Whatever It Takes

"Whatever it takes" is one of the most famous phrases in modern finance. Mario Draghi, then head of the European Central Bank, used it in 2012 to save the euro, and it worked — because a central bank has a genuinely unlimited balance sheet, and a promise to spend without limit, from an institution that can actually spend without limit, is self-fulfilling. The phrase has since been borrowed by founders and executives to convey conviction, and it has become, in that corporate usage, a warning sign rather than a reassurance — because a company, unlike a central bank, does not have infinite money, and "we will spend whatever it takes" from an entity with a finite balance sheet is not a guarantee of success but a declaration that the normal disciplines of budgeting and return-on-investment have been suspended in favor of a vision. In 2026, no company embodies this more completely than Meta, and no executive more than Mark Zuckerberg, who has bet the future of a roughly $1.5-trillion company on the conviction that artificial superintelligence is at hand and that he will spend, quite literally, whatever it takes to build it first.

The numbers are difficult to absorb because they are so large. Meta has guided its 2026 capital expenditures to a range of roughly $125 billion to $145 billion — raised, over the course of the year, from a prior range, as the ambition expanded — up from something like $72 billion in 2025, which was itself a record. To put $145 billion in perspective: it is more than the entire annual revenue of all but a handful of companies on earth, being spent in a single year, by one company, mostly on Nvidia GPUs, custom chips from Broadcom and Amazon, and the vast data centers and power infrastructure to house them. Meta spent roughly $19 billion on capital expenditures in the first quarter of 2026 alone. It has stood up a new division, Meta Superintelligence Labs, and gone to war for talent, poaching dozens of the world's top AI researchers with compensation packages reported to reach into nine figures per person. Zuckerberg's stated goal is nothing less than "personal superintelligence" — a machine smarter than humans, in everyone's hands, built by Meta. The conviction is total, and so is the spending.

The bill that arrives later

The first thing to understand about a capital-expenditure binge of this magnitude is that its cost does not hit the income statement when the money is spent; it arrives later, and it arrives relentlessly, as depreciation. When Meta spends $145 billion on chips and data centers, it does not record a $145 billion expense that year. It capitalizes the spending — puts the assets on its balance sheet — and then depreciates them over their useful lives, recognizing the cost as an expense spread across the following several years. This is ordinary accounting, but it has an extraordinary consequence when the spending is growing this fast: the depreciation expense compounds. Each year's enormous capex adds a new layer of depreciation that stacks on top of all the prior years' layers, so that even if Meta's spending plateaus, its depreciation charge keeps climbing for years, eventually reaching tens of billions of dollars annually — a permanent, growing drag on profits that arrives whether or not the AI infrastructure is generating commensurate revenue.

This is the trap hiding inside the "whatever it takes" spending, and it is the same trap that haunts every AI-capex story in this series. While the money is being spent and capitalized, it feels almost free — it flatters the present, building gleaming data centers and signaling boundless ambition, without yet denting reported earnings much. But the depreciation bill is deferred, not avoided, and it grows into a mountain. Within a couple of years, Meta will be carrying tens of billions of dollars of annual depreciation on AI infrastructure, and the only thing that justifies that drag is a commensurate flood of new AI revenue. If the revenue arrives, the depreciation is just the cost of a wildly profitable new business. If it does not — if superintelligence proves further away than Zuckerberg believes, or if it arrives but does not translate into the cash flows the spending assumed — then the depreciation alone, on a hundred-plus billion dollars a year of fast-obsolescing GPUs, becomes a crushing weight on margins, with no offsetting revenue to carry it. The spending is a promise to the future. The depreciation is the future collecting on the promise, on schedule, regardless of whether the bet paid off.

The single engine funding the moonshot

Here is the structural fact that the soaring stock price obscures: this entire superintelligence moonshot is funded by one business, and it is not an AI business. It is advertising. Meta's cash machine — the targeted ads served across Facebook, Instagram, and WhatsApp — generates the enormous, reliable, high-margin profits that pay for everything else, including the $145 billion of AI capex and the nine-figure researcher salaries. In a recent quarter Meta reported record total revenue above $56 billion, the overwhelming majority of it advertising, growing at a healthy clip. As long as that ad engine gushes cash, Zuckerberg can spend whatever it takes on superintelligence without raising a dollar of outside capital, because the core business simply produces the money. The moonshot is being funded out of pocket, from the deepest pocket in the consumer-internet world.

