropriations — an inspector general finding or a court ruling settles it, and until then the deduction is lawful until someone says otherwise.

There is also a reading where my complaint is just the tax system working as designed. Business credits routinely front-load benefits and get sorted out years later; the reserve exists precisely so that shareholders are not misled. And Elizabeth Warren's letter questioning AI tax breaks is a member of the minority party writing to companies she does not regulate — a press release dressed as oversight, with no enforcement behind it.

Why the take still holds

Because the reserve is the tell. A company that expects a position to hold does not describe it to investors as a source of uncertainty. Meta's own language concedes the contingency, which means the $6.7 billion is not a settled reduction in the public's bill — it is a bet placed with the public's money, where the downside scenario is a later assessment and the base case is nothing happens. And the scale of the credit pool is what makes the bet load-bearing: if claims of this kind became standard across the industry, the question of what AI infrastructure really costs the treasury would be answered by the tax bar rather than by anyone elected.

The 417-to-3 vote is the proof the country is willing to have this argument. Unanimity on an AI infrastructure bill is not a close-run thing — it is a chamber that has noticed that ratepayers are paying and does not want to be the one explaining why. The same instinct applies to the credit, and no one has asked the question yet in a forum where the answer would be recorded.

What would change my mind

• A Treasury or court ruling that compute purchased for model training and serving qualifies for the credit as a matter of law. If the classification is correct, then this is Congress's subsidy, properly priced, and my objection reduces to disclosure — Meta should say how much of the credit comes from serving inference, which is unambiguously running a business.
• Disclosure of the composition. A credit attached to training runs that might fail is arguable. A credit attached to chips serving paying customers is not. If the claim is overwhelmingly the former, I will soften this considerably.
• A pattern, or its absence. If Microsoft, Google and Amazon do not take the same position, the story is one company's aggressive filing. If they do, and none of them has said so, then the subsidy is an industry default nobody voted for — and the fix is narrower than a rate fight and cheaper than a rate cut.
• A reserve that unwinds without an assessment. If the uncertainty resolves in Meta's favour on the merits, with reasons published, I will take the answer.

Until one of those arrives, the honest accounting is this: the public is paying for the AI buildout twice, once through the power bill it argued about and once through the tax code it did not — and only the first has a feedback loop.

Should the public's share of the AI buildout be set by a vote or by a filed position? Tell us in the comments.

Sources: New York Times · Meta fiscal 2025 annual report (SEC) · IRS — research credit · CNBC — Warren questions AI tax breaks