From Cash Machines to Capital Consumers: The Hyperscaler Shift

The biggest technology companies are swapping one business model for another. For twenty years, Google and Meta ran asset light businesses that produced huge surplus cash. Now, to build for AI, they're becoming capital heavy businesses that consume cash and borrow more to keep going.

Both models can work, but they carry very different trade offs. Here's each one, side by side.

The old model: asset light and cash-rich

The traditional hyperscaler ran on software, data and scale. The companies spent upfront to build the technology, then grew revenue across millions of users without their costs rising at the same pace, and the marginal cost of serving one more customer was close to zero. That produced enormous, dependable cash flow, more than the businesses could reinvest, so the surplus went into new products, acquisitions, buybacks, or simply sat on net cash balance sheets. For most of the past decade, capital spending ran at roughly 40 percent of the cash these businesses generated, which left a comfortable cushion every year.

Pros

• Very high margins, because software scales without much added cost.
• Large surplus cash, enough to fund everything internally with plenty left over.
• Strong, flexible balance sheets holding more cash than debt.
• Low risk from any single bet, since growth didn't depend on heavy capital.

Cons

• All that spare cash arguably bred loose capital discipline, funding side projects that led nowhere.
• They owned relatively little of the physical layer beneath them, which mattered less then than it does now.

The new model: capital heavy and debt funded

AI broke the old pattern, because AI isn't light. Training and running large models needs chips, servers, data centres, power and cooling, all of it physical, expensive and owned rather than rented. So the companies that once grew off a small footprint are now among the largest infrastructure builders, and the numbers show it. JP Morgan estimates the five largest US hyperscalers will spend around $697 billion of capital in 2026, close to 100 percent of the cash they generate from operations, against that ten year average nearer 40 percent. Combined free cash flow for the four big names is heading for its lowest level since 2014, some of them have already turned free cash flow negative, and they raised well over $120 billion of new debt in 2025 to help fund the build.

Pros

• Owning the AI infrastructure could become a durable advantage, since compute is the scarce input everyone needs.
• It creates a real barrier to entry, because few others can afford to build at this scale.
• More control over their own capacity, and less dependence on outside providers.
• If AI demand shows up as hoped, these become the toll roads of the next computing era.

Cons

• The surplus cash that defined these companies is disappearing, and free cash flow is falling fast.
• They're taking on real debt and leverage for the first time, changing their risk profile.
• The main asset, AI chips, loses 30 to 40 percent of its value a year, so a lot of the spending has to be repeated just to stand still.
• The payback is uncertain, because AI revenue is still far smaller than the capital going in.
• The infrastructure is built to last decades, while the technology that makes it valuable can change in a couple of years, which risks a costly mismatch if demand disappoints.

How they're paying for it through creative financing

The big question with the new model isn't how much they're spending, it's how they're paying for it, hyperscalers have found ways to fund the build without carrying all the debt themselves. The main tool is a residual value guarantee. A separate company is set up to own the data centre and borrows the money, so the tech company doesn't take on the loan itself. Instead, it promises the asset will be worth at least a set amount later, and covers the gap if it sells for less. That way it backs the deal without the debt showing up as its own.

In the past year alone, tech companies have offered up to $300 billion of these guarantees, and Morgan Stanley counts more than $3 trillion of off-balance-sheet commitments and credit support across the major hyperscalers and chipmakers. Meta started it, backing its 2GW Hyperion project in Louisiana with a guarantee of around $28 billion, which helped that venture raise $27 billion of debt at rates close to Meta's own. Nvidia has since done the same for the projects buying its chips, including around $105 billion of guarantees tied to an Ohio campus being built for OpenAI, and Broadcom has taken on billions in exposure to help finance chips for Anthropic.

The catch is that all of this adds a layer of risk that's hard to see from the outside. Credit analysts warn that the rapid growth of these off balance sheet arrangements makes the companies' true risk profiles more complex, and the guarantees only stay costless as long as the AI bet pays off. If usage disappoints, or the industry builds more capacity than it can use, the value of these assets could fall below the guaranteed floor, and the shortfall would land back on the tech companies that promised to cover it. Rating agencies do adjust for some of this, adding to a company's leverage when they think the guaranteed value sits above what an asset would fetch in a stressed sale, though for now they often judge that gap to be small.

What this means

The old model worked precisely because it barely needed capital. The new one has to prove that all this spending turns into durable profit, not just a permanent cost of staying in business. It's a trade made at the largest scale in corporate history: the hyperscalers are giving up the asset light model that made them the most profitable companies, on the bet that owning AI infrastructure is worth more.

The wider lesson is that business models aren't permanent. The asset light machine looked untouchable for two decades, and then the next big opportunity turned out to be heavy, physical and expensive. We'll only know if it was a good trade once the revenue either catches up with the spending, or it doesn't.