Who Actually Pays for AI? Inside the $700 Billion Hyperscaler Capex Wave

Four companies — Amazon, Microsoft, Alphabet, and Meta — have told investors they will spend roughly $700 billion on capital expenditure in 2026, the vast majority of it on AI data centers. That is close to two billion dollars every day. It is the largest coordinated private-sector infrastructure buildout in history, and it is being paid for out of the cash flows of the most profitable business models ever created.

This article follows the money using the companies' own filings: who is spending what, what the spending is doing to their famously bulletproof cash flows, whose revenue it lands in, what evidence exists that the demand is real — and, because three of these quarters had earnings flattered by one-time items, why reading the footnotes matters more this year than most.

The spenders: four budgets, one direction

First, terms. Capital expenditure (capex) is money spent on long-lived physical assets — in this story, land, buildings, power, cooling, and above all the racks of AI chips inside. The hyperscalers are the handful of companies operating cloud data centers at global scale. Their 2026 budgets, from their own earnings calls: Amazon about $200 billion, Microsoft about $190 billion (calendar-year basis), Alphabet $180–190 billion, and Meta $125–145 billion — a range Meta raised mid-year from $115–135 billion, citing “higher component pricing.”

The scale is new, but so is the speed: combined, this is nearly double what the same four spent in 2025. Amazon spent about $44 billion in the first quarter alone — versus $25 billion in the same quarter a year earlier. Alphabet's Q1 capex of $35.7 billion was up 107% year over year. Microsoft guided its single fiscal fourth quarter to “over $40 billion.” And notably, two of the four now itemize memory prices inside those budgets — Microsoft attributes roughly $25 billion of its total to higher memory costs — which is how this article connects to the memory supercycle we covered separately.

The strain: what it's doing to cash flow

The cleanest way to see the cost is free cash flow — the cash a business generates after operating costs and capital spending; the money that funds buybacks and dividends. These four companies are history's greatest free-cash-flow machines, and the buildout is visibly bending them.

Amazon is the starkest case: trailing-twelve-month free cash flow collapsed from roughly $26 billion to $1.2 billion — operating cash flow of $148.5 billion almost entirely consumed by $147.3 billion of property-and-equipment purchases. Alphabet's quarterly free cash flow fell 47%; it paused share buybacks entirely ($0 versus $15 billion a year earlier) and issued $31 billion of new debt to help fund the program. Microsoft's fell 22%, with cloud gross margin compressing from 69% toward a guided ~64% as depreciation from all that hardware flows through. Meta also paused buybacks — but is the counter-example on cash generation: its free cash flow grew 20% despite capex up 45%, evidence that a 41%-operating-margin ad business can absorb the buildout, at least for now.

Pause on what this means. Companies that spent two decades returning cash to shareholders are now, in aggregate, choosing chips over buybacks — and in Alphabet's and Amazon's cases, borrowing or spending down to fund it. That is either the most rational land grab in business history or the setup for the biggest write-down cycle in tech. It cannot be neither.

Where the money lands

Follow a capex dollar out of a hyperscaler's budget and it lands, mostly, in one company's revenue line. NVIDIA — designer of the GPUs (graphics processing units, the general-purpose AI chips) that dominate AI training — reported $75.2 billion of data-center revenue in a single quarter, up 92% year over year: 92% of everything the company sells. Its data-center networking line grew 199%. Guidance for the following quarter: $91 billion, assuming zero contribution from China.

The rational response to writing checks that size to one vendor is to build an alternative, and that is Broadcom's franchise: it co-designs custom accelerators (ASICs — chips built for one customer's specific workload, like Google's TPU or Meta's MTIA) for six marquee customers, Anthropic, Google, Meta, and OpenAI among them. Broadcom's AI semiconductor revenue hit $10.8 billion (+143%), roughly half the company, and management reiterated guidance for over $100 billion of AI revenue in fiscal 2027. Beneath both sit the layers we covered in the memory piece — HBM memory (Micron) and ASML's lithography tools — plus a genuinely new entrant: SpaceX, which is simultaneously a new capex whale (~$20.7 billion in 2025) and a compute seller, with disclosed deals to supply Anthropic (~$1.25 billion) and Google.

The receipts: is the demand real?

