For three years, the AI trade ran on a simple promise: spend now, get paid later. The spending part is no longer in doubt. The getting paid part is where the debate has moved, and it is a better debate than the one we were having when everyone argued about whether chatbots were clever enough. Clever is settled. The question now is arithmetic.
The size of the bet
Start with the scale, because it is hard to hold in your head. According to a June Forbes analysis citing CreditSights estimates, the five largest hyperscalers (Amazon, Microsoft, Alphabet, Meta and Oracle) are projected to spend somewhere between $700 billion and $900 billion in 2026. Amazon alone is pointed at roughly $200 billion. Alphabet has guided to $175 billion to $185 billion. Meta has talked about as much as $145 billion, and Microsoft is above $120 billion for its fiscal year.
Those are not rounding errors on a normal budget. The same Forbes piece says roughly 75% of that money goes toward GPU clusters, custom accelerators, data centers, and the power and cooling gear that keeps them running. And the forecasting does not stop at this year. Analysts at Evercore and Bank of America, as cited there, see the combined figure potentially exceeding $1 trillion in 2027.
If you own a broad index fund, you own a slice of this whether you meant to or not. The biggest companies in the market are also the biggest spenders, and that makes the return on this money a question for almost every portfolio, not just the ones with a semiconductor tilt.
Where the money comes back
Here is the fair case for the spenders. Revenue is showing up. Forbes reports that AWS is running at roughly $150 billion annualized and growing 28% year over year, that Google Cloud grew 63% in its first quarter, and that Microsoft described its AI business as a $37 billion annual run rate, up 123%. A separate Yahoo Finance report from May put Google Cloud’s backlog at more than $460 billion. Backlog is not revenue, and it is not profit, but it is a signed indication that customers are lining up.
Then there is the supplier side. Nvidia’s data center revenue came in at $62.3 billion for its fourth quarter, up 75% from a year earlier, according to the same reports. When one company’s quarterly data center sales look like that, you are looking at someone else’s capital budget turned into a line on an income statement. Every dollar of hyperscaler spending is a dollar of someone’s revenue. That is the whole mechanism, and it is why the chip names have been the cleanest expression of the trade so far.
The catch is obvious once you say it out loud. Spending flows to suppliers immediately. Returns flow to the spenders slowly, if they arrive at all.
The part that should make you pause
UBS estimates, reported by 24/7 Wall St. in August, that Amazon, Alphabet and Microsoft will collectively spend about 102% of their cloud revenue on capital expenditures in 2026. Read that again. Not 102% of profit. Revenue. UBS expects the ratio to ease to 99% in 2027 and 94% in 2028, which would still leave the business reinvesting nearly everything the cloud brings in.
This is how a cash machine turns into a capital machine. A Yahoo Finance report in May noted that Amazon’s trailing twelve month free cash flow had fallen to $1.2 billion, down 95%. That is one company and one snapshot, and free cash flow can be lumpy for good reasons. But it shows what heavy building does to the number that dividends, buybacks and balance sheet strength depend on.
The financing picture is shifting too. Forbes reports the big five raised $108 billion in new debt in 2025 and cites projections of $1.5 trillion in tech sector debt issuance over several years. Companies that once funded everything from operating cash are leaning more on the bond market. That is not a crisis. It is a change in character, and investors who price these companies like asset light software businesses may be using the wrong mental model.
Then comes the gap argument. Sequoia’s David Cahn has updated his napkin math to roughly $1.5 trillion a year of end-customer revenue needed to justify current AI infrastructure spending, up from the $600 billion figure that circulated earlier in the cycle. Allianz Research, as cited by Forbes, measures the divergence between capex and revenue at 46%, versus 32% during the 2001 telecom buildout. Those figures come from analysts with a point of view, and each one rests on assumptions about what counts as AI revenue. Treat them as arguments, not verdicts.
The telecom comparison, and where it breaks
The telecom parallel is the one everybody reaches for, so it deserves a careful look. In the late 1990s, companies laid fiber far ahead of demand, and much of it sat dark for years. The infrastructure turned out to be useful eventually. The investors who financed it mostly did not live to see that.
There are real differences today. The spenders are some of the most profitable companies ever built, not leveraged startups. Demand for compute, by their own accounts, is running ahead of supply rather than behind it. And GPUs are not buried in the ground. They are in use now, earning rent.
But the similarities are not trivial either. Chips age quickly. A data center built around one generation of hardware does not stay current for free. Power is a hard physical limit. And there is a study that every skeptic cites for a reason: the MIT Project NANDA report from July 2025 found that 95% of enterprise generative AI pilots produced no measurable profit and loss impact, on roughly $30 billion to $40 billion of corporate spending. One study is not the whole economy, and the pilots it covered are early. Still, it is a reminder that buying the technology and profiting from it are separate events, and the second one belongs to the customers, not the vendors.
What a thoughtful investor watches
Forget the daily stock moves. A handful of indicators tell you more. The first is the capex to cloud revenue ratio, the one UBS flagged. If it drifts down as promised, the story holds together. If it climbs past 100% and stays there, the market will eventually ask for a different price.
The second is free cash flow at the spenders. Capex is a choice, and management teams can defend it. Cash flow is the scoreboard. The third is the source of funding. A buildout paid from operating cash is a different animal from one paid by debt, and a shift toward heavier borrowing deserves attention.
The fourth is customer results. Cloud growth rates and backlog are encouraging, but the durable proof is enterprises reporting real savings or real new revenue. Forbes cites an estimate that meaningful agentic AI deployment in the enterprise is 12 to 24 months away. That is a forecast, and forecasts in this field have a habit of arriving late or early.
The fifth is concentration. Forbes puts Nvidia at roughly 90% of AI accelerator spend. A supplier with that share is enormously profitable in a boom, and uniquely exposed if the buyers decide to pause, or if their custom chips take share. Amazon’s chip business was described in the May reporting as running at about a $20 billion annual revenue rate, which tells you the largest customers are building alternatives.
The honest uncertainty
Nobody knows how this ends, and anyone selling certainty in either direction is selling something. The bull case needs revenue to catch up with spending on a reasonable timeline. The bear case needs it to lag long enough for the financing to strain. Both are plausible, and the evidence so far fits each of them in places.
What the numbers do support is a narrower claim. The AI story has graduated from a technology question to a capital allocation question. That changes what a careful investor should read. Product demos matter less. Cash flow statements, debt schedules and capex guidance matter more.
The spenders are betting that the demand is real and the payback is coming. The suppliers are already being paid. And the rest of us are somewhere in the middle, holding shares in all of it, which is a good reason to understand the math before the narrative changes again.
This article is written for educational and informational purposes only and does not constitute financial or legal advice. The views and analytical frameworks presented draw on publicly available information and reported commentary from industry participants. Readers are encouraged to consult primary sources and form their own informed views on these complex topics.
