Meta’s $35 Billion Raise Shows AI Capex Has Outgrown Cash Flow
The company reported its first negative free cash flow in over a decade, then launched $35 billion of equity and preferred stock, with a $60 billion at-the-market programme to follow. Its chip suppliers are buying the shares.
Meta launched 35 billion dollars of concurrent equity and mandatory convertible preferred stock on Monday, and the announcement said plainly what the money is for: AI infrastructure and personal superintelligence. Alongside it, the company disclosed a financing plan that reaches further, a 60 billion dollar at-the-market programme to begin once its new Class C shares start trading in the first quarter of 2027. Taken together, the plan could raise as much as 95 billion dollars, against capital expenditure guidance of 135 to 145 billion dollars for 2026.
The number that gives those figures their meaning came three days earlier and does not appear in the release. Meta reported negative free cash flow of 2.12 billion dollars in the third quarter, against positive free cash flow of 10.63 billion dollars a year earlier. It was the first cash burn at the company in more than a decade. AI capex has crossed a line: the business no longer funds it.
The raise, and what it is for
The structure deserves a close reading because it is more sophisticated than a share sale. The common stock portion, 22.5 billion dollars, comes through forward sale agreements with Morgan Stanley and Goldman Sachs, which means Meta does not receive the cash now; it receives it on settlement, on dates it chooses up to 19 November 2027. The other 12.5 billion dollars is 250 million depositary shares, each a twentieth of a share of Series A mandatory convertible preferred stock with a 1,000 dollar liquidation preference, converting automatically around 15 August 2029 into Class A shares or, if conditions are met, Class C shares. The company intends to list the depositary shares on Nasdaq under the symbol METAM.
That combination buys time and flexibility. Mandatory convertible preferred pays a dividend and converts on a schedule, so it behaves like equity with a delayed dilution date and a known coupon. Forward sales let Meta sell shares at today’s price and take delivery of the cash later, which helps when capital is needed in tranches. Neither instrument is unusual on its own. What is unusual is the scale and the timing for AI capex, coming from a company that until this quarter funded its ambitions from operating cash flow.
The cash-flow pivot
The third-quarter numbers explain the urgency. Revenue reached 64.6 billion dollars, up 26 per cent year on year, with net income of 9.9 billion dollars. Operating cash flow for the first three quarters totalled 97 billion dollars, and the company held roughly 82 billion dollars in cash and marketable securities. Those are extraordinary figures, and they are why the raise is not a rescue. But capital expenditure runs at 135 to 145 billion dollars for the year, the company owns data centres in 23 American states, it is building four superclusters of more than a gigawatt each, and it expects to nearly double total infrastructure capacity in 2026 alone. When capex outruns operating cash flow, the difference comes from somewhere, and this quarter it came from the balance sheet.
The Class C decision
One structural choice deserves more attention than it has received. The at-the-market programme will sell Class C shares, a new class with no voting power, established by an amended certificate of incorporation filed in Delaware on 21 October. That is a governance decision dressed as a financing one. Selling 60 billion dollars of non-voting stock lets Meta raise capital at scale without diluting founder control, and it means the investors funding the buildout get an economic claim without a say in how the money is spent. Shareholders accepted dual-class structures long ago; a third class created specifically to fund a spending programme is a further step, and it is worth asking what it implies about how much capital the company expects to need.
The circular loop
The detail that will draw the most scrutiny sits near the end of the release. NVIDIA, AMD and Broadcom have each indicated a non-binding interest in buying up to an aggregate 5 billion dollars of Meta’s Class A stock, at the same price as every other purchaser. Those three companies are Meta’s principal silicon suppliers. NVIDIA signed a multiyear deal in February covering millions of Blackwell and Rubin GPUs. Broadcom is co-developing Meta’s custom MTIA accelerators, including what the two companies describe as the industry’s first 2-nanometre AI compute accelerator. AMD agreed in February to sell Meta up to 60 billion dollars of chips over five years, in a deal that lets Meta buy as much as 10 per cent of AMD.
Taken together, the pattern is circular in a way that deserves a name. Suppliers sell hardware to a customer, take equity in that customer and, in AMD’s case, hand the customer equity in themselves. Reuters described the AMD arrangement as a closed-loop hyperscaler pattern at the time, and this offering extends it. The accounting is legitimate and the strategic logic is obvious, because a vendor holding equity in its largest customers has an interest in their continued spending. But it also means that part of the capital funding AI infrastructure now originates inside the AI supply chain, which makes the whole system more sensitive to a single shift in sentiment. If the vendors’ shares fall, the capital they might have deployed into their customers falls with them.
What the money buys
The release is unusually specific about the asset base, which helps. Meta owns data centres in 23 states and is building four superclusters of more than a gigawatt each, with a 600 billion dollar American commitment by 2028 behind them. Two of the largest are already under construction: Prometheus, a gigawatt site in Ohio, and Hyperion, a five-gigawatt site in Louisiana. On the product side the company reports 3.64 billion daily users across its apps, more than 10 million Muse users in the United States and Canada within weeks of launch, with shopping the leading use case, and more than two million businesses using its Business Agent weekly across 30 million conversations. It also flags a Meta Enterprise Platform aimed at high-margin inference, and Muse Wonder, its first frontier model.
That is the bull case, and a fair account has to state it. Meta’s AI investment is not speculative in the way a startup’s is. Advertising revenue is growing 26 per cent on the strength of recommendation and ranking improvements the company attributes to AI, and engagement sits at all-time highs on Facebook, Instagram and WhatsApp. The argument that AI already pays for itself inside the core business is credible. The open question is whether it pays fast enough to fund a capital programme of this size without permanently changing the company’s relationship with capital markets. The wider context makes the structure easier to read. Every hyperscaler is now competing for the same capital, the same power and the same chips, and the largest of them have begun to look less like software companies and more like utilities with a construction programme.
Plausible Growth Rate
That commitment cannot come from retained earnings alone at any plausible growth rate. The question analysts have been circling for a year, whether AI capex can be funded internally, now has an answer from the largest advertiser in the world, and the answer is that it cannot, at least not at this pace.
What to watch
Four things will show whether this financing was prudent or the top of a cycle. The first is free cash flow: whether the third quarter’s negative figure proves an anomaly or the start of a run. The second is the pricing of the preferred, because the dividend rate set at pricing tells you what investors demand to hold Meta in equity form for three years. The third is whether the at-the-market programme actually launches in the first quarter of 2027 at 60 billion dollars, or whether the company quietly shrinks it. The fourth is whether the vendor purchases materialise, because a non-binding indication of interest is not a commitment, and the release is careful to say so twice.
A fifth marker matters for the industry rather than the company. If Meta’s offering prices well, other hyperscalers will follow with their own equity, and the buildout will be funded increasingly by public markets rather than by cash flow. If it prices badly, the capital available for data centres tightens for everyone, and the smaller neoclouds feel it first.
What it signals
The offering marks a phase change in how AI capex gets paid for. For three years the hyperscalers funded their data centres from the cash their existing businesses threw off, which made the spending look self-limiting and therefore safe. Meta has now told the market that its ambitions exceed that envelope, and it chose equity over debt despite investment-grade ratings that would let it borrow cheaply. That choice is revealing. It suggests management wants the risk shared, and it suggests the capex curve keeps rising past 2026. The market’s reaction to the Class C announcement, a seven per cent fall in the shares, says investors understood the message. The next twelve months will establish whether markets will finance the AI buildout on these terms, and every other company building at this scale is watching the answer.