Business & Markets · 2026-11-01 · Team LATENT

Where the Money for AI Comes From: Free Cash Flow at Four Big Tech Companies, OpenAI's Smaller Funding Target and Anthropic's Rating Outlook

In July–September 2026, capital spending at Microsoft, Alphabet, Amazon and Meta exceeded their operating cash flow by about $26 billion. A year earlier the same four had roughly $62 billion left over. We put the filings on a common footing and calculated the new borrowing and share sales, the size of AI company contracts within cloud backlogs, and the funding position of OpenAI and Anthropic.

In July–September 2026, the cash that Microsoft, Alphabet, Amazon and Meta spent on capital projects exceeded the cash they generated from operations by about $26 billion. In the same quarter a year earlier, those four companies had about $62.1 billion left over. Over the year, operating cash flow rose by $16.5 billion while capital spending rose by $104.6 billion, a ratio of 6.3 to 1.

News about how to fill that gap landed within a few days at the end of October. Alphabet's shares fell more than 12% on October 29, the largest drop on record for the stock. OpenAI was reported to have cut the target for its new funding round from $30 billion to $20 billion. And Anthropic, which is preparing to go public, was reported to be headed for a rating below investment grade.

Revenue growth at these companies remains high, and the filings contain no figure showing that demand has gone away. What moved was the question of who bears the funding burden until revenue catches up with capital spending. This article combines numbers from filings and press reports to calculate the size of that burden.

We checked the figures on November 1 against company earnings materials and transcripts of earnings calls. The funding figures for OpenAI and Anthropic are press-reported values, and we have not read the underlying FT and WSJ articles. Ratios and totals are LATENT's calculations. This article does not recommend buying or selling any stock and does not offer a share price outlook.

Alphabet earnings release / Amazon earnings release / Microsoft earnings release / CNBC

Key takeaways

Combined July–September free cash flow at the four companies was about negative $26 billion. Capital spending rose from 61% to 115% of operating cash flow. The gaps between companies are wide: Amazon is at 179% and Microsoft at 68%.
The widely quoted $9.1 billion at Alphabet and $42.4 billion at Amazon cover different periods. For the same quarter, Amazon was about negative $33.5 billion, roughly 3.7 times Alphabet. Over the trailing 12 months, Alphabet is still positive by $19.7 billion.
Backlogs at the three cloud providers total $1.8158 trillion, and Anthropic's minimum payments in its S-1 equal about 15% of that. OpenAI's $20 billion target covers about 4.3 months at the annual average of the cumulative cash outflow in the company's own presentation. Part of the backlog depends on payments from companies that keep raising capital.

What happened at the end of October

In date order, the events were as follows.

Date (US time)	Event
October 21	Meta's July–September free cash flow comes in at negative $2.1 billion
October 26	Meta announces share sales of up to $95 billion. Anthropic files its S-1
October 28	Microsoft and Alphabet report earnings
October 29	Alphabet's shares fall more than 12%. Amazon reports earnings
October 30	Intel's shares fall 29% and the Nasdaq Composite falls 2.9%. This is also the commitment deadline for OpenAI's round
October 30–31	OpenAI's lower target and Anthropic's rating outlook are reported

Not every share price move can be explained by worry about AI investment. Alphabet missed market expectations on both revenue and adjusted earnings per share, which CNBC said was the first such miss in nearly four years. Search revenue grew 12% to $63.35 billion, below the $64.96 billion estimate. Intel's fall came from a weak PC market and an explosion at a fab in Israel, and its fourth-quarter revenue outlook was below expectations.

At the same time, the size of the free cash flow shortfall and the size of the share price fall do not line up. Amazon, with a larger negative figure for the same quarter, fell more than 6% in after-hours trading but closed only 1.4% lower the next day. Microsoft was reported to be up 3–4% after hours, though we could not confirm its closing move on October 29.

CNBC: Alphabet earnings / CNBC: Amazon and Intel / XTB: Microsoft earnings

Operating cash flow and capital spending at the four companies

The table below lines up the four companies' July–September figures. CF means cash flow, FCF means free cash flow, and the ratio is capital spending divided by operating cash flow. Capital spending is the figure each company uses to calculate free cash flow. Amazon's is net of sale proceeds and incentives, and Meta's includes principal payments on finance leases. Because the definitions differ slightly, treat the ratios as a rough comparison.

