Record Data Sheet

Amount
$231.5B vs $198.6B
Entity
Amazon, Alphabet, Meta, Microsoft, Oracle
Vertical
Tech · AI infrastructure
Record Type
Capex vs operating cash flow, Q3 2026
Previous Record
Q3 2025: $97.3B of capex at Amazon, Alphabet, Microsoft and Meta
Held By
Amazon: $77.3B of capex in one quarter
Period
July–September 2026 (Oracle: June–August 2026)
Confirmed
October–November 2026 (Q3 results and SEC filings)

Amazon, Alphabet, Meta and Oracle all reported negative free cash flow for the third quarter of 2026. Microsoft was the exception. Together, the five spent $231.5 billion on capital expenditure against $198.6 billion of operating cash flow. The $32.9 billion gap was filled with bonds, new shares and leases that do not yet appear on any balance sheet.

This is not a solvency story. These are among the most profitable companies in history. It is a story about who now carries the risk of the AI buildout, and what that does to the price of capital for everyone else.

THE NUMBERS

Q3 2026 at a glance
$231.5B
Capex, five AI builders, July–September 2026
$198.6B
Operating cash flow, same quarter
1 of 5
Companies with positive free cash flow (Microsoft)
$115.1B
Long-term debt issued by Amazon in nine months
$25B
Microsoft bond sale, its first since 2017
7.50%
Coupon on Meta’s new 30-year bond
$820B
Data-center leases not yet on balance sheets (Moody’s)
$1.3T
S&P 2027 capex forecast, six hyperscalers
Capex for Microsoft, Alphabet and Amazon is cash purchases of property and equipment; Meta’s includes finance-lease principal. Oracle’s quarter is June–August 2026.

The cash gap

The cash gap
Four of five AI builders spent more than their operations generated
Capex vs operating cash flow, Q3 2026, $ billions. Hover a row for the ratio.
Operating cash flowCapexFree cash flow
AmazonJul–Sep 2026
42.4
77.3
−$34.9B
OracleJun–Aug 2026
23.1
28.5
−$5.4B
AlphabetJul–Sep 2026
42.3
51.4
−$9.1B
MetaJul–Sep 2026
32.8
34.9
−$2.1B
MicrosoftJul–Sep 2026
58.0
39.4
+$18.6B
Sources: company results releases and SEC filings. Free cash flow = operating cash flow minus capex; Amazon’s single-quarter figure is our arithmetic. Combined: $231.5B spent, $198.6B generated.

Amazon is the outlier. It spent $77.3 billion in one quarter, 1.8 times what its operations produced, and its trailing twelve-month free cash flow swung to an outflow of $42.4 billion, from an inflow of $14.8 billion a year earlier. It happened in the quarter with the highest operating margin in Amazon’s history.

Alphabet’s free cash flow was negative for the second quarter running, the first negative quarters in its 22 years as a listed company. Meta’s was negative for the first time since its 2012 IPO; it bought back no shares in the quarter, against $3.3 billion a year earlier. Microsoft kept $18.6 billion after spending $39.4 billion, partly because it leases more of its capacity than it builds.

The pressure grows from here. Meta’s full-year guide of $135–145 billion implies $49–59 billion of capex in the fourth quarter. Microsoft says its October–December capex will exceed $50 billion. Alphabet’s CFO says spending will “increase significantly in 2027”. Our earlier analysis of the Microsoft–Alphabet spending race covers how these budgets were set.

The bond wave: $67 billion in eight days

Between 29 October and 5 November, Microsoft, Meta, Oracle and Alphabet raised about $67 billion in bonds, counting the euro deals at issue-date exchange rates. Each said the money was for “general corporate purposes”. Each had just reported capex above, or close to, its operating cash.

