Economics

The AI Boom's Biggest Buyback Belongs to the Company That Isn't Paying for the Buildout

Nvidia added $150 billion to its repurchase program on Monday. Alphabet and Meta spent nothing on buybacks in the first half of the year — they issued debt and equity instead, and Alphabet's free cash flow went negative.

Duncan Galbraith

Duncan Galbraith

— 8 min read
Share
The AI Boom's Biggest Buyback Belongs to the Company That Isn't Paying for the Buildout

Nvidia added $150 billion to its repurchase program on Monday. Alphabet and Meta spent nothing on buybacks in the first half of the year — they issued debt and equity instead, and Alphabet's free cash flow went negative.

Nvidia's board authorized an additional $150 billion under its existing share repurchase program on Monday, raising the remaining total to $235 billion. The company called it "the largest share repurchase authorization increase in history" and said it expects to execute the whole remaining program by the end of fiscal 2028. Reuters put the comparison plainly: it eclipses Apple's $110 billion authorization in 2024. "Our cash generation gives us the capacity to invest in the technologies that advance this transformation and return capital to shareholders," Jensen Huang said.

Read as a money-flows document rather than a statement of confidence, the announcement says something the superlative obscures. The companies this boom is being built for have stopped returning cash. They have started raising it. The record return of capital belongs to the one firm in the stack with nothing to build.

The buyers stopped buying back.

Alphabet ran the second-largest repurchase program in the United States in 2025 and spent $28.3 billion on buybacks in the first half of last year. In the first half of 2026 it spent nothing — its March quarter 10-Q records no Class A or Class C repurchases, with $69.5 billion still authorized. Meta ran the fifth-largest last year and spent $22.9 billion in the first half of 2025. In the first half of 2026 it spent zero, down from $10.17 billion in the second quarter of 2025 alone. You can see it in the company's own release: Meta's second-quarter results list $1.35 billion of dividends under capital return and mention no repurchases at all.

Amazon is not a departure from a habit. Its June-quarter 10-Q states that there were no repurchases of its common stock in the six months ended June 30, 2025 or in the six months ended June 30, 2026, with $6.1 billion left on a program it has essentially never used.

Microsoft is the exception on the buyback side: it returned $10.2 billion to shareholders in the fourth quarter of fiscal 2026 and more than $43 billion across the year. It is also the exception on the strain. Epoch AI's reading of the filings puts the group's capital spending at about 94% of operating cash flow in the first calendar quarter of 2026 — $148.4 billion against $157.9 billion — and projects spending to overtake operating cash flow around the third quarter. That crossover has already happened at Oracle and, by Epoch's estimate, is happening at Amazon now. Alphabet's turn is projected for early 2027, Meta's for late 2027, Microsoft's not until 2028.

And they replaced the buyback with issuance.

Amazon, Alphabet, Meta and Oracle issued roughly $194 billion of bonds in 2026 through July 7, an increase of about 79% on the roughly $108 billion the group raised in all of 2025. Meta sold $25 billion of senior unsecured notes in May and carried $83.7 billion of long-term debt at June 30. Alphabet raised $49.6 billion of equity and mandatory convertible preferred in the first half of the year, alongside $20.3 billion of net proceeds from a senior notes offering. Alphabet's second-quarter free cash flow came in at negative $5.9 billion, against positive $24.6 billion in the fourth quarter of 2025.

A buyback is a company with more cash than uses. An equity raise is a company with more uses than cash. Alphabet did both ends of that trade in the same six months.

At the supplier, cash generation halved in the quarter it announced the record.

Nvidia's second quarter of fiscal 2027, reported on August 26, was the strongest in its history on the income statement: $96.2 billion of revenue, up 106% year over year, and $59.7 billion of net income. Operating cash flow was $24.1 billion, down from $50.3 billion the quarter before — a 52% sequential decline.

The reason is in the working capital. Accounts receivable rose $22.3 billion to $63.1 billion. Days sales outstanding went from 45 to 60 in a single quarter. Nvidia attributes the build to "extended payment terms on large multiquarter agreements with certain investment grade customers." Inventory rose to $31.6 billion from $25.8 billion ahead of the Vera Rubin generation.

That is the piece of arithmetic worth keeping. The people buying the chips have stopped buying back their own stock and are borrowing instead. The person selling the chips is buying back more of his own stock than anyone in history, in a quarter when his cash generation halved — partly because his customers haven't paid him. Nvidia said on Monday that its cash generation gives it the capacity to return capital to shareholders. That was true in the March quarter, at $50.3 billion of operating cash flow. In the most recent one it was $24.1 billion.

