We are entering the tail end of Wall Street’s second-quarter earnings season, and, as usual, the majority of companies have managed to beat already sky-high expectations. This is largely thanks to strong earnings growth, with S&P 500 earnings per share now at record highs.

And analysts don’t anticipate a slowdown. They expect S&P 500 earnings to grow faster than 27 percent in the next 12 months, driven largely by the so-called hyperscalers, including Alphabet, Microsoft, Meta and Amazon, as well as semiconductor giant Nvidia, which are all benefiting from the AI boom.

If that is correct, EPS would be more than 85 percent above trend next year and 100 percent above trend by the middle of 2028. Notably, earnings have never been more than 44 percent above trend since the end of World War Two.

The problem with these earnings projections is that the business model of the hyperscalers driving them has materially changed in the last few years.

In the past, these huge US tech firms were capital-light businesses with large “moats” driven by high switching costs. This, in turn, helped them generate supernormal earnings for more than a decade without the dreaded mean reversion one would expect from increased competition.

But now the hyperscalers have become highly capital-intensive businesses, because of the massive AI infrastructure spending spree and sections of those moats have gotten a lot slimmer.

This reflects the fierce competition across the AI value chain.

Hyperscalers typically have long-term agreements with specific developers of large language models (LLMs), most notably Anthropic and OpenAI.

Currently, LLM users appear willing to switch between AI platforms depending on the quality and cost. For example, when OpenAI's ChatGPT was overtaken in performance by Anthropic's Claude in March, users could switch with little friction, based on data from the Ramp AI Index about changes in AI spending share. Such consumer moves are expected to continue as AI rapidly develops, which will result in a shift in data center load from one provider to another.

This means that hyperscalers are either at the mercy of users’ preference for one specific model provider, or they will have to compete with other data center providers for the business of each new market leader. In the former case, their revenue growth will slow if their partner model falls behind. In the latter, their margins will shrink from increased competition.

Alphabet may be in a somewhat advantaged position here, since it has its own competitive model that it can run on its own stack, though it currently appears to be getting more closely linked to Anthropic, based on reports from the Information in June.

The result of all this may not just be negative free cash flow, which is already on the horizon (or already here, in certain instances), but negative net profits, unless the hyperscalers cut back their capex.

And if they cut back on capex, we can expect a large drop in revenue for the semiconductor firms that have benefited from all this spending – and thus a decline in earnings in that part of the market.

This is just one example of the codependencies among companies in the AI ecosystem. In its Annual Economic Report 2026, the Bank for International Settlements (BIS) analyzed the revenue of hyperscalers and semiconductor companies, separating circular financing arrangements — where companies from different parts of the AI value chain are financing each other — from true arm’s-length contracts. They found that in 2025, over half of hyperscalers’ revenue and almost all of the chipmakers’ revenue could be traced to circular financing arrangements.

If that merry-go-round stops because hyperscalers cut back on capex or their margins shrink, earnings for the S&P 500 overall could slip, potentially dropping back toward trend.

Such an earnings inflection is not inevitable. AI optimists and many equity analysts will argue that this technology will continue to grow exponentially, creating hundreds of billions of dollars in additional revenue that will justify hyperscalers’ massive spending.

This is entirely possible. After all, AI is revolutionizing the world, and new use cases emerge with every model update.

On the other hand, the cost of building data centers is skyrocketing, forcing most hyperscalers and many others to constantly increase their projected capex. Two numbers put the challenge of making money with newly built data centers in stark contrast.

Nvidia CEO Jensen Huang recently estimated that the cost of building a 1-gigawatt (GW) data center could soon reach US$80 billion to $100 billion. This compares with an estimated $10 billion to $12 billion in revenue generated by a 1-GW data center running AI models, as estimated by Cleanview, a market intelligence firm focused on US power infrastructure, and Lancium, an energy tech and infrastructure company.

If accurate, this means that it would take eight to 10 years for a newly built data center to amortize its initial costs, let alone pay for operating expenses or make a profit. Clearly, this is longer than the reasonable life of a cutting-edge GPU or TPU.

Taken at face value, that turns new data centers into a very unattractive proposition.

If hyperscalers start to conclude that this is indeed the case, this could send today’s entire investment boom into a tailspin.

---

The writer is an investment strategist at independent investment bank Panmure Liberum. The views expressed are personal.