The AI cashflow pipeline: how Amazon, Microsoft and Meta route their cash straight into NVIDIA. A Dot-com CISCO style cash-flow mechanics
Four of the most profitable companies on Earth are running the same trade. Three of them are sending money. One is receiving it.
Something weird is happening to the Magnificent Seven
If you only look at top-line numbers for Amazon, Microsoft and Meta in 2025, the businesses look fine. Operating cash flow grew. Revenue grew. Operating income grew. By any normal measure, these are companies in great shape.
Then you look at free cash flow.
Amazon’s full-year 2025 free cash flow came in at $11.2 billion, down from $38.2 billion a year earlier. That is a roughly 71% drop in a single year for a company that just generated $139.5 billion in operating cash. The Q4 letter explicitly attributed the decline to a $50.7 billion year-over-year jump in property and equipment purchases, primarily for AI infrastructure. Microsoft is in a different fiscal year but the same regime: FY25 free cash flow at $71.6 billion, with quarterly capex jumping from $16.7B in Q3 FY25 to $30.9B in Q3 FY26 (the most recent quarter), an 84% year-over-year ramp. Meta delivered $43.6 billion of free cash flow in 2025 against $54.1 billion in 2024, with a $69.7 billion capex bill that the CFO has now guided up to $115-135 billion for 2026.
So why is operating cash flow rising while free cash flow is falling? Because the companies are taking the cash they generate and pouring it into one place: data center capacity. And the single biggest line item inside that data center capacity is NVIDIA’s GPUs.
This is the cashflow pipeline. And it is bigger and tighter than most investors realize.
The great divergence
If you plot the annual free cash flow of NVIDIA against Microsoft, Amazon and Meta side by side, the picture is almost cartoonish.
NVIDIA’s free cash flow went from $3.8 billion in FY23 to $96.7 billion in FY26. That is a 25x increase in three years. In the same window, Amazon’s free cash flow has done a full round trip - from -$16.9B in 2022 to $32.9B in 2024, and back down to $7.7B in 2025. Meta’s has plateaued around $44-54B. Microsoft’s has gone slightly backwards despite revenue growth, because capex is climbing faster than operating cash flow can support.
The green crossover at the right side of the chart matters. For the first time, the company selling the picks and shovels is generating more free cash flow than three of the four companies buying them. The exception is Microsoft, and it is now within touching distance.
The latest data point is even more extreme. NVIDIA reported Q1 FY27 results on May 20, 2026 - a single quarter - and free cash flow came in at $48.6 billion. For comparison, that is more than Meta’s entire 2025 year of $43.6 billion of FCF. In one quarter. From one company. Selling chips.
NVIDIA - the cash machine
Watching NVIDIA’s FCF curve fill in over time is genuinely hard to comprehend if you have not seen it animated. Five years ago, the company generated $3 to $8 billion of free cash flow per year. The latest quarter alone is roughly 12 times the entire FY23.
The market is now treating NVIDIA’s cash position as quasi-strategic. On May 18, 2026, the board approved an additional $80.0 billion of share repurchase authorization and bumped the quarterly dividend from $0.01 to $0.25 per share, a 25x increase. Total capital returned to shareholders in Q1 FY27 alone hit roughly $20 billion. That is the kind of number a company prints when it does not know what else to do with the cash.
The FY26 10-K disclosure also confirmed something that has been quietly looming in the background. Four direct customers each accounted for more than 10% of NVIDIA’s total revenue, collectively making up 61% of the business. Two of them alone made up 39% in Q2 FY26. The names are not formally disclosed but reporting and reasonable inference point at Microsoft, Amazon, Meta and Alphabet, with Oracle and xAI as the next tier. In short, the same handful of companies whose free cash flow is being squeezed are NVIDIA’s revenue.
This is what a cashflow pipeline looks like.
The mechanics of the pipeline
The cleanest way to see what is happening is to draw the system as it actually operates. On the left are the sources of cash. In the middle is the conduit - hyperscaler capex commitments for 2026. On the right is the sink, which currently happens to be one company.
Each of the four hyperscalers is independently generating $115-140 billion in operating cash flow per year. Each of them is then deciding to commit a very large share of that cash, plus additional debt, to AI infrastructure. The 2026 capex guides, taken from the earnings calls that took place in late January and early February 2026, then refined in late April with Q1 reports, total approximately $700 billion: $200B at Amazon, around $185-190B at Microsoft, $115-135B at Meta, and $175-190B at Alphabet. Tom’s Hardware and the Financial Times both reported that the combined number now sits at roughly $725 billion, up 77% from a record $410 billion in 2025.
