Senators Warren, Smith and Merkley have sent information requests to four major tech companies about statutory tax deductions tied to AI infrastructure — and the numbers that don’t yet exist publicly are the whole point.
These are letters requesting information — not findings, not charges, and not evidence of wrongdoing. The deductions at issue are provided by statute and claiming them is lawful. The company-level figures below are the senators’ office’s own estimates, not company disclosures. The Joint Committee on Taxation’s $67 billion is an economy-wide figure for all companies, not an AI or data-centre figure. Meta also received a letter, but no estimate for Meta is published and none is inferred here.Nothing here is tax advice or investment advice.
Senators Elizabeth Warren, Tina Smith and Jeff Merkley sent letters to the chief executives of Meta, Alphabet, Amazon and Microsoft requesting information about tax deductions claimed under provisions of the 2025 legislation known as the One Big Beautiful Bill Act — specifically permanent 100% bonus depreciation and retroactive research-and-development expensing. Responses are due October 12. These are information requests. They are not findings, not charges, and not evidence of wrongdoing. Claiming a deduction that a statute explicitly permits is lawful.
The senators’ office published its own arithmetic alongside the letters: estimates of roughly $17.9 billion for Alphabet, roughly $15.7 billion for Amazon and roughly $12.5 billion for Microsoft. These are the senators’ office’s own calculations — not company disclosures, not findings by any tax authority, and not figures from the Joint Committee on Taxation. No methodology for those estimates has been published. No company’s actual claimed deduction has been established anywhere, and no company has responded in the reporting. The absence of a public figure is a limit of available reporting, not evidence that any specific amount does or does not exist.
The Joint Committee on Taxation’s separate $67 billion figure is economy-wide — covering all companies, for retroactive breaks, in 2026. It is not an AI figure, not a data-centre figure, and not a figure for these four companies. The two sets of numbers measure different things across different scopes and are not components of one another. Nothing in this article is tax advice or investment advice, and no position is taken here on whether the underlying tax provisions are sound policy.
The key insight: The reason these letters exist at all is structural: accelerated depreciation produces no public entry. A grant has a named recipient and a searchable register. A smaller tax bill has neither — which means the only way to learn a company-level number is to ask the company directly, because no filing requires it to be broken out and no agency publishes it.
No public register carries these numbers, which is why a Senate office is the one publishing them.
The Structural Read
The fiscal architecture of bonus depreciation is built differently from a grant or a subsidy line item. When a government awards a grant, it creates a record: a named recipient, a dollar amount, a date, a register that a journalist or a researcher can search. Accelerated depreciation creates none of that. It is simply a reduction in taxable income — and a smaller tax bill is not an entry in any public list.
That invisibility is not concealment. Nobody is hiding anything here. Companies are under no legal obligation to report their depreciation by asset class or by business function in public filings. The tax code simply does not generate the field an observer would want to see. The senators’ letters are, in a precise sense, an attempt to retrofit a disclosure requirement onto an instrument that was never designed to produce one.
The result — a Senate office publishing its own arithmetic in place of a disclosure that does not exist — is the reporting problem made visible. When the only authoritative figure is economy-wide and the only company-level figures are estimates produced by an interested party, there is no number anyone can cite without attaching a qualifier. That condition is exactly what the letters are attempting to end, and it is a transparency problem before it is a political one.
Permission Layer — Business Engineer Framework
“The Permission Layer is not only regulation that blocks deployment — it is also the absence of a reporting structure that would make deployment legible. When the instrument of subsidy leaves no public trace, the oversight apparatus has to improvise. These letters are improvised oversight.”
This is where the fiscal side of the AI infrastructure buildout becomes structurally opaque. The gross capital expenditure flowing into data centres is partly a matter of public record — companies report capex in aggregate. The portion of that capex offset against tax through bonus depreciation is not public. One side of the ledger is visible; the other is not. No depreciation rate is applied to any capital figure here, and no estimate of how much of any buildout was deducted is offered — that number does not exist in any public source.
Three Implications
IMPLICATION 1 — The Disclosure Gap Is the Story
Accelerated depreciation is structurally different from named subsidies precisely because it produces no public record. The information requests reveal that the only path to a company-level figure currently runs through the company itself — not through any public filing, agency database or regulatory disclosure. That is a reporting infrastructure problem that predates this particular political moment.
IMPLICATION 2 — Estimates Without Methodology Are a Weak Foundation
The senators’ office’s company-level estimates — roughly $17.9 billion for Alphabet, roughly $15.7 billion for Amazon, roughly $12.5 billion for Microsoft — carry no published methodology. Until companies respond, or until a disclosure standard exists, these figures are the best available proxies and simultaneously the least verifiable ones. Any analysis built on them inherits that uncertainty in full.
IMPLICATION 3 — The Permission Layer Operates on Both Sides of the Ledger
Regulatory frameworks that govern AI tend to be understood as deployment controls — what can ship, where, under what conditions. This episode illustrates that the Permission Layer also operates on the fiscal side: what tax treatment infrastructure receives, and whether that treatment is visible to the public. Both dimensions of permission shape the economics of building AI at scale.
Business Engineer Framework
The Permission Layer
The Permission Layer maps how government and regulatory structures determine which AI capabilities can be built, deployed and — as this episode makes clear — fiscally legible. When the instrument of public support leaves no public trace, the oversight apparatus is operating blind. The Business Engineer Map of AI traces where each layer sits and where the structural gaps are largest.
Three senators sent four letters requesting information about statutory, lawful tax deductions — and the most structurally significant fact is not any of the numbers attached to those letters, but the reason the letters had to be sent at all: no public filing, no agency database and no regulatory disclosure currently produces the figure an observer would need to evaluate what the AI infrastructure buildout actually costs the public fisc. The October 12 deadline is a request for data that should, by any reasonable transparency standard, already exist.
These are letters requesting information. They are not findings, not charges and not evidence that any company did anything improper — the deductions at issue are provided by statute, and claiming them is lawful. The company-level figures above are the senators’ office’s own calculations, not company disclosures and not Joint Committee on Taxation figures.The Joint Committee’s $67 billion is an economy-wide estimate covering all companies for retroactive breaks in 2026 — it is not an AI figure, not a data-centre figure and not a figure for these four firms. The two sets of numbers measure different things over different scopes, are not summed above, and no proportion between them is calculated. No estimate for Meta is published, and none is supplied or inferred above. No company’s actual claimed deduction, no company response, the senators’ methodology, the data-centre share of any deduction and lobbying spend figures are not established and do not appear — a limit of this reporting rather than evidence that none exist. Nothing above is tax advice or investment advice, nothing above takes a position on whether the underlying tax provisions are sound policy, and nothing above predicts responses, legislation, enforcement or any company’s behaviour.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