The bond wave
About $67 billion of Big Tech bonds in eight days
29 October–5 November 2026. Click a column header to sort.
IssuerWhat was soldSize ($B)CouponDemand
Microsoft$25B of notes, 10 tranches, due 2028–2066
Priced Oct 29
25.04.85–6.35%
~$115B (4.6x)
Meta$15B of notes, 5 tranches, due 2031–2056
Settled Nov 5
15.05.85–7.50%
~3.9x
Meta€9B, its first euro bonds
Early Nov
10.4n/d>€62B (6.9x)
Oracle$10B of notes, plus $10B of new shares
Filed Nov 2
10.0n/dn/d
Alphabet€6B of euro bonds (~$6.7B)
Nov 2
6.7n/dn/d
Euro amounts converted at rates reported at issue. Coupon bars run on a 4%–8% scale. Sources: prospectus supplements and Bloomberg/Reuters reporting.

Three details matter more than the totals.

Microsoft had not sold US investment-grade bonds since 2017. Its last deal was $17 billion that January. The new issue takes its total debt from $35.5 billion to roughly $60.5 billion, according to the CFO’s figures and the filing. The company with the strongest cash position in the group chose to borrow anyway.

Meta is paying more, not less, as it borrows more. Its new 2056 bond carries 7.50%. Its 2055 bond, sold in November 2025, carries 5.625%: about 190 basis points less for almost the same maturity. Meta’s five-year credit default swaps hit record highs after its results. Moody’s still rates the notes Aa3.

The bonds are only one layer. In October Meta also priced $25 billion of forward stock sales and $12.5 billion of mandatory convertible preferred shares, plus a program to sell up to $60 billion more. Wells Fargo estimates Meta needs $60–80 billion of financing over the next 12 months. Amazon issued $48.1 billion of long-term debt in the third quarter alone and $115.1 billion over nine months; its long-term debt went from $65.6 billion to $173.1 billion. Alphabet covered its quarter with $29 billion of new notes and $12.4 billion of new stock.

The market is absorbing it. Order books of four to seven times the deal size show investors want this paper. But the coupons show what they want in return.

The debt you don’t see

Reported debt understates the commitment. Moody’s counts about $460 billion of outstanding hyperscaler debt, double the end-2024 level. Data-center leases signed by Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave add up to $1.2 trillion, and $820 billion of that is not yet recognized on their balance sheets.

Three mechanisms keep it there.

  1. Leases not yet commenced. A signed lease for a building still under construction is disclosed in a footnote, not booked as a liability. Oracle alone disclosed $288 billion of such commitments, almost all data centers, at 31 August. S&P notes Microsoft has more than $329 billion of future lease obligations and Meta $279 billion.
  2. Special purpose vehicles. Meta’s Hyperion campus in Louisiana is owned by a joint venture in which Blue Owl funds hold 80%. The vehicle, Beignet Investor, borrowed about $27.3 billion. Meta leases the campus and does not consolidate the debt. A second deal for El Paso, with BlackRock, had to offer yields above 7%.
  3. Residual value guarantees. Landlords will only build if the tenant promises to cover any fall in the building’s value at the end of a short lease. Moody’s has warned these guarantees can hide a material liability.

None of this is hidden in the legal sense. All of it is disclosed. But a CFO comparing leverage across these companies is comparing different things. Meta routes much of its build through vehicles. Oracle borrows directly. Their headline debt numbers cannot be read side by side.

There is a second accounting effect. Servers are typically depreciated over five to six years. If AI chips lose economic value faster, today’s operating profits are overstated and future write-downs are larger. That is a judgment, not a disclosure, and it is where the next argument about these accounts will happen.

2027: who can pay for it

2027 outlook
S&P expects only Microsoft to generate free cash flow in 2027
Projected 2027 free cash flow and capex, $ billions. Six companies: $1,321B of capex, a $268.7B net cash shortfall.
2027 capex
Microsoft
+$33.6B
$189B
Meta
−$3.5B
$164B
Oracle
−$41.6B
$95B
Amazon
−$60.1B
$319.1B
Alphabet
−$82.7B
$357B
SpaceX
−$114.4B
$197.2B
Source: S&P Global projections, as reported by The Motley Fool, September 2026. SpaceX is included by S&P after its merger with xAI.

S&P expects the six largest AI builders to spend $1.3 trillion in 2027, up from about $870 billion this year. On its numbers, only Microsoft generates free cash. Alphabet and Amazon, the two biggest spenders, would each need $60–83 billion of outside money in a single year.