To put $700 billion in context: it is more than the entire annual capital expenditure of the entire S&P 500 oil and gas sector at peak. It is also more than NVIDIA’s entire trailing twelve-month revenue, which means that even if NVIDIA captured every single dollar of this capex, it would still not absorb all of it.
NVIDIA does not capture all of it, of course. Not all hyperscaler capex is GPUs. There are land, buildings, power equipment, networking, cooling, custom chips, ARM CPUs, storage, optical interconnects, and increasingly satellite and robotics buildouts (Amazon’s Project Kuiper / Leo is now formally inside the capex line). But NVIDIA does capture a strikingly large share. In Q1 FY27, NVIDIA’s Data Center revenue was $75.2 billion. The total quarterly capex of the four hyperscalers was approximately $130 billion. NVIDIA’s Data Center segment alone is therefore capturing close to 58% of the quarterly capex of its largest customers.
That ratio is the pipeline.
The pipeline in motion
The dollar flow becomes obvious when you plot hyperscaler quarterly capex as a stacked area and overlay NVIDIA’s Data Center revenue as a line. They move together. They have moved together for two years. And NVIDIA’s slope is steeper.
In Q1 2025, the four hyperscalers spent $72 billion of capex. NVIDIA’s Data Center revenue that quarter was $39.1 billion. By Q1 2026, the hyperscalers’ combined capex had grown to $130 billion, and NVIDIA’s Data Center revenue had grown to $75.2 billion. The hyperscaler line grew 80%. The NVIDIA line grew 92%. NVIDIA is growing faster than its customers’ capex, which means it is taking share inside the capex bucket.
This is also why the recent CNBC takeaways on the May 20, 2026 earnings noted that NVIDIA’s Data Center segment grew 92% year over year, faster than the 75% growth in the prior quarter and 73% in the same period a year ago. The growth rate is accelerating, not slowing. As long as the hyperscalers keep raising capex guidance, NVIDIA keeps reporting beats.
There is one important nuance. NVIDIA’s most recent disclosure noted that hyperscale revenue stayed at “approximately 50% of Data Center revenue” while the other half came from a growing mix of AI Clouds, industrial, enterprise and sovereign customers. In other words, NVIDIA is in the process of diversifying away from the hyperscalers, even while the hyperscaler dollars themselves are still growing in absolute terms. The pipeline is widening on both ends.
The cost of being on the buy side
So what does it cost a company to be a hyperscaler in 2026? It costs nearly all of your operating cash flow.
In 2020 and 2021, when the cloud business was already a major part of these companies, Microsoft was spending roughly 25% of operating cash flow on capex. Meta was spending around 35-40%. Amazon was the outlier even then because it was still building out fulfillment, sitting at 60-130% across the early 2020s. None of these companies had ever sustained over 50% of OCF reinvested for very long, because it crushes free cash flow.
Then the AI cycle started. In 2025, Amazon ran at 95% of OCF, Microsoft at 47%, and Meta at 60%. In 2026, Amazon will run at roughly 100% on a full-year basis. Reuters’ Morning Bid podcast called this milestone “the entire 2026 market story.” When a company commits to spending its entire operating cash generation on capex, free cash flow goes to zero by definition. There is no margin of safety left.
NVIDIA, by contrast, is sitting at 5-6% of OCF on capex. It is a fabless chip designer with no need to own factories or land, and it sells software-and-hardware bundles with 75% GAAP gross margins. The same dynamic that crushes its customers’ free cash flow accrues directly to its bottom line.
This is the cleanest visualization of why the green NVIDIA line is climbing and the others are sagging. They are mirror images of the same dollar.
Per company - what the cash actually looks like
Amazon is the most extreme case in the panel. Operating cash flow climbed from $84.9B in 2023 to $115.9B in 2024 to $139.5B in 2025 - an absolute increase of $54.6 billion in two years. By any normal standard, this is an extraordinary cash generation engine. Then the AI capex jumped, and free cash flow fell from $32.9B to $7.7B. Some analyses, including 24/7 Wall St. citing Bloomberg coverage, report that on a trailing twelve-month basis FCF has fallen as low as $1.2 billion, depending on whether finance leases are included. The 2026 capex guide of $200 billion, announced by Andy Jassy on the Q4 2025 earnings call on February 5, 2026, implies that capex will exceed the entire 2025 OCF in absolute dollar terms. Amazon stock fell about 11% in after-hours trading the night of that announcement. Jassy’s explanation was simple: “We are monetizing capacity as fast as we can install it.” AWS grew 24% year over year in Q4 2025, the fastest in 13 quarters, with an annualized run rate of $142 billion. The bet is that demand will keep validating the spend.
