Executive Summary

Key Numbers

Four numbers carry this announcement. The first three are amounts written into filings and company statements. The last one is a slot where nothing has been written.

Sources: Meta SEC Form FWP · Reuters · CNBC (October 26–27, 2026)

The filing Meta made on October 26 had two slots. First, about $35 billion raised through a public offering: roughly $22.5 billion of common stock, plus about $12.5 billion of preferred depositary shares that convert into common stock when their date arrives. The remaining $60 billion is to be filled by selling shares at market prices, a little at a time, from the first quarter of 2027 onward. That program opens only after the new non-voting shares have started trading.

The next day the first slot grew. Orders came to more than five times the stock on offer, so Meta raised $22.5 billion to $25 billion, and the price was set at $625 a share. The preferred depositary shares sell at $50 each, 250 million of them. Add the two and you get $37.5 billion; add the $60 billion of at-market sales and you get the $97.5 billion total of the plan.

The structure printed next to the figure deserves as much attention as the figure. On the common stock side, Meta is not issuing shares and selling them today. According to the filing, it is entering forward sale agreements with Morgan Stanley and Goldman Sachs. Meta picks the day it hands over the shares and collects the money, any time up to November 19, 2027, and settles at the offering price less the underwriting discount. Which is to say that the day $25 billion reaches Meta's account has not been fixed yet.

Other terms were not all written down on the first day either. The dividend rate on the preferred depositary shares was left to pricing in the October 26 filing, and the figure that it settled at 6.50% a year came out in the reporting that followed. There is also a separate allotment the underwriters may take up within 30 days: $3.375 billion of common stock and 37.5 million depositary shares. So $97.5 billion is not a ceiling either.

Why equity, then. Meta has been paying for its facilities out of the cash it earns. Capital spending planned for this year runs from $135 billion to $145 billion, close to double last year's $72.2 billion. As of September 30 the company held $82.45 billion in cash and equivalents against $80.92 billion of long-term debt. Free cash flow for the third quarter came in at negative $2.12 billion, the first negative reading since 2012. Money left the books faster than it came in.

How that happened is written plainly in the quarterly results. Cash from operations in the third quarter was $32.786 billion, and capital spending was $34.91 billion. The gap between the two is the negative free cash flow. Widen the window to January through September and cash from operations stands at $96.873 billion, which is still thick. Meta chose to sell stock anyway. Over those same nine months it bought back none of its own shares, where a year earlier it had repurchased $26.248 billion. The company that was buying its stock back has moved to the seat where it sells it.

Three companies appear in the filing under the heading of cornerstone investors: Nvidia, AMD, and Broadcom. The filing records that the three have, severally and not jointly, indicated an interest in buying up to $5 billion of the common stock at the same price and on the same terms as every other investor.

Two things have to be kept apart here. First, this is an indication of interest and not a commitment to buy. The filing states outright that it carries no legal obligation, that any of them may buy more, buy less, or buy nothing at all, and that the underwriters may adjust allocations as well. Second, there is no discount. Where cornerstone investors usually come in a few percent below the market, these three pay the same $625 a share as everyone else.

No discount, though, means only that these three were not marked down separately. It does not mean $625 was the going price. That level sits 2.7% below the close on the Friday before the announcement. Everyone who came into the offering bought that much below the market, and the underwriting discount applies to the shares handed to the three chip companies just as it does to the rest.

On October 27 two of the three named their amounts. Nvidia put in $3 billion and AMD $1 billion. Broadcom is on the cornerstone list by name only and has not disclosed a figure. The filing caps the three together at $5 billion and $4 billion has been disclosed, which leaves $1 billion of room, but because an indication of interest binds no one, that subtraction cannot settle what Broadcom will buy.

How large a place that $4 billion holds in the order book can be estimated. With the offering more than five times oversubscribed, more than half of the allocated shares went to long-only funds and existing shareholders, according to the reporting. The fifteen largest recipients took roughly 70% of the deal. Split that share evenly across fifteen and each comes to a little over $1.2 billion. Nvidia alone disclosed $3 billion, more than double that average.

What the three have in common sets the character of the list. Every chip and system going into Meta's AI facilities comes from one of these three.

What separates this from an ordinary share purchase is that the investment news arrived on the same day as the supply news. Nvidia disclosed its $3 billion alongside word that deployments of its Vera Rubin systems and Vera CPUs at Meta are growing, and added that it is widening its work on Muse, Meta's AI agent. AMD disclosed its $1 billion alongside word that its Helios racks are installed at Meta and running Muse, with EPYC processors beside them. Both are continuations of supply agreements signed in February of this year.