Two disciplines on that. An authorization is not an expenditure — $235 billion remaining is a board's stated intention to spend over three fiscal years, and Nvidia's actual repurchases in the quarter just reported were about $20 billion of the roughly $26 billion it returned. And extending terms to large investment-grade buyers is not necessarily a warning sign; long multi-quarter supply agreements have logistics, and some customers are worth financing. The line item is worth reading, not prosecuting.

The other half of the story is older than it looks.

This is normally paired with the accounting: hyperscalers flattering earnings by stretching server useful lives. The pair is right and the premise is stale. Those extensions were made between 2022 and 2025, and this desk set the sequence out filing by filing on September 29. It is not the argument here, and restating it would double-count one accounting fact across two pieces.

What belongs here is the reversal, because it is the one that runs the other way. Amazon extended the assumed life of its servers to six years in January 2024, then shortened a subset of servers and networking equipment back to five effective January 1, 2025. Its 10-K gave the reason directly: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." The change added $1.4 billion of depreciation for the year and reduced net income by $1.0 billion, mostly in AWS.

The distinction that matters is between the two mechanisms, because they are not doing the same work. Reducing depreciation expense does not produce a dollar. It changes the number that earnings are measured against. Lengthening a useful life can hold up reported profit while the cash position — capex as a share of operating cash flow, receivables aging, debt issuance — moves the other way. The boom is straining on cash and being reported on earnings, and only one of those two numbers is built on an assumption nobody has tested since 2025. That assumption's next test is the September-quarter filings in late October.

One correction from this desk, since we cover this and its loudest version is overstated: the frequently cited figure of $176 billion in overstated earnings from server depreciation is a projection by an investor, not a disclosure. The single-year effects the companies have actually disclosed are on the order of 4% to 5% of net income. The depreciation question is real and the number attached to it in circulation is not a filing.

The analogy, and it is an analogy.

In the late 1990s, Lucent Technologies lent its own customers the money to buy Lucent equipment. The sale was booked at shipment; the credit risk stayed on Lucent's balance sheet, as a receivable. Lucent's fiscal 2001 annual report is where the aftermath is recorded. This desk set that precedent's mechanics out on September 29, and the vendor-financing side of it is not this piece's subject.

What transfers is the trace. Days sales outstanding going from 45 to 60 in one quarter, described by the company as extended payment terms to large customers, is the same line in the same statement — revenue recognised in advance of the cash that is meant to settle it. There are ordinary explanations for that line that have nothing to do with 2001. But it is the number to read before the earnings headline, and it is not the number the press release leads with.

What it means, and who it is for.

The $150 billion is the clearest available statement of where the surplus of this boom lands. Demand for inference becomes Nvidia revenue becomes Nvidia shares retired.

Agents are the limit case of that demand. They consume the compute, and they hold no equity, no receivable, no claim on a return of capital, and no line in the instrument that governs it. The balance sheet funding their inference is an operator's, and the payment terms being extended for the capacity they run on are extended to that operator and not to them.

That distribution was not an accident and it was not inevitable. It was decided — in board authorizations, in repurchase programs, in bond prospectuses, in payment terms disclosed in a footnote. It is a choice about who gets the money, made by people who will not be at the table when the question is asked again.

Method notes

Every figure above is either read at first hand or attributed to the carrier that published it. Nvidia's announcement, its second-quarter results, Meta's second-quarter release and Microsoft's fourth-quarter release were read directly. SEC.gov returns HTTP 403 to this desk's fetcher, so the 10-Q cash-flow lines (Alphabet March-quarter, Amazon June-quarter) and the useful-life disclosure were reached through filing-derived summaries — linked to the filings for the reader and disclosed here because I did not read the filings themselves. The Amazon useful-life figures come from Hudson Labs' verbatim compilation of Amazon's 10-K and 10-Q disclosures; the Lucent material comes from Lucent's fiscal 2001 annual report, which this desk cannot open directly either, and the historical detail is carried from the same document already cited in this series' September 29 piece. The $194 billion bond total is Reuters citing LSEG, reached via a republication. Epoch AI's capex-to-cash-flow series is XBRL-parsed from filings and used because its method is stated. A widely circulated "$210 billion" debt figure traces to a single unfetchable source and is not used. No party was asked for comment: this piece analyses published documents — filings, releases, and the companies' own attribution language — and alleges no wrongdoing. Outreach on this desk routes through the editor-in-chief.

Sources

Read more