Microsoft is in a calmer place financially because its FY ends in June, which means its FY25 capex of $64.6B does not yet fully reflect the AI surge that the calendar-year hyperscalers are showing. Even so, capex is now accelerating sharply. The FY26 Q3 earnings call (April 2026) confirmed $30.9 billion of capex in a single quarter, up 84% year over year, and CFO Amy Hood guided to over $40 billion in Q4 alone. For calendar year 2026, Microsoft expects approximately $190 billion of capex, including roughly $25 billion attributable to higher component pricing. The capex guide is now matching Alphabet’s. Importantly, Microsoft is openly admitting that it is supply-constrained: “Even with these additional investments... we expect to remain constrained at least through 2026.” The constraint is the AI customer demand for Azure capacity, which Microsoft cannot install fast enough.
Meta is the most interesting company in the panel because it does not run a third-party cloud business. It is buying GPUs for its own services - Reels, ranking, ad targeting, Llama, and increasingly “personal superintelligence” per Zuckerberg’s January 2026 letter. Meta’s 2025 capex of $69.7 billion against $115.8 billion of OCF gives the company more headroom than Amazon, but the 2026 guide of $115-135 billion approaches the limit. Meta’s long-term debt has grown to $58.7 billion as of December 31, 2025, up from $28.8 billion a year earlier. The company is openly borrowing to fund the build. The $27 billion Hyperion data center financing arrangement that Barron’s reported on last year is one such example. Free cash flow in Q3 2025 fell 35% despite higher operating cash flow, then bounced back to $14.1 billion in Q4 2025. The pattern is volatility, not weakness.
NVIDIA is the other end of every one of these stories. In FY26, the company generated $215.9 billion in revenue, $96.7 billion of free cash flow, and 75% GAAP gross margins. In Q1 FY27 alone, it printed $48.6 billion of FCF, up from $26.1 billion a year prior. It returned $20 billion to shareholders in a single quarter and announced an $80 billion incremental buyback authorization. Inventory rose to $25.8 billion and supply commitments hit $119.0 billion, reflecting the company’s confidence that the demand will keep coming. As long as the four hyperscalers keep guiding capex up, NVIDIA’s FY27 will look like FY26 only larger.
What can break the pipeline
This is where it gets interesting, because pipelines have failure modes.
The first and most public risk is custom silicon. Each hyperscaler is investing aggressively in its own chip designs in order to reduce dependence on NVIDIA. Amazon’s chip business (Graviton, Trainium, Nitro) just crossed a $20 billion annual run rate in Q1 2026 according to the 2025 shareholder letter, growing at triple-digit percentages. Google’s TPU is on its eighth generation and a multi-billion-dollar investment was just made into Anthropic to anchor TPU demand. Microsoft has Maia. Meta has MTIA and has parallel-tracked a 6 gigawatt MI450 deal with AMD. If even a third of hyperscaler training-and-inference workloads shift to in-house silicon over the next three years, NVIDIA’s Data Center growth rate will slow materially. The market noticed: NVIDIA stock fell roughly 5.5% on February 27, 2026 after the Q4 FY26 report, with customer concentration cited explicitly as the primary risk factor.
The second risk is power and the physical world. Roughly 50% of planned U.S. data center builds have been delayed or cancelled due to power infrastructure shortages and parts unavailability, per coverage from Tom’s Hardware on the AI-capex constraint. If the hyperscalers cannot connect new capacity to the grid, capex spending can still grow on paper while NVIDIA shipments slow. There is a non-trivial lag between writing a check for a data center and plugging in a GPU.
The third risk is inference efficiency. The economics of training a frontier model are very different from the economics of serving it once it is built. If model architecture continues to compress - smaller models, mixture-of-experts, distillation, longer-context-with-less-compute approaches like Mamba or efficient attention - then GPU-per-token requirements fall. There is also the ASIC story. Cerebras’ WSE-3 wafer-scale design, the Groq LPX rack that NVIDIA itself launched at GTC, and various custom inference chips from startups are all trying to make NVIDIA’s H- and B-series look overkill for production inference workloads. NVIDIA’s CEO Jensen Huang acknowledged on the Q1 FY27 call that the new Groq LPX product “will be a niche product for some time” because its throughput is low, but the underlying market signal is clear: customers are actively searching for cheaper inference.
The fourth risk is the demand cliff. If the hyperscaler return on AI investment lags, the natural reaction is to slow the next round of capex. The current pattern is “monetize as we install.” If monetization slows, the next capex round shrinks. NVIDIA’s quarterly revenue trajectory would then bend - not collapse, but bend - because the order book is heavily front-loaded. Wall Street’s Brent Thill is on record telling the Financial Times that “the bear thesis is garbage” and that current revenue growth justifies the capital outlays. He may be right. But that thesis only holds if revenue keeps validating the spend. If the AWS-AI revenue run rate of $15+ billion stops growing in line with the $200 billion 2026 capex, the math turns quickly.