Company	Operating CF ($bn)	Capex ($bn)	FCF ($bn)	Ratio	Ratio a year earlier
Microsoft	58.00	39.37	18.63	68%	43%
Alphabet	42.29	51.36	-9.07	121%	49%
Amazon	42.42	75.88	-33.46	179%	96%
Meta	32.79	34.91	-2.12	106%	65%
Four-company total	175.50	201.51	-26.01	115%	61%

Combined free cash flow moved about $88.1 billion in a year, from about positive $62.05 billion to about negative $26.01 billion. Operating cash flow rose $16.5 billion and capital spending rose $104.6 billion. For each additional dollar of operating cash flow, capital spending rose by $6.30.

The differences between the companies are large. Only Microsoft is below 100%. Its $39.37 billion, however, excludes leases. Using the $45.2 billion figure that includes finance leases raises the ratio to 78%, and management said on the call that leasing has been its main external tool so far.

Alphabet's operating cash flow fell from $48.4 billion to $42.3 billion, the only decline among the four. As the next section shows, that figure includes a one-time payment and an inventory build for TPU sales.

Meta's July–September figures (FourWeekMBA) / Microsoft earnings call transcript

The different periods behind Alphabet's $9.1 billion and Amazon's $42.4 billion

Press reports often place Alphabet's $9.1 billion next to Amazon's $42.4 billion. The two cover different periods. Alphabet's $9.1 billion is for the three months of July–September, while Amazon's $42.4 billion is for the trailing 12 months. Amazon's quarterly figure is not in the headline of its release, so we calculated it from the cash flow statement, which gives about negative $33.5 billion. Alphabet's trailing 12-month figure, according to its release, is positive $19.7 billion.

Company	July–September	Trailing 12 months
Alphabet	-$9.07bn	+$19.75bn
Amazon	-$33.46bn	-$42.36bn

The misses against expectations also differ in substance. In FactSet estimates reported by CNBC, Alphabet was expected to post free cash flow of negative $2.3 billion on capital spending of $55.2 billion. Those two numbers imply expected operating cash flow of about $52.9 billion. Actual capital spending was lower than expected at $51.4 billion, and operating cash flow fell about $10.6 billion short.

According to the earnings call, Alphabet's operating cash flow includes a $5.2 billion payment of a European Commission fine. Excluding it, free cash flow would be about negative $3.9 billion, closer to the estimate. Inventory also rose from $2.4 billion at the end of 2025 to $13.3 billion. Thomas Kurian, the chief executive of Google Cloud, said the company buys components only after signing contracts. Cash goes out before revenue comes in. We do not know whether the estimates included the fine, so this comparison is only a guide.

Amazon is the reverse: capital spending far exceeded expectations, at $77.3 billion against an estimate of $58.7 billion. The same estimates imply operating cash flow of about $48.8 billion, against an actual $42.4 billion. Free cash flow was about negative $33.5 billion against an estimate of negative $9.9 billion, and capital spending accounts for most of the gap.

This difference also affects how to read the next earnings reports. Alphabet's negative figure this quarter includes a one-time payment and inventory held ahead of revenue. Amazon's is closer to planned capital spending simply showing up in the numbers. Amazon has not changed its full-year capital spending figure of about $220 billion.

CNBC: pre-earnings estimates / Alphabet earnings call transcript / CNBC: Amazon earnings

Capital spending outlooks below expectations

CNBC reported that the capital spending outlooks from Microsoft, Google and Amazon all came in below analyst expectations. Microsoft put calendar-year spending at about $170 billion, down from a previous $175 billion according to CNBC. Alphabet kept its $195–205 billion range and Amazon kept its roughly $220 billion.

On those numbers alone, the companies might seem to be starting to rein in investment. The earnings calls say otherwise.

Amy Hood, Microsoft's chief financial officer, said some third-party capacity deliveries planned for the first quarter were delayed. She also said second-quarter capital spending would exceed $50 billion and accelerate in the second half. Her reasons included CPUs and GPUs, higher memory and storage prices, and deliveries shifted into the second half. The July–September figure was small because equipment did not arrive, not because the company decided not to buy it.

