Two Numbers, One Lie: A Forensic Audit of Meta's Q3 Earnings Print
Meta's fiscal Q3 print contains two numbers that cannot both be describing the same company. Operating income: $12.99 billion, down 37%. Earnings per share: $3.83, up 265% from $1.05 a year earlier. Both figures appear in the release. Both are arithmetically correct. Only one of them is informative.
This is not a rounding dispute. It is the oldest failure mode in financial reporting wearing new clothes โ a comparison between a distorted base and a normalized numerator, presented without the adjustment that would make the comparison valid. Anyone who has read a token vesting schedule recognizes the shape of this problem. The art is the hash; the value is the proof. When the proof does not reconcile, you do not defer to the narrative. You recompute.
Start with the reconciliation, because the numbers do resolve cleanly. Revenue of $64.6 billion, up 26%, implies a prior-year base near $51.27 billion. Meta's actual FY2025 Q3 revenue was $51.24 billion. Operating income of $12.99 billion, down 37%, implies a prior-year figure of roughly $20.62 billion against an actual $20.53 billion. The arithmetic is self-consistent. What is missing is the fiscal year label โ this is FY2026 Q3, and the source document never states it. That omission is the first fingerprint of a translated secondary release, and it matters, because the translation preserved the money and discarded the users.
Now the EPS line. Prior-year Q3 GAAP earnings per share came in at $1.05. That figure absorbed a one-time, non-cash income tax charge of approximately $15.9 billion. Strip the charge out and the comparable base lands in the neighborhood of $7 per share. Against that base, this quarter's $3.83 represents a decline north of 40 percent โ not a 265 percent expansion. The release ran both signals side by side and explained neither.
That is the largest logical discontinuity in the document. It is the exact seam where a professional reader should stop reading and start recomputing. Non-cash does not mean non-informative. It means the distortion lives in the comparability layer rather than the liquidity layer, and comparability is precisely what a year-over-year percentage claims to measure.
Follow the capital expenditure line, because that is where the actual architecture change lives. Full-year guidance was raised to $135โ145 billion. Reality Labs is guided to a 2026 operating loss of $18โ19 billion. Run the ratio: capital intensity near $140 billion against operating cash flow that Meta has historically generated in the $80โ100 billion band. The ratio approaches 1.0. It may cross it.
I spent four months benchmarking proof-generation cost against L2 gas expenditure during the last bear market, and I recognize what a ratio crossing 1.0 means. It means the business no longer funds its own buildout. Free cash flow converges on zero, then turns negative. The Class C share issuance is not a growth signal. It is a balance-sheet signal dressed as a growth signal, and the two are separated by exactly one line item.
This is a structural migration from light-asset software to heavy-asset infrastructure, executed at a velocity that leaves no margin for error in demand forecasting. Depreciation is not discretionary. Once the compute is booked, the amortization schedule runs whether or not inference demand arrives on time. In protocol terms, the state transition has already been committed, and no subsequent upgrade reverts it.
Here is what the source material does not contain, and the absence is more instructive than anything it does. No daily active users. No daily active people. No time spent. No segment revenue. No free cash flow statement. No earnings call guidance. No regulatory disclosure. Seven data points in total, every one of them denominated in dollars.
For a company whose entire revenue is a monotonic function of captured attention, deleting the attention variable is not an editorial oversight. It is a filter. What survives translation is what circulates. What circulates is what moves price. The mechanism is identical to the one I documented in 2021, when I traced how sixty percent of popular NFT collections failed on gateway cache-policy changes โ not because the assets were fake, but because the reporting layer never audited the storage layer. The mint was on-chain. The metadata was not. The ownership claim and the ownership proof had decoupled, and nobody noticed until a provider changed a header.
The same decoupling is running here. The earnings claim is on-chain. The user metrics sit behind a gateway nobody queried. Reentrancy doesn't announce itself. It arrives inside a function that looked correct in isolation, in a contract that passed review, on the day the balance moved.
I have a resilience score I apply to infrastructure claims, and it has three inputs: whether the funding source is internal or external, whether the cost curve is fixed or variable, and whether the reporting layer is auditable by the reader. Meta fails the third input completely in this document and is mid-failure on the first two. A project that cannot show you its user curve while showing you its tax-adjusted EPS has told you which variable it believes you will not check.
It will not be checked by the outlets that ran the 265 percent headline. It will be checked by the cash flow statement, four quarters from now, when the depreciation schedule comes due and the demand curve is still in transit.
We do not build for today. We build for the quarter when the amortization lands and the guidance has already been spent.