The obvious question — is anyone actually buying this much AI? — has an unusually documentable answer this cycle, because so much of the demand is contracted in advance. A backlog (formally, “remaining performance obligations”) is revenue a customer has already signed for but not yet consumed. Microsoft's commercial backlog stands at $627 billion, up 99% in a year. Google Cloud's backlog is $460 billion, nearly doubling in a single quarter. Microsoft's AI business alone runs at $37 billion of annualized revenue, up 123%; Google Cloud grew 63% with its operating margin expanding from 17.8% to 32.9%; AWS grew 28%, its fastest in fifteen quarters.

Two honest caveats belong next to that chart. Part of Microsoft's backlog surge reflects OpenAI's enormous Azure commitments — one counterparty; excluding OpenAI, bookings grew a much more ordinary 7%. And a backlog is a promise to spend, not proof the end products (copilots, agents, AI features) will earn their keep — the spend is contracted, the payoff is not.

Read the footnotes: three flattered quarters

A pattern worth knowing about this exact earnings season: three of the headline numbers in this story were inflated by one-time items, and the flattering was largest where the capex strain was greatest.

• Alphabet: EPS $5.11, up 82% — but roughly $2.35 of it came from a $36.9 billion one-time gain on equity securities it holds. Operating income, the cleaner read, grew 30%.
• Meta: EPS $10.44, up 62% — flattered by an $8.0 billion one-time tax benefit. Excluding it, EPS was about $7.31.
• NVIDIA: GAAP EPS $2.39, up 214% — inflated by a ~$15.9 billion gain on its own equity investments. The company's non-GAAP figure, $1.87 (up 140%), is the number that reflects the chip business.

None of this is improper — it's all disclosed. But a reader who stops at the headline EPS this season overstates how fast the underlying businesses grew, right at the moment those businesses are spending historically. When the same report contains a record capex number and a flattered earnings number, the flattered number is the one that gets quoted.

The contrarian: Apple sits it out

Not everyone is playing. Apple spends about $2 billion per quarter on capex — barely one percent of the big four's combined budget — investing in AI through R&D (up 34%) and renting compute rather than pouring concrete. Its reward for sitting out the buildout: a $100 billion buyback while rivals pause theirs. Its punishment: the buildout reached it anyway, through component prices — memory costs have forced Apple into what its CEO called “unavoidable” price increases, a story we cover in the memory-supercycle deep dive.

What a crack would look like

If the wave breaks, it will show up in a specific order. Signals worth tracking:

1. Capex guidance itself — the leading indicator for the entire chain. These budgets are re-guided quarterly; the first broad guide-down would hit chipmakers, memory, and toolmakers within quarters. (So far the revisions have gone the other way — Meta raised.)
2. AI revenue versus depreciation. Data-center hardware depreciates over roughly five to six years; hundreds of billions of capex become tens of billions of annual expense. Watch whether AI revenue lines (Microsoft's $37B ARR, cloud growth rates) keep compounding faster than that expense arrives.
3. The application layer's willingness to pay. The end revenue behind all this capex is software companies paying for inference — the cost of actually running AI models inside their products. App companies cutting AI features to protect their margins would be the early warning that the end demand is thinner than the backlogs suggest.
4. Cash-return restarts. Alphabet and Meta pausing buybacks was the tell that spending had priority. Buybacks resuming at scale would signal managements see the buildout cresting.

How to research this chain yourself

Every layer of this story is a public company you can study: the spenders (Amazon, Microsoft, Alphabet, Meta), the chipmakers collecting (NVIDIA, Broadcom), the layers beneath (Micron, ASML), and the abstainer (Apple). They are one chain: the same dollar flows through all of them, and the same guidance cut would echo through all of them. That's the frame — one theme, many tickers, shared risk.

Sources

• Company earnings releases, investor decks, and calls: Alphabet Q1 2026 (Apr 29), Microsoft FY26 Q3 (Apr 29), Meta Q1 2026 (Apr 29), Amazon Q1 2026 (Apr 29), NVIDIA Q1 FY2027 (May 20), Broadcom Q2 FY2026 (Jun 3), Apple FQ2 2026 (Apr 30), Micron FQ3 2026 (Jun 24), SpaceX IPO prospectus (Jun 2026).
• CNBC on Amazon's Q1 2026 and the ~$200B capex plan · 24/7 Wall St. on Amazon's free-cash-flow collapse
• Related deep dive: The 2026 Memory Supercycle, Explained