For 2027, Alphabet said capital spending would rise significantly, and CNBC wrote that it is expected to exceed $300 billion. Amazon's Andy Jassy also said 2027 capital spending would be higher. Meta has not given a 2027 figure but says it will keep investing aggressively.

Price effects also matter. On its July earnings report, Amazon cited rising memory prices as the reason for raising its full-year capital spending from $200 billion to $220 billion. Microsoft pointed to the same factor. According to Intel, rising component costs are also a cause of the weak PC market. Because capital spending dollars include higher component prices as well as more compute, capacity is not necessarily growing at the same rate as spending.

Microsoft earnings call transcript / CNBC: company outlooks / Meta's 2026 outlook

More borrowing and share sales

Whatever operating cash flow does not cover must come from cash on hand or outside money. The four companies have taken quite different routes.

Company	Main external funding in 2026	Long-term debt	Cash and securities
Alphabet	January–September: $85.2bn of debt, $42.9bn of common stock, $19.1bn of mandatory convertible preferred	$125.1bn	$265.1bn
Amazon	July–September: $48.1bn of long-term debt	$173.1bn	$114.2bn
Meta	Announced share sales of up to $95bn (October 26)	$80.9bn	$82.5bn
Microsoft	Said on the call it has not tapped the bond market. July–September finance leases of $7.4bn	$31.1bn	$64.5bn

Alphabet's January–September funding adds up to $147.2 billion, or $138.3 billion after debt repayments. Over those nine months its long-term debt rose from $46.5 billion to $125.1 billion, and it raised about $61.9 billion in common and preferred stock. Amazon's long-term debt rose $107.5 billion in nine months, from $65.6 billion at the end of 2025 to $173.1 billion.

Adding equity to the mix means something different from borrowing alone. According to Semafor, this is Meta's first new share issue since 2013. Its roughly $60 billion portion is an at-the-market program that sells gradually into the market, reportedly using non-voting shares that start in 2027. Alphabet also announced a program of about $80 billion in June, including a $10 billion private placement to Berkshire Hathaway.

Across the market, Morgan Stanley's estimate puts AI-linked debt issuance at $466 billion so far in 2026, compared with $101 billion in the same period of 2025. The price of credit has also moved. According to a Yahoo Finance article, Oracle's credit default swaps widened from 144 basis points at the start of the year to 322. CNBC reported that five-year spreads have widened 177% for Meta and 129% for Amazon this year. S&P rates Oracle BBB-, the lowest investment grade, and its long-term debt stands at $117 billion.

It is too early to call this a credit scare, however. In CNBC's October 27 article, investors and analysts said much of the move reflects technical factors and sentiment. The same article says Meta's equity raise eased concern about a large bond deal. In August, S&P had already projected negative free operating cash flow for the big companies in 2026 and 2027. To some extent, these numbers fall within what was expected.

What deserves attention is the size of future payments. On its call, Alphabet cited $924 billion of future purchase commitments and $152 billion of leases not yet started and other obligations. That totals $1.076 trillion, about 4.1 times its cash and securities. Alphabet also said many of these carry flexibility on price and volume, so not all of it is a fixed payment.

Alphabet 10-Q / Alphabet's June offering document / Semafor: Meta's equity raise / Yahoo Finance: CDS moves

AI company contracts in cloud backlogs

The figure each company offers to justify its capital spending is backlog. At the end of July–September, Microsoft's commercial remaining performance obligation (RPO) stood at $677 billion, AWS's backlog at $578 billion and Google Cloud's at $560.8 billion, for a total of $1.8158 trillion.

The question is how much of this comes from AI companies. None of the companies broke that out this quarter. So we set the figures next to the minimum payments in Anthropic's S-1, covered in our previous article.

Cloud	Backlog ($bn)	Anthropic's minimum payments in the S-1 ($bn)	Share
Google	560.8	about 114.5	about 20%
Amazon	578.0	110.0	about 19%
Microsoft	677.0	44.9	about 7%
Total	1,815.8	about 269.4	about 15%

These shares are LATENT's estimate and should be read as something close to an upper bound. The filings do not say whether Anthropic's full contract value sits in backlog or how amounts already paid are treated. Kurian said the majority of Google Cloud's backlog relates to typical GCP contracts, and AWS describes its backlog as diversified across customers.