But look more closely at that funding engine and the comfort frays. The advertising business is mature, not young; it is the dominant incumbent in a market that grows with the digital ad economy but is not going to multiply many times over. It faces its own questions about AI disruption — the same generative-AI forces reshaping every software category. And in the most recent results, two warning signs appeared together: Meta reported its first-ever decline in daily users — daily active people slipping about 5% to roughly 3.56 billion, a drop the company attributed partly to external shocks like the Iran war and a WhatsApp ban in Russia rather than pure engagement decay — even as it announced it would cut roughly 8,000 jobs, about 10% of its workforce. The user dip may prove temporary and externally driven; but the fact that the streak ever broke, in the same quarter as the layoffs, is the kind of hairline crack worth noticing. Sit with the juxtaposition. The company is spending $145 billion to build a superintelligence while its core user base ticks down for the first time in its history and it lays off thousands of employees to control costs. The moonshot is being financed by a mature advertising business that is, at the margin, showing the first faint signs of strain — and the entire superintelligence bet rests on the assumption that this one engine keeps producing the tens of billions a year required to feed it. If the ad business stumbles, even modestly, the funding for "whatever it takes" tightens, and the bet that depends on it becomes far harder to sustain. The moonshot has a single point of failure, and it is the comfortable old advertising business everyone has stopped worrying about.

The money does not vanish — it becomes someone else's revenue

There is a systemic dimension to Meta's spending that connects it to nearly every other story in this collection, and it is worth tracing the dollars. Meta's $145 billion does not disappear into a void; it flows, in enormous quantities, to a small set of other companies — to Nvidia for GPUs, to Broadcom for the custom AI chips examined in an earlier chapter, to the data-center developers and the power utilities that the AI-electricity chapter described. Meta is one of the single largest nodes in the great circular machine of AI capital spending, in which the hyperscalers' enormous outlays become the chipmakers' and power producers' enormous revenues, which justify their soaring valuations, which the market then extrapolates as proof that the AI boom is real and durable. Meta's capex is, quite literally, a line item in Nvidia's and Broadcom's growth stories.

This means Meta's bet is not contained to Meta. If Zuckerberg's conviction holds and the spending continues, it props up the revenues and valuations of the entire AI-infrastructure complex. But if Meta were ever to blink — to conclude that superintelligence is further off than hoped, or that the returns are not materializing, and to cut its capital spending sharply — the effect would not stop at Meta's own income statement. It would ripple straight into the order books of Nvidia and Broadcom and the demand forecasts of the utilities, because one buyer's $145 billion of spending is a meaningful slice of those sellers' growth. The hyperscalers' AI capex is mutually reinforcing on the way up and mutually reinforcing on the way down, and Meta is large enough that its individual decision is a systemic variable. "Whatever it takes," spent by a company this size, is not just a bet on Meta's future; it is one of the pillars holding up the valuations of the chip and power names that the rest of this series has questioned. Pull the pillar and the building leans.

The asset that ages before it is paid for

There is a further, technical danger buried in the depreciation problem, and it concerns the kind of asset Meta is buying. The bulk of this spending goes into AI accelerators — GPUs and custom chips — that are advancing so rapidly that each generation is dramatically more capable than the last. That relentless progress, which is wonderful for the technology, is treacherous for the balance sheet, because it means the chips Meta buys this year may be functionally obsolete in two or three years, superseded by hardware that does the same work far more cheaply. Yet companies typically depreciate these assets over longer schedules — five or six years — which assumes the chips remain useful and valuable across that span. If the real economic life of an AI accelerator is shorter than the accounting life, then Meta (and every hyperscaler) is under-depreciating — carrying assets on the books at values the market would no longer pay, with a hidden write-down waiting to be recognized. The faster AI hardware improves, the worse this mismatch becomes, and the more the reported profits of the whole sector may be flattering reality.

Zuckerberg himself has, perhaps inadvertently, signaled awareness of the overbuild risk. He has floated the idea that Meta could enter the cloud-computing business — renting out its AI infrastructure to others — and described it explicitly as a "natural hedge against overinvestment in AI infrastructure." Read that phrase carefully, because it is a tell. A man spending $145 billion a year with total conviction does not, in the same breath, casually mention a "hedge against overinvestment" unless the possibility of overinvestment is real enough to require hedging. The cloud idea is Meta quietly acknowledging that it might be building more capacity than its own products can use — that "whatever it takes" might take more than it needs — and looking for a way to sell the excess. It is the faint sound of a budget constraint reasserting itself inside the mind of the man who declared he had suspended it.