Inside one week, then, money moved in two directions. When Meta buys chips, that is revenue for Nvidia and AMD. When those two buy Meta stock, that is capital for Meta.

The two directions are not the same size. The $4 billion disclosed comes to less than 3% of what Meta says it will spend on facilities this year. The amount going the other way has not been disclosed company by company. Six gigawatts is a measure of electrical capacity, not of money.

A total does exist. Money Meta has committed to spend under leases and contracts adds up to roughly $825 billion once the line items in the quarterly report are summed, up from close to $700 billion as of the July filing. That value, though, is not written on any single line. It is a figure assembled from items scattered across the report, and it is not broken out by counterparty. How much of it goes to Nvidia, to AMD, and to Broadcom cannot be read off it. That is why the upper arrow carries no amount.

Two names are in circulation for this structure. The critics call it circular financing. Nigel Green of the asset manager deVere, quoted across several outlets, points out that the companies making the chips are now buying stock in the companies buying those chips, and that money going out as investment comes back as revenue. He warns that this makes it hard for an outsider to tell demand a customer pays for out of its own pocket from demand funded inside the industry, and that someone outside the circle ends up holding the bill.

Those pushing back call it vendor financing. The economist Noah Smith sees it as different from the fake revenue inflation of the dot-com years. Back then investors were handed money that had not been earned and the round trip back into revenue was hidden, whereas these amounts are disclosed and who put what where is public. Nvidia's Jensen Huang rejects the phrase circular financing outright, calling the equity stake a very small fraction of the capital a customer needs and a mark of long-term conviction. Put that in numbers and it reads as follows. Nvidia's revenue in its most recent quarter was $96.2 billion, and the $3 billion going into Meta is a little over 3% of it.

This argument did not open with Meta. Nvidia has said it would invest up to $100 billion in OpenAI. AMD gave OpenAI warrants over as many as 160 million of its own shares and tied a 6-gigawatt GPU supply agreement to them. OpenAI is reported to have agreed to pay Oracle roughly $300 billion over five years. In all three, the seller and the buyer are bound together a second time through equity or rights, and in all three the figure is a ceiling, an intention, or a condition. The $5 billion in this cornerstone arrangement is the same kind of number.

Both sides have a case. The point is that the numbers both sides use come from the same place. Whether you call it circular or call it vendor financing, the investment amounts and the scale of the supply agreements come out of statements written by the parties to the deal. William Blair wrote that the raise signals Meta's conviction in the long-term returns on its AI investment and that chasing the opportunity is growing more capital-intensive, adding that investors will keep asking about this financing and the possibility of more.

That the total Meta has committed to future spending is scattered across the quarterly report rather than gathered on one line is something this blog has written about once before. Which signals show up first when AI capital spending runs ahead of itself was the subject of a report on surplus servers.

Sort the numbers that poured out around this announcement by where they came from and three layers appear. At the bottom are the ones priced from outside or already audited. The $625 offering price was set against the orders that came in. The $82.45 billion of cash and equivalents as of September 30, the $80.92 billion of long-term debt, and the negative $2.12 billion of third-quarter free cash flow sit in the financial statements of a quarter already closed. Get those wrong and the company answers for it.

This layer also holds the values the market reprices every day. On October 22, when third-quarter results came out, Meta's stock fell nearly 7%. The same day, the premium on five-year credit default swaps on Meta rose 7.6 basis points to 114.79, the highest since those contracts began trading. On October 27, the day after the offering was priced, the stock traded as low as $621.47 intraday, below the $625 offering price.

Reading those values as a risk of default would be too much. Around 115 basis points means insuring Meta's debt for a year costs a little over 1.15%, which for a company holding $82.4 billion in cash at the end of September still sits in investment-grade territory. What to read here is direction, not height. And what these values measure is not AI demand. A share price and a credit premium are prices put on what the market believes about Meta's spending plans, not a reading taken off the data centers themselves. They belong at the bottom layer because they are priced from outside, but they do not stand in for the top layer.

In the middle are the numbers the parties to the deal wrote down themselves. Nvidia's $3 billion and AMD's $1 billion came from each company's own announcement. What the filing states is a $5 billion cap across all three, and even that binds no one. AMD's up-to-6 gigawatts, Nvidia's millions of GPUs, and the MTIA built with Broadcom are likewise scales each company wrote into its own statement. This year's $135 billion to $145 billion of capital spending is a plan for money not yet spent.

That range alone has been rewritten several times inside this year. When Meta reported last year's fourth quarter it stood at $115 billion to $135 billion. In the first quarter both ends moved up, to $125 billion to $145 billion, and after that the upper end stopped at $145 billion while the lower end climbed to $130 billion and then to $135 billion. Numbers in a financial statement are not rewritten once the quarter closes. Numbers that get rewritten sit side by side in the same table with numbers that do not. The two do not belong in the same layer.