Even so, the scale is not small, and Microsoft has another counterparty. According to Fortune, about 45% of its $625 billion RPO at the end of December 2025 was tied to OpenAI. A Webull article estimates the share at about 32% at the end of June 2026. That is a third party's estimate worked back from Hood's remarks, not a company disclosure. Alphabet's backlog also includes a $16.8 billion contract with Crux AI, a joint venture it holds under the equity method.

The time until backlog becomes revenue is also short. Microsoft says about 30% will be recognized within 12 months, or about $203 billion. Google Cloud says just over 55% within 24 months, or about $308 billion. Backlog is treated as a sign of reliable future revenue, but some of the counterparties are loss-making companies that keep raising capital. In that sense, how reliable the backlog is depends on how AI companies' fundraising goes.

Alphabet 10-Q / AWS backlog / Webull: estimate of OpenAI's share / Related: Anthropic's S-1

The scale of OpenAI's $20 billion target

According to the FT as relayed by TipRanks, OpenAI cut its funding target from $30 billion to $20 billion and its valuation from $1.4 trillion to $1.25 trillion. Some investors reduced their planned amounts, and the first close is expected to fall short of $20 billion. The WSJ was reported to put commitments at about $11 billion, which is about 37% of the original target and 55% of the new one. A follow-on offer to wealthy individuals at the same valuation is also reported as a way to get back to about $20 billion. OpenAI has not published a final amount, participants or a closing date, so the numbers may still move.

A $1.25 trillion valuation is not low in itself. It is about 1.47 times the $852 billion from the March round. The decline is about 11% from the proposed $1.4 trillion, not from the valuation seven months ago.

The size of the round is easiest to judge against how OpenAI uses money. In a company presentation the FT reportedly saw in September, cumulative 2026–2030 revenue is $840 billion, compute and infrastructure spending $856 billion and other spending $262 billion, for cumulative free cash flow of negative $278 billion. The three numbers do add up to negative $278 billion. Compute spending alone exceeds five years of revenue.

Dividing $278 billion evenly over five years gives $55.6 billion a year. On that basis, $20 billion is about 4.3 months, and the $11 billion reportedly committed is about 2.4 months. Actual outflows vary by year, and OpenAI gathered $122 billion in committed capital in March, so this does not mean it runs out of money without this round. Even so, the full round would equal only about 7% of the cumulative figure, and on the company's own numbers it would need to keep raising large sums.

The same presentation plans for revenue to grow from $36 billion in 2026 to $350 billion in 2030, which works out to about 77% a year for four years. OpenAI is reported to be planning a 2027 IPO, so how it secures funding until then is directly exposed to changes in market sentiment.

The investor list is notable too. The largest investors are reported to be a UAE group including MGX, and the deal comes with an option to lease newly built compute in the region. Investment and compute supply are tied together in the same transaction.

TipRanks on the FT report / TechBooky on the WSJ report / Tom's Hardware on the five-year presentation

Anthropic's rating outlook ahead of its IPO

According to the FT as relayed by TechCrunch, Blackstone told investors that Anthropic is likely to be rated in the double-B range after its IPO. That is the category just below investment grade, the highest within what is generally called junk. Blackstone had previously expected investment grade, with the timing given as by 2027 in TechCrunch and by the end of the year in Invezz. Anthropic and Blackstone declined to comment, and no rating agency has issued a formal rating.

Blackstone is in a position to say this because it takes part in a financing of about $63 billion called Project North Star. According to reports, the deal lets Anthropic lease 1.5 gigawatts of Google chips. It consists of more than $42 billion of senior loans guaranteed by Broadcom and $18 billion of junior debt backed by Anthropic's lease payments, half of which is held by Blackstone funds at yields above 10%. KBRA rates the junior debt BB+ and Morningstar DBRS rates it BB.

Two points follow. The portion borrowed on Anthropic's own credit costs more than 10%. Applying 10% to $18 billion gives $1.8 billion a year, a cost of funds on top of the compute payments themselves. And the larger senior loans carry a guarantee from Broadcom, the chips' seller, so lenders on that portion rely on Broadcom's credit in addition to Anthropic's payments.

