The Cloud Business Just Stopped Paying for Itself — and That Changes Everything

The Cloud Business Just Stopped Paying for Itself — and That Changes Everything

Let me tell you something that's been sitting wrong with me all week. I've been running hosting infrastructure for over a decade, which means I've spent ten years watching the cloud giants do two things at once: sell me and everyone else on the idea that scale is destiny, and quietly build a machine that generates more cash than any industry in history. The cloud was the money printer. It funded the rest of the tech world's toys. And this week I ran the numbers again and realized something I should have seen coming: the printer has stopped paying for itself.

UBS dropped a forecast on August 22 that should have been front-page news. The bank now estimates that Amazon, Alphabet, and Microsoft — the three biggest cloud companies on earth — will collectively spend about 102% of their cloud revenue on capital expenditures in 2026. Not 60%. Not 80%. More than every dollar they bring in from the business that supposedly prints money. The cloud isn't funding the AI buildout anymore. The AI buildout is eating the cloud.

One Number Should Stop You Cold: 102 Percent

Let me be precise about what UBS is saying, because the details matter more than the headline. The 102% figure doesn't mean these companies are losing money overall — Amazon still sells groceries and runs logistics, Alphabet still owns YouTube and Search, Microsoft still sells Office and Windows. The ratio is specifically cloud revenue being recycled into capex. But that's exactly why the number is so damning. The single most profitable, most cash-generative business model of the last two decades is now pouring every dollar it earns back into building more of itself — and then some.

And the ratio isn't even the scariest part. UBS projects total hyperscaler capex at $492 billion in 2025, $1.009 trillion in 2026, $1.447 trillion in 2027, and $1.619 trillion in 2028. Cumulative, that's roughly $4.1 trillion over three years — more than three times the $1.292 trillion these companies spent across the previous six years combined. The ratio eases to about 99% of cloud revenue in 2027 and 94% in 2028, sure. But the dollar amount keeps climbing. You can spend a smaller percentage of a much bigger revenue base and still spend more absolute money every single year.

Look at the company-level numbers and it gets even more surreal. UBS puts Alphabet at roughly $938 billion of capex over 2026-2028. Meta at $683 billion. Microsoft at $672 billion. Amazon at $628 billion. Then SpaceX at $335 billion. Oracle at $276 billion. And the neoclouds — CoreWeave at $130 billion, Nebius at $93 billion. Eight companies spending the better part of four trillion dollars in three years. That's not an upgrade cycle. That's an industrial reset.

The Crossover Is Happening Right Now

Here's the part that keeps me up at night: this isn't a forecast about 2028. The moment is now. Epoch AI has been tracking the quarterly cash flows of Microsoft, Amazon, Alphabet, Meta, and Oracle from their SEC filings, and their data shows aggregate cash capex across the hyperscalers is on track to overtake operating cash flow around Q3 2026. That's this quarter. The quarter we're living in right now.

The growth rates tell the story. Operating cash flow is growing about 23% a year across those five companies. Cash capex is growing about 70% a year. When one line grows three times faster than the other, they cross — and they're crossing right about now. After that crossing, free cash flow — operating cash flow minus capex — goes to zero, then negative. The Epoch analysis puts it plainly: most hyperscalers have already turned to external financing to fund the buildout, or are considering it. The biggest cash machines in economic history are now structurally dependent on outside money.

Reading One: The Machine Is Working as Designed

Now, before I get accused of being one of those guys who's called the bubble for four years straight, let me steelman the bull case, because it's not stupid. There's a reading where 102% is exactly the right number. If AI demand is real — and Gartner's Monday forecast says semiconductor revenue will pass $1.6 trillion this year, nearly doubling year over year, with AI data center spending becoming more than half of all semiconductor revenue within four years — then the companies that build the most infrastructure fastest win. Recycling 100% of cloud revenue into the buildout isn't reckless. It's the only rational move when the prize is a market nobody's fully captured yet.

The other thing the bulls get right: this isn't one company's speculative bet. When Amazon, Alphabet, Microsoft, Meta, SpaceX, Oracle, and a dozen neoclouds are all spending in the same direction, that's an ecosystem making a collective commitment, not a single balance sheet gambling. Spread across that many players, the risk of any one company's miscalculation killing the whole thing drops a lot. The market is betting AI demand becomes large enough to keep this infrastructure productive for years. That's not crazy. It's just expensive.

Reading Two: The Self-Funding Engine Just Broke

But here's the other reading, and it's the one I can't shake. When a business spends more than it generates, every dollar beyond operating cash has to come from somewhere else. And we're watching that "somewhere else" light up like a pinball machine. CNBC's Fast Money flagged it back in July: across the hyperscaler group, roughly 92% of operating profit is now spent on capex. Meta is raising $12 billion at 7.5% — a rate the show's host called "just below like freaking junk" — with weak bond demand. Jim Chanos keeps pointing at the mismatch between long-duration debt and the short depreciation schedule of the GPUs those bonds finance. And the off-balance-sheet obligations — the leasing structures, the take-or-pay contracts, the commitments nobody can see from a P&L — go largely undiscussed.