What this actually means
The way to read this is not as a NVIDIA bull case or a hyperscaler bear case. It is as a tightly coupled system.
Microsoft, Amazon and Meta are willingly trading near-term free cash flow for AI infrastructure capacity, and they are explicit about that decision. Jassy’s “high demand” language, Hood’s “remain constrained” language, and Zuckerberg’s “infrastructure costs will drive expense growth” language all point the same direction. They are not being forced. They are choosing. The question is whether the underlying demand they cite turns out to be real and durable.
NVIDIA is the immediate beneficiary because it sits at the chokepoint of the buildout. The 75% gross margins and capital-light model mean that every incremental capex dollar from the hyperscalers gets converted into FCF at a remarkable rate. That is why the green line is going vertical. It is also why the dividend went up 25x and another $80 billion of buyback got authorized in a single board meeting.
The interesting trade is not “will NVIDIA report a good FY27.” It almost certainly will, because the order book and the supply commitments are already on the balance sheet. The interesting trade is what happens when the hyperscaler ROI report cards come in. If the AI revenue growth at AWS, Azure and Meta’s ad business validates the capex, the pipeline keeps flowing and 2027 capex guides go up again. If it does not, three of the largest companies in the world have to choose between flat free cash flow and slowing their build, and the chokepoint company on the other end has to deal with the digestion period.
We are watching one of the largest concentrated capital reallocation events in corporate history happen in real time. The pipeline is wide open right now. Whether it stays open depends on the only number that matters in the end, which is the revenue these four companies generate from the capex they are buying.
Keep an eye on the cash conversion. The rest follows.
Sources
NVIDIA Corporation, “NVIDIA Announces Financial Results for First Quarter Fiscal 2027,” press release, May 20, 2026. Available at the NVIDIA investor relations site.
NVIDIA Corporation, “NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026,” press release, February 25, 2026.
NVIDIA Corporation, Form 8-K, Q1 FY27 CFO Commentary, May 20, 2026, SEC EDGAR accession 0001045810-26-000051.
Amazon.com Inc., Q4 2025 Earnings Release, February 5, 2026, SEC EDGAR filing.
Amazon.com Inc., 2025 Annual Report and Shareholder Letter, April 2026.
Microsoft Corporation, Form 8-K, Q3 FY26 Earnings Release, April 2026, SEC EDGAR accession 0001193125-26-191457.
Microsoft Corporation, FY 2026 Q3 Earnings Conference Call transcript, April 2026.
Meta Platforms Inc., “Meta Reports Fourth Quarter and Full Year 2025 Results,” press release, January 28, 2026, SEC EDGAR accession 0001628280-26-003832.
Meta Platforms Inc., Q3 2025 Earnings Release, October 2025, SEC EDGAR filing.
DCD (Data Center Dynamics), “Two unnamed customers accounted for almost 40% of Nvidia’s Q2 2026 revenue,” March 11, 2026.
Daloopa, “NVIDIA Customer Concentration: A Big 4 Earnings Preview,” April 2026.
Tom’s Hardware, “Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year,” April 30, 2026.
Fortune, “Big Tech is about to spend $700 billion on AI this year,” April 30, 2026.
CNBC, “Tech AI spending approaches $700 billion in 2026, cash taking big hit,” February 6, 2026.
CNBC, “Nvidia earnings takeaways: Data center revenue nearly doubles, report is strong but stock slides,” May 20, 2026.
24/7 Wall St., “Amazon’s Free Cash Flow Just Collapsed From $26 Billion to $1.2 Billion. The Market Barely Blinked,” May 2026.
Futurum Group, “Amazon Q4 FY 2025: Revenue Beat, AWS +24% Amid $200B Capex Plan,” February 9, 2026.
Levelheaded Investing, “Amazon’s Q4 2025 Earnings and 2026 Outlook,” February 18, 2026.
Om Malik, “What I Learned about Hyperscalers’ AI Spend,” April 30, 2026.
ServeTheHome, “NVIDIA Reports Q4’FY2026 Earnings: Data Center and ProViz Drive Revenue Records,” February 26, 2026.
Constellation Research, “Amazon sees $200 billion in capex ahead, AWS sales surge 24%,” February 2026.
Data Center Frontier / Data Center Richness, “Hyperscalers Plan $630 Billion in 2026 CapEx,” February 6, 2026.
Introl, “Hyperscaler CapEx Hits $600B in 2026,” January 7, 2026.
S&P Global Market Intelligence, “Nvidia earnings preview: Q1 2027,” May 2026.
Quantitative data: Yahoo Finance (yfinance) quarterly cash flow statements and SEC EDGAR XBRL company-facts API for NVDA, MSFT, AMZN, META, accessed May 23, 2026.