AI Boom’s Profits Are Currently Coming From Investors Rather Than Customers, Economist Says It...

The debate over whether artificial intelligence is in a bubble may be asking the wrong question.

A more consequential issue is emerging within the AI economy itself: the companies generating the highest profits from the boom are largely selling the infrastructure needed to build AI, while the companies developing the models and applications that are supposed to generate the ultimate economic returns are still losing substantial amounts of money.

That imbalance could leave the AI investment cycle increasingly dependent on continued access to capital rather than on revenue generated by end users.

Torsten Slok, chief economist at Apollo, highlighted the mismatch in a blog post on Friday - first reported by Fortune, dividing the AI value chain into four broad segments: models and applications, cloud and compute, energy and grid, and silicon and equipment.

Using data from PitchBook and Bloomberg covering companies including OpenAI, Anthropic, Microsoft, Amazon, Constellation Energy, Nvidia, AMD and Micron, Slok found a striking difference in profitability across the industry.

The silicon and equipment segment, which includes chipmakers and other semiconductor suppliers, had an operating margin of about 41%, the highest in the AI ecosystem. By contrast, the models and applications segment had an operating margin of negative 59%.

The result turns the conventional logic of a technology supply chain on its head. In many industries, the companies closest to the end customer capture the strongest margins because they control the products and services for which consumers and businesses ultimately pay. In the current AI cycle, much of the financial value is instead accruing to the companies selling the chips, servers, networking equipment and other infrastructure required to build and operate AI systems.

Slok's concern is not that AI lacks commercial value. It is that the current level of infrastructure spending may be running ahead of the revenue that AI applications are capable of generating.

"AI boom's profits are currently being funded by investors rather than earned from customers," Slok said. "The upstream margins are real, but they are paid for out of capital raised by the layer losing money, not out of cash generated by end demand."

That distinction is critical.

Nvidia, AMD, memory manufacturers and data-center infrastructure providers can record substantial revenue as cloud companies and AI developers spend aggressively on computing capacity. But their customers must ultimately generate enough cash from AI services to justify those expenditures. If that does not happen, the economic pressure moves back up the supply chain.

The potential vulnerability has become more important as the scale of investment expands.

Goldman Sachs expects AI investment to exceed $1 trillion in 2026. The spending encompasses semiconductors, data centers, electricity generation, networking equipment and cloud infrastructure, creating a powerful investment cycle that has benefited a wide range of technology and industrial companies.

The problem is that the financial returns from AI applications have so far not expanded at the same pace.

Slok argues that there has been limited evidence of a broad increase in productivity or corporate profit margins attributable to AI outside the largest technology companies. That creates a widening gap between the capital being deployed to build AI infrastructure and the economic returns being generated by the applications using that infrastructure.

"The bottom line is that the most profitable part of the AI value chain depends on the least profitable part continuing to grow revenue or raise capital," Slok wrote. "Capital can bridge the gap for a while, but not indefinitely."

The Bank for International Settlements has raised a similar concern.

In its annual report published in June, the BIS warned that AI investment by major hyperscalers was running ahead of earnings and free cash flow, prompting the companies to rely more heavily on debt financing.

The scale of that borrowing has already increased sharply. A Bank of America analysis found that the five largest hyperscalers issued about $121 billion of debt in 2025, roughly four times their average annual issuance during the previous five years.

The concern is not simply that some AI companies could fail. It is that a slowdown in AI investment could propagate through a highly interconnected financing and supply chain.

"If disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust," the BIS warned, "with potential knock-on effects on financial conditions."

That scenario would put pressure on semiconductor manufacturers, equipment suppliers, data-center operators, power producers and other businesses that have expanded capacity based on expectations of sustained AI spending.

Oracle provides one of the clearest examples of the financing risk.

The company has committed enormous resources to data-center infrastructure and has become a major infrastructure partner for OpenAI. At the end of fiscal 2026, Oracle had negative free cash flow of about $23.7 billion, according to technology writer Ed Zitron, alongside nearly $130 billion of debt and roughly $260 billion in uncommenced lease commitments for AI infrastructure.

Those future commitments have become relevant because they illustrate how much of the AI infrastructure build-out has yet to appear as conventional balance-sheet debt.

Oracle has said its uncommenced data-center leases generally begin between fiscal 2027 and fiscal 2029 and run for 15 to 19 years. The company has also warned that the duration and pricing of those leases may not match the length of its customer contracts.

That creates a form of duration risk. Oracle can commit to infrastructure for decades while its customers may have contracts that last considerably less time. If demand weakens, the infrastructure cannot necessarily be scaled down as quickly as the financial commitments supporting it.

OpenAI is central to that equation. Oracle signed a $300 billion deal with the AI company last September, creating a major expected source of future demand for its infrastructure.

But the broader risk is larger than any single customer.

The AI infrastructure cycle depends heavily on the continued spending of Microsoft, Alphabet, Amazon and Meta, alongside other major technology companies. These companies are among the world's largest buyers of GPUs, memory, networking equipment and data-center capacity.

If they conclude that the returns from AI are insufficient, even temporarily slowing capital expenditure could have an outsized impact on suppliers.

"If Microsoft, Google, Amazon, and Meta decide that it's time to stop spending $30 billion or more a quarter on GPUs, RAM, storage, and data center construction," Zitron argued, "that'll tear a hole in the side of what people assume is a permanent supercycle."

This is where the AI boom begins to resemble a traditional capital cycle.

When expectations of future demand are strong, companies invest ahead of actual demand. Suppliers expand capacity, investors provide financing, and asset prices rise. That investment generates further orders, reinforcing the perception that demand is structurally increasing.

The cycle can become self-reinforcing on the way up. But it can also work in reverse.

A reduction in expected AI returns could cause hyperscalers to slow capital expenditure. That would reduce orders for chips and equipment, leaving suppliers with excess capacity. Data-center operators could then face lower utilization while still carrying large financing and lease obligations. Lower cash flows could make debt more expensive or harder to refinance, creating additional pressure to reduce spending.

That does not necessarily mean an AI crash is imminent. It does, however, mean that the sustainability of the current boom increasingly depends on a crucial question: Can AI applications generate enough revenue and productivity gains to justify the extraordinary infrastructure investment being made today?

There are reasons for investors to remain cautious about treating infrastructure demand as proof of end-market demand. The rapid increase in AI computing capacity demonstrates that companies are willing to spend heavily on the technology. It does not, by itself, demonstrate that those investments will earn attractive returns.

The distinction matters to the semiconductor industry. Chipmakers can generate extraordinary profits while AI developers continue to lose money because the former are paid immediately for infrastructure while the latter must spend heavily before establishing durable business models.

That makes the current AI boom unusual. The most profitable companies are effectively monetizing the investment cycle itself, while the businesses expected to create the ultimate economic value are still building their customer bases and searching for sustainable margins.

Against the backdrop of heavy investment in AI infrastructure, analysts believe the next stage of the AI cycle will therefore be judged less by how many GPUs are installed or how many data centers are announced and more by whether companies can convert that infrastructure into recurring revenue and durable free cash flow.