We have seen this conviction before — at this company

The most important fact in evaluating Meta's superintelligence bet is that we do not have to imagine how a Zuckerberg "whatever it takes" wager on an imminent future plays out, because we have just watched one. Beginning in 2020, Zuckerberg became convinced that the metaverse — an immersive virtual world — was the next computing platform, imminent and inevitable, and that whoever built it first would win the future. He renamed the company Meta to signal the conviction. And he spent. Meta's Reality Labs division, the metaverse effort, has now accumulated operating losses of more than $80 billion since 2020 (around $83.5 billion through 2025) — a sum larger than the entire market value of most public companies — burning roughly $4 to $6 billion per quarter, with the cumulative losses still mounting and the immersive-computing future Zuckerberg promised still nowhere near arriving at the scale he predicted. The metaverse was a "whatever it takes" bet on an imminent platform shift, made by the same man, at the same company, funded by the same ad business, justified by the same logic that whoever builds it first wins everything. It has been, by any honest financial accounting, one of the largest value-destroying corporate bets in history.

This is the precedent the superintelligence bet must be weighed against, and it cuts in two directions, both of which matter. On one hand, it is a stark warning: Zuckerberg has demonstrated, with eighty billion dollars of evidence, that his conviction about an imminent technological future can be both total and wrong about the timing — that he will pour fortunes into a vision on a schedule the world does not keep, and that the public shareholders, who cannot outvote him, will absorb the losses. The superintelligence bet is the metaverse bet an order of magnitude larger, made by a person whose last bet of this exact shape destroyed tens of billions of dollars. On the other hand — and intellectual honesty requires stating it — the bull will argue that AI is different, that unlike the speculative metaverse, artificial intelligence is already demonstrably real and revenue-generating, that the analogy flatters the skeptic. There is something to that. But the relevant lesson of Reality Labs is not that the technology was fake; it is that Zuckerberg's certainty about timing and his willingness to spend unlimited sums on it are not reliable guides to returns — and that a founder who controls the company through supervoting shares can sustain a money-losing conviction far longer, and far more expensively, than a normal corporate structure would permit. The man is making the same kind of bet, for the second time, with ten times the chips, and the house he is betting belongs partly to shareholders who cannot fold his hand.

The conviction that cannot be voted down

That governance point deserves its own emphasis, because it is what makes Meta's bet different in kind from an ordinary corporate capital-allocation decision. Through Meta's dual-class share structure, Mark Zuckerberg controls a majority of the voting power despite owning a minority of the economics. This means the decision to spend $145 billion a year on superintelligence — to bet the company's future on his personal conviction — is, in the end, his alone. Public shareholders can sell if they disagree, but they cannot vote it down, cannot force a more disciplined capital-allocation policy, cannot install a board that would rein it in. They are passengers on a wager whose size and direction are set by one man's belief about the future, and their only choice is to ride or to leave. When that man's conviction is right, the structure looks like visionary founder-leadership immune to short-term market pressure — exactly the freedom that let Amazon and others invest through skepticism. When it is wrong, it looks like an autocrat spending other people's money on a private obsession with no check. Reality Labs was the second story. Superintelligence will be one or the other, and the shareholders do not get a say in which.

None of this is a prediction that Meta's bet fails. It might be the defining triumph of the age. Artificial intelligence is real, advancing fast, and genuinely capable of transforming Meta's advertising business and creating entirely new ones; if superintelligence is in fact close, and if Meta's enormous head start in compute, data, and talent lets it win, then $145 billion a year will look like the cheapest fortune ever spent, and the doubters — as with every genuine technological revolution — will look foolish. The conviction may be vindicated. That possibility is real and should be weighted honestly.

But the phrase itself remains the warning. "Whatever it takes" is not a strategy; it is the suspension of strategy — the decision to stop asking what a thing is worth and start asking only whether it can be won. It worked for Draghi because a central bank truly has no budget constraint. It is working, for now, for Zuckerberg because the advertising business has been generous enough to make the constraint feel distant. But a corporation does have a budget, denominated in the depreciation that is even now compounding toward a mountain, in the ad profits that are showing their first hairline cracks, and in the patience of shareholders who have watched this exact man make this exact kind of bet before and lose eighty billion dollars on it. The metaverse asked whether Meta would spend a fortune chasing a future that arrives on its own schedule rather than Zuckerberg's. The answer was yes. The superintelligence bet asks whether anyone — the market, the shareholders, the man himself — learned anything from the asking. We will find out the way we always do: later, when the depreciation comes due, on a future that will arrive exactly when it is ready, and not one quarter sooner because someone was willing to spend whatever it takes to hurry it.