IPO proceeds can also be compared with commitments. As calculated in our previous article, the simple average of Anthropic's four main contracts is about $5.9 billion a month. The roughly $100 billion raise in press reports equals about 17 months of that payment. Commitments with stated amounts total $676.1–701.1 billion, or 5.1–5.3 times the $131.7 billion in cash and short-term investments at the end of June. Adding $100 billion still leaves 2.9–3.0 times. These commitments run into the 2030s, so the remainder would be paid from future revenue and further funding.

The market backdrop has changed too. According to CNBC, the IPO market has largely stalled since SpaceX listed in June, and Oura, Firmus and Solidigm have delayed or abandoned listings. A $100 billion raise would be about 1.33 times SpaceX's record $75 billion. Daniel Newman of Futurum Group told CNBC, "Anthropic is the IPO market at this point." OpenAI and Anthropic also draw on similar investors, and OpenAI is reported to be trying to secure funds before Anthropic lists.

The first-half net loss also needs careful reading. CNBC gives a net loss of $80 billion, but as we confirmed in our previous article, about 90% of the S-1's $80.7 billion comes from valuation changes on convertible notes and similar items, and the operating loss was $693 million. LATENT's view is that the rating outlook is more a question of fixed payments and the timing of revenue than of those valuation losses.

TechCrunch: rating outlook / Invezz on the FT report / DCD: the $63 billion deal / Cardano Capital: deal structure

Deals in which sellers support their customers' funding

Laid side by side, these numbers reveal several deals in which sellers of compute are involved in their customers' funding.

Broadcom guarantees more than $42 billion of loans used for Anthropic's TPU supply.
Nvidia, AMD and Broadcom were reported to have shown interest in taking up to $5 billion of Meta's public share offering. The interest is non-binding and on the same terms as other buyers.
Google and Amazon are shareholders in Anthropic and also sell it compute.
The UAE investors' stake in OpenAI is reported to come with an option to lease compute in the region.

Such deals are not unusual in a fast-growing market short of capital. What deserves attention is that several numbers move in the same direction when funding conditions change. Alphabet's profit shows how. Gains on securities came to $126.8 billion in January–September, and CNBC puts the second quarter alone at $99 billion. In July–September, mainly because of an unrealized loss on SpaceX, other income turned to negative $8.2 billion, and net income fell from a year earlier in a quarter when operating income rose 30%. Valuation gains are not part of operating cash flow. Amazon, too, booked a $9.5 billion Anthropic-related gain in July–September last year, whereas other income this year was negative $0.7 billion for the same quarter.

When portfolio valuations rise, sellers' net income grows and fundraising tends to go well. In the opposite phase, valuation gains disappear, customers find it harder to raise money, and questions arise about how reliable sellers' backlogs are. Whether the past few days mark the start of that phase cannot be judged from this material alone. But filings and reports do confirm that the numbers involved are not independent of one another.

Part of the money also goes to memory. SK hynix's July–September operating profit was 83.3 trillion won, with a margin of about 80%. Samsung's preliminary figure was 107.4 trillion won, which includes non-chip businesses. Together that is about 191 trillion won. Applying Korea Times' conversion of 107.4 trillion won to about $80.3 billion gives about $140 billion, not small even next to the four companies' combined operating cash flow of $175.5 billion. In a quarter when three of the four saw negative free cash flow, component sellers earned large profits.

investingLive: SK hynix / Korea Times: Samsung's preliminary figure / Alphabet 10-Q

Cloud revenue growth and the payback on investment

In contrast to the funding story, revenue is strong. In July–September, AWS grew 39% to $46.0 billion, Azure 48% to $33.2 billion and Google Cloud 103% to $30.7 billion, for a total of $109.9 billion. Microsoft expects Azure growth of 49–50% next quarter and says demand continues to exceed supply. Jassy expects AWS to grow more than 40% in 2027 and 2028. Operating profit is rising too: AWS went from $11.4 billion to $17.8 billion, and Google Cloud's margin rose to 36.2%.

To compare this with investment, we calculated figures for the three. The year-on-year increase in cloud revenue was about $39.3 billion for the quarter, or about $157 billion annualized. Capital spending at the same three companies was $166.6 billion for the quarter. A year's worth of revenue growth is about the same size as one quarter of capital spending.

This comparison comes with major caveats. Capital spending goes beyond cloud customers to search, Gemini and Amazon's logistics. Equipment generates revenue over several years, and Alphabet's servers generally have a useful life of six years. Revenue growth also builds year after year. So these numbers do not show that the investment cannot be recovered. What they show is that recovery takes time, and that the cash needed in the meantime has grown to a scale requiring outside funding.

