The dispatch reached me as six bullet points and a vibe. No dollar figure. No timestamp. No named source. Just a headline claiming Meta had raised its "AI spending floor" and was reviving a non-voting Class C share class โ the kind of thin, unsourced item that lands in my aggregator feed a dozen times before breakfast and usually dies there.
I pulled it anyway. Because when I was auditing ERC-20 whitepapers in 2017, the ones that scared me most weren't the elaborate frauds. They were the two-paragraph ones. Sloppy disclosure is not the absence of a structure. It's a structure someone didn't want you to read closely.
So I read it closely. And the actual story isn't the capex number. It's that two separate instruments โ a spending commitment and a share class with no vote โ appeared in the same news cycle, and one of them is a governance token wearing a corporate law costume.
The context you need, briefly. Meta has done this before. In 2016 it proposed a non-voting Class C recapitalization, explicitly so the founder could honor a pledge to give away 99% of his shares without surrendering control. Shareholders sued. The company withdrew the plan in early 2017 rather than face an "entire fairness" review in Delaware. The controlling shareholder holds roughly 55โ61% of voting power through high-vote Class B stock. A no-vote C class does not dilute that. It creates a painless exit ramp underneath it.
Now the other half. Public guidance puts 2025 capital expenditure near $70โ72 billion, revised upward more than once from an opening range around $60โ65 billion, with management flagging a substantially larger dollar increase in 2026. That is a one-directional ladder. Every rung higher, never lower.
Here's where my audit instincts kick in, and where almost every write-up I've seen this week gets it wrong.
A floor is not a floor. There are two versions of this word and they differ by an order of magnitude in risk. Version one is guidance โ a number the CFO repeats on an earnings call with the word "approximately" stapled to it. Version two is a contractual, non-cancellable purchase commitment, the line item buried in the 10-Q that survives even if the strategy that justified it collapses. If Meta means version two, then the technical roadmap is legally locked. You cannot pivot to small models and distillation when you have already signed for the megawatts. Based on my experience reading filing language, the phrase "floor" in a financing context is usually doing the work of version two while sounding like version one. Verify it against the purchase commitments disclosure before you believe anything else.
The structural constraint nobody is pricing. Meta is the only hyperscaler with no compute resale channel. Google rents TPUs through Cloud. Microsoft resells through Azure. Amazon through AWS. Meta has no cloud business and no paid model API of consequence. Every dollar of capex must be amortized against advertising revenue โ full stop. That is a wildly different risk profile from a company that can dump excess capacity onto a spot market.
And the accounting is the tell. Server and networking depreciation schedules are the softest lever in the entire income statement. Lengthen them by a year and the earnings hit from a hundred-billion-dollar buildout quietly softens. When you see capex guidance laddering up while operating margins hold flat, ask which assumption moved. It's rarely the revenue line.
Now zoom out, because the single-company framing is too small. Meta has assembled the full financing kit this cycle: roughly $30 billion in bonds, an off-balance-sheet data-center joint venture reported around $27 billion (the Hyperion structure), and now an equity instrument that raises capital without touching control. Debt, SPV, stock. Three pieces.
I have seen this movie. In 1999 and 2000, telecom equipment vendors financed their own customers' fiber buildouts, those customers bought more equipment, and everyone booked revenue on both sides of the same transaction. The AI supply chain has the same silhouette: chipmakers investing in the companies buying their chips, cloud providers signing decade-long compute contracts with each other. That is a risk transmission chain, not proof of endless demand โ and the party holding the bag if demand softens is not Meta. It's the power producers, the transformer makers, the optics suppliers, the second-tier cloud operators who expanded against a promise.
Which brings me to the angle that all the financial coverage missed entirely.
A Class C share with no vote is a token with its governance stripped out. The crypto industry spent a decade litigating this exact design question โ whether you can sell someone the economic upside of a system while permanently denying them any say in how it runs. We called them governance tokens. Meta calls it a share class. Same engineering, different regulator.
The detail that decides everything is the sunset clause. A no-vote class with a ten-year sunset is a financing tool. A no-vote class with no sunset is a control perpetuity device โ a machine that lets the founder sell, donate, or pledge shares forever while his voting position never moves. If there is no sunset, treat every future governance commitment from the company as non-binding.
And one honest counterpoint, because I don't want to hand you a one-sided ledger: Meta's lack of a cloud arm is also its shock absorber. Pure compute landlords face a brutal cliff where idle capacity means zero revenue. Meta's spare cycles still make the ad auction slightly better, and ad revenue doesn't evaporate on a demand shock. That's why its return on invested capital looks worst on paper and its survival math looks best in practice.
So what do I watch next? Three things, in order. The purchase commitments disclosure in the next quarterly filing โ that number tells you whether the floor is a promise or a contract. The presence or absence of a sunset provision on the Class C. And whether that share class gets used as acquisition currency, because a vote-free stock is the cheapest possible rifle for a company that is trying to buy back years of research it never built.
Scanning the noise for the signal, the signal here is not the seventy billion. It's the fine print wrapped around it. From ICO hype to on-chain truth โ the instrument changed, the question didn't. Who holds the vote, and when does it expire?