Meta is the cleanest example. CFO Susan Li lifted full-year capex guidance to $125-145 billion, Q1 capex alone hit $19.8 billion, and the company disclosed a $107 billion step-up in contractual commitments from multiyear cloud deals. That's a $107 billion obligation sitting outside the day-to-day income statement. Zuckerberg says every signal gives him confidence. Fine. But the bill is real, and it's due, and the interest rate on it just went up.

Here's where it connects to everything else I've been writing about for the past month: the GPU-backed loans, the $13.7 billion deals signed by companies that can't afford them, the junk-bond crossover, the Treasury buying its own bonds to keep the buildout cheap. They're all the same story. The cloud business used to be the money printer that funded the tech industry. Now the tech industry's entire AI bet requires external capital at every layer — and the layers are all borrowing from the same pool of nervous lenders.

The Secondary Bottleneck Nobody's Pricing In: External Capital Dependence

And here's the kicker — the demand side is pushing back at the exact same moment. Gartner's data shows CIOs are entering a cautious deployment era as the bills pile up. Providers have already shifted from flat-rate subscriptions to usage- and outcome-based pricing, and enterprises hate it. Info-Tech's analysts say continuously looping AI agents are the biggest cost driver — the technology double-guesses itself and reprocesses data, burning money in ways nobody predicted. Travelers Insurance built its own LLM just to keep costs down. CIOs are openly talking about using cheaper open-source models for routine work. Snowflake, AWS, and Oracle are all shipping cost-control features like lifeboats.

Put those two forces together and you get the real structural bottleneck of this buildout, and it's not chips and it's not power. It's capital dependence. The spenders need more external money than ever, while the people paying the bills — the CIOs — are actively looking for ways to spend less. That's a two-sided squeeze on the same equation. When 102% of revenue goes to capex, price increases are not optional, they're arithmetic. And when prices go up, the customers who can run on cheaper models will. The gap between what the buildout costs and what the market will pay for compute is the gap that has to be filled with borrowed money — and borrowed money is getting more expensive and harder to find.

What This Actually Means for Independent Hosting Providers

If you're running an independent hosting business, this is the most important setup in a decade. Let me give you the playbook straight.

First, watch the OCF-to-capex ratio, not the revenue headlines. The revenue numbers are going to keep smashing records — Google Cloud growing 80%, Azure growing 40%, everyone beating estimates. None of that tells you anything about the machine's health anymore. The ratio of operating cash flow to capex is the canary. It's about to cross zero for the first time. That's the number that will tell you when the music stops, and it's public data from SEC filings. Watch it quarterly like your business depends on it — because it does.

Second, position as the capital-light alternative — now, before the price increases land. The hyperscalers have no choice but to raise cloud prices. When you spend 102% of revenue on capex, depreciation alone forces your pricing up. That's your opening. Every dollar of price increase they push onto their customers is a customer who suddenly finds independent hosting very interesting. Make sure your capacity, your support, and your story are ready for that conversation before the rate card changes, not after.

Third, don't borrow to compete with people who borrow. The temptation when you see this much capital sloshing around is to lever up and grab market share. Resist it. The hyperscalers are borrowing at 7.5% and near-junk spreads because they're building trillion-dollar campuses. Your margin of safety is that you don't have a $683 billion capex plan. A hosting business that's cash-flow positive, capital-light, and debt-free is going to be the rarest asset in tech when the credit cycle turns. Don't trade that away for a bigger rack footprint.

Fourth, lock your own financing now. If you do need capital for capacity expansion, get it while the credit window is still open. Right now lenders still treat data center-adjacent businesses as AI-adjacent and price accordingly. When the junk market blinks — and it will — it won't discriminate between a $100 billion hyperscaler and a $2 million hosting company. Both will find the door shut. Borrow while the door is open, or don't borrow at all.

The Bottom Line

For twenty years, the cloud was the industry's cash cow. It generated the profits that funded everything else — the startups, the R&D, the next wave of expansion. That era just ended. The cow is now a conduit: money goes in one end, comes out the other end as a data center, and whatever doesn't fit in the middle gets borrowed at 7.5%.

I'm not calling the top, and I'm not telling you AI is a bubble. I'm telling you the physics changed. The buildout now costs more than the business that's supposed to pay for it, and the difference is being financed by a nervous credit market. That's not a crash warning. It's a math warning. And the math says the companies that survive this cycle won't be the ones spending the most — they'll be the ones who understood that when the printer stops paying for itself, the only thing that matters is who still has cash when the borrowing gets expensive. Plan accordingly.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: UBS via 24/7 Wall St. (Aug 22), Epoch AI Data Insights (Jun 16), CNBC Fast Money via AOL (Jul 29), Gartner via CIO Dive (Aug 24).