AI investment enters the "Trough of Disillusionment", the next 18 months will be a life-or-death test!

Every large-scale wave of investment in computing infrastructure eventually reaches the same crossroads—the market starts questioning returns.

According to a Goldman Sachs research report dated November 10, current AI investments are in the "trough of disillusionment." Recent fluctuations in equity and credit markets, increasingly complex circular financing structures, and new regulatory uncertainties following the US midterm elections have collectively led the market to question the sustainability of AI capital expenditure.

Eric Sheridan, Goldman Sachs’ Head of US Internet Equity Research, said directly: "Every computing cycle I have analyzed follows the same curve—initial excitement about the product, expansion of investment scale, and then scrutiny of returns."

He emphasized that this is not the first appearance of disillusionment in this cycle. Over the past 12 to 18 months, there have already been three AI-driven market pullbacks: last autumn and winter, this summer, and now. However, he admitted: "The stakes are undoubtedly higher, as we've gone further into the cycle and the amount of invested capital has risen significantly."

Social Capital founder and CEO Chamath Palihapitiya stated in a report: "What I see is a severe mismatch between data center committed capacity, actual capacity, and returns." He warned that the next 18 months will be a "reckoning"—the market will gradually differentiate which AI business models and infrastructure commitments are sustainable and which are not.

The Fatal Vulnerability of the Neoclouds

Chamath Palihapitiya conducted a detailed analysis of the area under the most pressure—the so-called Neoclouds.

"I see a mismatch between data center committed capacity, actual capacity, and returns," he said. "This is especially critical for the Neoclouds, who have made massive commitments in GPUs, electricity, and data center capacity. The demand needed to support these commitments ultimately depends on whether end customers can derive meaningful returns from their AI spending."

Palihapitiya used the airline industry as an analogy: GPUs are the engines, data centers are the fuselage, power and land are the airports, and compute power hours are the seats for sale—an unsold compute hour is like an empty seat, gone forever. Neoclouds must maintain a breakeven utilization rate of around 70-75% and only exceed that can they become truly profitable.

He further noted that the current biggest problem lies on the demand side: enterprise AI project rollouts are difficult, there is resistance from middle management, and productivity gains are often lost elsewhere along the value chain. "Today, companies that attribute profitability or profit growth to AI are basically zero."

18 Months: The Window for Reckoning

Palihapitiya’s judgment on the time window is clear: "The next 18 months are critical. While we won’t have the final answer by then, we will lay the groundwork for a reckoning that will differentiate sustainable AI business models and infrastructure commitments from those that are not."

He explained that the key to this window is that more major players in the AI industry chain will come under the public market’s scrutiny, and investors will be able to better assess whether these companies' business models can continue to scale.

A noteworthy metric to track is: Will leading model providers continue to derive revenue primarily from "per token consumption," or will they transition to "outcome-based pricing"? "Subscription models are not the same as outcome-based pricing, but the direction is closer," he said.

Hyperscalers: Still Able to Achieve 15%+ Returns?

Despite the many challenges, Goldman Sachs holds a relatively optimistic view of the Hyperscalers.

Sheridan said, "Even so, we estimate that the six major US Hyperscalers can easily achieve a return on invested capital (ROIC) exceeding 15%." These six are Alphabet, Amazon, Microsoft, Meta, Oracle, and SpaceX, with cumulative revenues projected at approximately $1.42 trillion from 2028 to 2030.

By Q3 2026, AWS, Azure, and Google Cloud have an aggregate backlog of $1.8 trillion in revenue, year-on-year growth of about 140%. Some Hyperscalers have disclosed that about 50% of their backlog orders are expected to convert into revenue over the next two years.

"Visibility never emerges as quickly as investors would like," Sheridan noted, "but companies have made substantial progress."

Capital Markets: Cracks Have Emerged

Signals of stress in the capital markets should not be overlooked.

Goldman Sachs' Global Head of Leveraged Finance, Miriam Wheeler, pointed out that AI-related transaction activity in the high-yield debt market has slowed sharply. This summer, the market expected $60 billion in new issuances in the fall, which led to a significant widening of spreads. Over the past month, there have been virtually no new transactions in the non-investment-grade market.

"Divergence will likely remain the dominant theme in the high-yield market," Wheeler said, "in line with the dynamics playing out in the broader market landscape." She expects top developers and projects to continue to access more favorable financing, while others face bigger challenges.

Goldman Sachs Chief Credit Strategist Amanda Lynam's data reveals a deeper issue: market saturation, not financing costs, is now the biggest limiting factor. As of early November 2026, the five major Hyperscalers had issued $296 billion in debt during the year, already exceeding Goldman’s previous full-year forecast of $250 billion. In addition, these five companies have $1.7 trillion in lease commitments (75% of which have yet to take effect) and $1.9 trillion in other procurement commitments—of which $685 billion will mature by the end of 2027.

"Ratings agencies have already adopted a more cautious stance," Lynam wrote. In September 2026, Moody’s noted the Hyperscalers’ aggregate financial commitments have grown eightfold since 2023; in October, S&P stated that "the margin for error is narrowing in the AI capital ecosystem."

If Demand Fails to Materialize

Palihapitiya has also considered the worst-case scenario, but he does not expect a 1999-style crash.

"What’s more likely is a rationalization period," he said. End customers may refuse to expand spending; if the debt market loses confidence in utilization assumptions, funding costs will rise—"the weakest credits could face borrowing costs approaching 20%, CDS spreads could hit 1,000 basis points, and even investment-grade borrowers might face financing costs close to double digits."

He believes a more likely outcome is a redistribution of assets rather than a collapse. Hyperscalers, with their strong balance sheets and core businesses, will naturally become asset acquirers.

Sheridan agrees, adding three conditions that could trigger a cyclical slowdown: "If demand disappoints, capital access is impeded, or the AI ecosystem faces power shortages—the current trough of disillusionment could become a more persistent contraction."