The $825 billion of commitments from earlier straddles the two layers. It carries responsibility because it appears in a quarterly filing, which sets it apart from a figure in a press release, and it is money not yet spent, which puts it alongside one. Sort things into layers and values like this turn up on the boundary. Rather than split the taxonomy finer, note how much of each character the value carries.

The top is empty. How much Meta's AI data centers are actually running, how much power they draw, how much the company earns on a token sold: none of it appears anywhere in this announcement. Confirming from outside whether the demand is real takes numbers from this layer, and this layer has nothing in it right now.

What marks this round of coverage is that all three layers are mixed together in a single story. Numbers recording what already happened have something outside holding them up. Numbers recording what is going to happen have only the word of whoever wrote them. Most of the amounts that drew the biggest headlines belong to the second kind.

Meta's financing calendar is not something Pebblous needs to follow. The reason to look at this announcement is where the numbers sit. Amounts released in the same week all look like official disclosures from the outside, and yet what you can do to check them differs by layer. Behind a value carried in the financial statements stand auditors and regulators; behind a value carried in a company announcement stands only the company that wrote it.

Anyone who has worked on data quality knows this distinction. The columns in a single table do not all carry the same reliability. A value a sensor recorded, a value someone filled in by hand, and a value pulled in from another system sit side by side in one row, and the schema alone will not tell the three apart. Unless that distinction is written down somewhere, a model trusts all three equally.

Talk about data being ready for AI often stops at format and missing values, when what decides the judgment is where a value came from. In this announcement the amounts are not scattered; the character of their sources is mixed. If your organization publishes its own metrics outward, three things are worth settling.

• Is there a column recording who produced each number. When a value priced from outside and a value written by a party to the deal share a table, something has to mark which is which
• Is there someone who answers for it if the value is wrong. A number in a financial statement and a number in a press release differ on exactly this point
• Is there a first-hand measurement anywhere that could falsify the claim. Without one, the claim stays a matter of belief for the time being

That last item is the question this announcement leaves behind. The numbers that would settle whether AI demand is real are utilization, power draw, and revenue earned against what was spent, and everyone holding those numbers is a party to the deal. What an outsider can see, for now, is the amounts they sent each other and the prices the market put on those amounts.

Thank you for reading this far. The structure of the raise and the cornerstone investor provisions were checked against the October 26 filing, and the quarterly results and capital spending plan against the third-quarter results Meta filed the same month. The offering price, the allocation of the book, the share price, and the credit default swap levels were taken from the reporting that carried them. The remarks from Nigel Green and Noah Smith are relayed from quotations in several outlets, since the original sources could not be confirmed, and the sorting of these numbers into three layers is this article's reading. If your organization has put values from different sources into one table and lived with the result, we would be glad to hear about it.

Pebblous Data Communication Team
October 28, 2026

R.1Official Documents

• 1.Meta Platforms, Inc. (2026). "Form FWP — Class A Common Stock Offering." U.S. Securities and Exchange Commission, October 26, 2026.
• 2.Meta Platforms, Inc. (2026). "Form 8-K Exhibit 99.1 — Third Quarter 2026 Results." U.S. Securities and Exchange Commission, October 21, 2026.

R.2Industry and Press

• 3.CNBC. (2026). "Meta launches plan for $95 billion stock offering to fund AI expansion." October 26, 2026.
• 4.CNBC. (2026). "Nvidia, AMD to invest $4 billion in Meta as tech giants deepen AI partnership." October 27, 2026.
• 5.CNBC. (2026). "Meta stock sinks nearly 7% as AI spend pushes free cash flow negative." October 22, 2026.
• 6.Reuters / The Globe and Mail. (2026). "Meta increases size of planned equity offering to $97.5-billion." October 27, 2026.
• 7.Bloomberg / Yahoo Finance. (2026). "Meta's $25 Billion Stock Sale Ended Five Times Oversubscribed." October 27, 2026.
• 8.Reuters. (2026). "Meta's credit risk indicator hits record high on increased AI spending plan." October 22, 2026.
• 9.The Irish Times. (2026). "Meta plans $95bn stock offering to bankroll AI expansion." October 26, 2026.
• 10.Yahoo Finance. (2026). "Nvidia to invest $3 billion in Meta, expand partnership on Muse." October 27, 2026.
• 11.Business Recorder. (2026). "Meta eyes record $97.5 billion equity raise to fund AI expansion." October 27, 2026.