The offering was enlarged before the market opened. On Monday, Meta Platforms set out a plan to sell $97.5 billion worth of stock, the largest equity raise in its history, and by Tuesday morning Nvidia had agreed to buy $3 billion of it.
The structure came in three parts. Meta is selling $12.5 billion of preferred shares priced at $50 apiece, $25 billion of common stock at $625 a share, and arranging a roughly $60 billion at-the-market program through which it can sell more common stock into the market over time. The preferred carries conversion terms tied to the common, which is why its issue price sits so far below the common price.
The at-the-market portion depends on a new class of shares. Meta’s existing structure has Class A and Class B stock, with the second carrying extra votes and held largely by Mark Zuckerberg. The filing creates a Class C, non-voting, which lets the company issue equity without diluting the founder’s control.
That detail matters more than the size of the sale. Zuckerberg has said for years that he intends to run the company for decades, and every large capital raise before this one was structured to preserve his voting position. A non-voting class lets Meta raise money from index funds and institutions without handing them a say.
Demand appeared to be there. Nvidia’s $3 billion commitment, announced on Tuesday, is a strategic purchase as much as a financial one: the chipmaker said it would expand its work with Meta on the Muse family of assistant models and on inference and training performance. Meta has been among the largest buyers of Nvidia’s accelerators.
Two smaller commitments had been disclosed earlier. Nvidia, Advanced Micro Devices and Broadcom together agreed to take up to $5 billion of Meta’s common stock, according to the company, and Nvidia’s new purchase sits on top of that allocation rather than replacing it.
The proceeds are earmarked for data centers and consumer AI products. Meta is building two clusters large enough to be described in gigawatts: Prometheus, in Ohio, and Hyperion, in Louisiana, the second of which is planned to scale to five gigawatts across an area the size of a small airport.
Gigawatts are expensive in a way that is easy to state and hard to finance. A single gigawatt of AI-capable data center capacity costs roughly $30 billion to $40 billion once land, power, cooling and the chips are counted, according to estimates published by several brokerages this year. Two clusters of that scale absorb more than Meta’s entire cash flow.
The company has therefore been raising money on both sides of the balance sheet. Last October it sold $30 billion of bonds in its first-ever investment-grade debt offering. It has also brought outside investors into specific projects, including a joint venture with Blue Owl Capital for the Louisiana site that shifted part of the construction cost off its own books.
Adding equity to that mix is unusual for a company of Meta’s profitability. Debt is cheaper, and interest is deductible, which is why most large technology companies have funded their buildouts with bonds. Equity dilutes existing holders and signals either that debt capacity is exhausted or that management wants room for more borrowing later.
Investors made their judgment quickly. Meta’s shares have been volatile this year as spending guidance rose faster than revenue guidance, and the raise was read in two ways: as a company that sees demand it cannot fund internally, or as one whose capital budget has outrun its business.
The preferred shares offer a clue. Buyers of a $50 instrument that converts into a $625 common stock are being paid for time and for the risk of a decline. The terms, which include a liquidation preference ahead of the common, suggest the book was built with institutions that wanted downside protection rather than upside exposure.
One detail sits awkwardly with the rest. Meta pays a quarterly dividend and has bought back tens of billions of its own stock in recent years. A company that repurchases shares while simultaneously selling new ones is moving capital in two directions at once, and the filing does not say whether the buyback will pause. Investors who own the stock for the payout will want an answer before the next declaration.
Meta is not alone in going to the market. Oracle has raised billions in debt to build capacity for OpenAI. CoreWeave borrows against contracts. Broadcom has arranged vendor financing for its own customers, a structure that arrived the same day as Meta’s filing and involves a $63 billion lease facility for Anthropic.
The common thread is that the compute buildout is being financed like infrastructure rather than like software. Software companies convert revenue into cash at high margins and need little capital. Data centers look more like utilities: long-lived assets, predictable depreciation, and a permanent need for outside money.
For Zuckerberg, the trade is control for capital. The Class C shares let him take the second without giving up the first, which is the reason the structure exists. The question investors will test over the next year is whether the assets those shares pay for earn more than the cost of the money raised to build them.
