Apollo’s Slok finds declining returns on invested capital at five cloud giants

The largest cloud operators are adding invested capital faster than they are generating earnings, pushing returns on that capital lower, Apollo chief economist Torsten Slok found in an October 20 analysis. His warning connects the infrastructure boom to a customer-side constraint: demand for AI computing depends on what businesses are willing to pay.

Slok points to slowing growth in corporate token spending as evidence against the assumption that demand for computing is effectively unlimited. Tokens are the usage measure behind many AI bills. As employees make more requests, or AI agents process data and tackle more complex tasks, usage-based charges can climb. More activity therefore creates an expense that customers must justify, not simply a larger market for suppliers.

His argument is that buyers are highly sensitive to price. Even modest changes in token prices can prompt them to reduce usage or switch to cheaper open-source alternatives. That gives customers two ways to respond to a bill they consider too high: buy less computing, or obtain it more cheaply. Neither response supports a simple assumption that growing interest guarantees growing spending.

Slok puts the commercial test ahead of the physical constraints that dominate many infrastructure discussions. In his assessment, the decisive variable for data-center construction is the price customers will pay, rather than capacity or power alone. Higher prices require business benefits: stronger revenue growth, larger profit margins or substantial cost savings. He argues those benefits are not yet appearing at scale.

The core economic flaw in the AI bull case is that the marginal cost of producing compute exceeds the marginal benefit it delivers to users, raising questions about claims of unlimited demand for data centers.

Torsten Slok, Apollo chief economist, October 20, 2026

The supply-side evidence concerns five major cloud operators: Amazon, Google, Meta, Microsoft and Oracle. Apollo’s analysis finds their return on invested capital declining as the capital committed grows faster than earnings. Its calculation divides trailing 12-month operating earnings after taxes by equity plus debt, less cash and short-term investments. The measure asks how much earnings the capital base generates, rather than simply whether earnings are growing.

That distinction helps reconcile this finding with Slok’s October 24 comparison of the Magnificent Seven and the other 493 S&P 500 companies. That assessment found rising revenue per worker and profit margins for the seven, but flat results for the rest. It describes a divide in company performance; the October 20 capital-return measure tests the earnings generated relative to the money committed.

Rising margins and falling returns on invested capital are therefore not contradictory, nor do the two analyses cover identical groups. A business can improve its margins while expanding its capital base faster than earnings. Slok’s concern is not merely whether the suppliers are profitable, but whether the expanding investment earns enough in return.

Slok also questions what current spending reveals about durability. Fortune’s October 24 account of his notes reports his finding that 80% of the top 1% of AI spenders are venture-backed. He argues that these startups are financing compute bills with venture capital rather than organic earnings. That makes the source of customers’ money part of his challenge to the demand thesis.

The unresolved issue is whether business benefits are delayed or will remain insufficient to support the prices suppliers need. Slok explicitly leaves both possibilities open in the October 20 note. His analysis challenges the economics of the buildout today; it does not settle whether companies will eventually turn AI use into substantially higher revenue, profits or savings.