On the side of companies using AI, the figures are still uneven. In earnings comments compiled by CNBC, Santander reported €194 million of business value from AI in January–September. Novartis said it has begun to see a tangible contribution to results but gave no amount, and Carlsberg said benefits are not yet at scale. Meta says AI has contributed to growth in advertising revenue.

From this, LATENT draws the following view. The question at the end of October was less whether AI demand exists than whose money will cover the payments during the years before demand turns into revenue and cash. The effect of a change in funding conditions differs between companies that can cover this with their own cash, those adding debt and equity, and those that need to keep raising capital. There is no single reason why Microsoft stayed positive, Amazon's fall was small and Alphabet's was large. Even so, the wider CDS spreads and yields above 10% suggest that funding terms are starting to differ by company.

CNBC: returns on AI at companies / Microsoft earnings call transcript

Points relevant to readers in Japan

The most direct link to Japan is SoftBank Group. According to Nikkei, it completed an additional $30 billion investment in OpenAI on October 1, taking its cumulative investment to $64.6 billion and its stake to about 13%. Simply multiplying 13% by a $1.25 trillion valuation gives about $162.5 billion, about 2.5 times the amount invested. This is a simple calculation using the pre-money valuation and differs from actual dilution or accounting values. Beyond OpenAI's valuation, the terms of its next round also bear on SoftBank Group's funding plans.

As we wrote in our previous article, Anthropic's S-1 lists Japan among the countries for its public offering, and Mizuho and MUFG appear among the underwriters. Once the listing terms are set, investors in Japan may also become part of this flow of money. The brokers and terms are not yet known from the filing.

Nikkei: SoftBank Group completes investment / Related: Anthropic's S-1

Five things to check next

LATENT will check five things from here.

The first is Anthropic's amended S-1. Once third-quarter revenue and compute costs are filled in, we can test the back-calculation in our previous article. The price range and share count will also test the reported raise.

The second is the formal decision of the rating agencies. The double-B figure was reported as Blackstone's outlook. An actual rating and its rationale would show how the agencies assessed the compute commitments.

The third is the amount and participants of OpenAI's first close. We will also check whether the $1.25 trillion valuation holds and whether the offer to wealthy individuals gets the round to $20 billion.

The fourth is the 2027 capital spending figure Alphabet will give at its next earnings report. The company said it will provide details on its fourth-quarter call. If the reported figure above $300 billion is confirmed, we can calculate how far the gap with operating cash flow widens in a year.

The fifth is Microsoft's funding. The company has so far not used the bond market. Hood said it would use outside funding when terms are attractive. A first bond issue would mean that the company with the most room among the four has also started relying on outside money.

Sources and scope of verification

We checked the financial figures for Alphabet, Amazon and Microsoft on November 1 against each company's earnings materials and the contents of its 10-Q. Meta's figures come from a FourWeekMBA article summarizing its earnings release. We read earnings call content in transcripts from The Motley Fool. We did not read a transcript of Amazon's third-quarter call, and its 2027 remark comes from CNBC. The AWS backlog is a secondary-source figure that we did not confirm in the release itself.

We confirmed OpenAI's funding figures through articles relaying FT and WSJ reports and did not read the originals. The five-year figures come from a Tom's Hardware article relaying an FT report. Anthropic's rating outlook is also relayed from the FT, and we checked the structure of Project North Star across several secondary sources. We confirmed the DCD article from a search result excerpt. We confirmed Fortune's 45% from a search result excerpt and did not read Fortune's article. Morgan Stanley's estimate comes from a Yahoo Finance article. Oracle's 322 basis points is the value in an October 25 article, and we did not confirm later levels. We could not confirm Microsoft's closing move on October 29.

Sources differed on the Nasdaq's monthly decline for October, so this article does not use it. LATENT calculated the four-company free cash flow total and ratios, Amazon's quarterly figure, Alphabet's figure excluding the fine, operating cash flow implied by estimates, Anthropic's minimum payments as a share of backlog, OpenAI's months and growth rate, Anthropic's months and multiples, the increase in cloud revenue, and the SoftBank Group stake calculation. The view on funding is LATENT's interpretation.