Forecasts diverge, and that is the point. S&P puts Meta at a $3.5 billion deficit; sell-side consensus puts it closer to $35 billion. Morgan Stanley’s broader count of hyperscalers and AI clouds reaches $1.6 trillion for 2027. Whatever the exact figure, the funding gap is measured in hundreds of billions, and it has to come from lenders, shareholders or landlords. For the full bill to the end of the decade, see the $31.6 trillion AI infrastructure bill.

Ranked by their ability to absorb it, the order is clear. Microsoft funds itself and keeps its AAA rating. Alphabet and Amazon have the cash and margins to borrow cheaply. Meta has the cash flow but is paying a rising price for debt. Oracle, with debt above $100 billion and a rating two notches above junk, has the least room.

What it means if you don’t run a hyperscaler

The AI buildout used to be financed by the biggest cash machines in the world. It is now financed by the same bond market everyone else uses. That has four practical consequences.

  1. Competition for long-dated money. Hyperscalers issued about $182 billion of senior notes in 2026, against $93.3 billion in 2025, much of it at 20 to 40 years. Corporate treasurers refinancing in 2027 will share the same investor base. Expect pricing to reflect it.
  2. Your cloud contract is part of their financing. Hyperscalers added about $700 billion of contracted backlog in two quarters, and that backlog is what lenders underwrite. Multi-year minimum commitments are worth more to the seller than ever. That is leverage at the negotiating table, for the buyer.
  3. Capacity is the constraint, not price. Morgan Stanley expects a shortage of more than 30% in powered data-center capacity through 2028. Spot prices for top GPUs show it: Ornn’s B200 rental index rose from $5.66 to $9.94 per GPU-hour between August and November. Budgets for AI projects should assume compute gets dearer before it gets cheaper. The power side of the same problem is in who pays for the electricity behind AI.
  4. Counterparty risk is uneven. Microsoft is rated AAA. Oracle sits two notches above junk with negative outlooks. A ten-year AI or cloud dependency deserves the same credit review as any other long-term supplier.

Verdict

The question is no longer whether Big Tech can afford AI. It is whether AI revenue arrives before the next refinancing does. In the third quarter, only Microsoft funded its buildout from its own operations. Everyone else borrowed, sold shares, or signed leases that will land on the balance sheet later.

That is a normal way to build infrastructure. Railways, telecoms and utilities were all built on debt. But it changes the nature of these companies. They used to be valued as software businesses with spare cash. They are now capital-intensive, partly leveraged operators. Investors, lenders and customers should read their accounts that way.

The number to watch is not 2027 capex. It is the gap between capex and operating cash flow, quarter by quarter, and how much of it lands off the balance sheet.

Frequently asked questions

Which Big Tech companies have negative free cash flow?

In July–September 2026, Amazon, Alphabet and Meta all reported negative free cash flow, as did Oracle in June–August 2026. Microsoft was the only one of the five with positive free cash flow, at $18.6 billion.

How much are hyperscalers borrowing for AI?

About $182 billion of senior notes in 2026 so far, nearly double 2025, plus equity sales, convertible preferred shares and off-balance-sheet vehicles. Microsoft, Meta, Oracle and Alphabet sold about $67 billion of bonds between 29 October and 5 November alone.

What is off-balance-sheet AI debt?

Debt raised by vehicles the tech company does not consolidate, such as the $27.3 billion borrowed by Beignet Investor for Meta’s Hyperion campus, plus leases signed for data centers not yet built. Moody’s estimates $820 billion of such leases are not yet on balance sheets.

How much will Big Tech spend on AI in 2027?

S&P projects $1.3 trillion for six companies, including SpaceX. Morgan Stanley projects $1.6 trillion for hyperscalers and AI cloud providers combined.

Sources

Free cash flow is operating cash flow minus capital expenditure, as each company defines it. Amazon’s quarterly figure is our arithmetic; the company reports free cash flow on a trailing twelve-month basis. Quarters are calendar July–September 2026, except Oracle (June–August 2026). Figures for 2027 are third-party forecasts, not company guidance.

Poll

Who is taking the biggest risk with AI debt?