For many investors, the huge increase in spending on the artificial intelligence buildout is proof that there is no bubble in the stock market. But at GQG Partners, this is more evidence informing their belief that this boom will give way to a bust, just like with telecom in the 1990s and shale oil a decade ago.

The portfolio managers and analysts at GQG, which manages $160 billion, say the huge orders for semiconductors, memory hardware, and related data center spending in recent quarters are based on unproven hopes for future demand and pricing for AI businesses and usage. They point to evidence of overspending, such as recent developments around potential price wars—with business users tightening budgets as they seek cheaper ways to run AI models—and slowing gains in the effectiveness of new large language models. Additionally, they think the continued use of circular investments and opaque accounting methods raises risks.

Even the massive orders for semiconductors, which have powered huge gains in chip stocks, are a “feature, not a bug,” according to GQG portfolio manager Brian Kersmanc. “Bubbles are often characterized by real demand in the near term, massive capital deployment, investor belief that demand is effectively unlimited, and weak visibility into long-term returns. Strong chip orders do not disprove a bubble; they may actually be the strongest signs of it,” he says.

GQG’s fund managers and analysts are not perma-bears on technology or the stock market overall. As recently as 2024, upwards of 70% of its portfolios were in tech or tech-adjacent stocks, such as Uber. In 2017, the firm made a bet on then out-of-favor semiconductor chips, including Nvidia NVDA.

But starting in late 2024 and into early 2025, GQG positioned its portfolios away from the AI technology and infrastructure stocks that have staged huge rallies and helped lift the overall stock market. The bet has hurt the performance of its strategies, such as the Gold-rated GQG Partners US Select Quality Equity Fund GQEIX, but the firm is sticking to its guns.

A key factor in driving these stocks higher (and a central argument of bulls) is the massive surge in AI-driven orders at semiconductor chip companies like Nvidia, Broadcom AVGO, and most recently Micron, or for other computer memory hardware with stocks such as SanDisk SNDK and Western Digital WDC.

However, Kersmanc says this is where investors should be skeptical, starting with the mismatch in how purchases are accounted for. “You are seeing orders and sales booked today and the costs (to the buyer) are spread over three, five, six, or even ten years,” he explains. The issue isn’t with this accounting method itself, which is standard, but with the implications, given the high uncertainty about the economics of AI usage. “The key issue is not whether demand is strong today, but whether that demand is economically justified and sustainable once return on investment, utilization, and pricing power are tested.”

GQG’s bear case is built on those fronts. They argue that capex spending is taking place far ahead of proven economic value. Kersmanc starts with an assumption being made by the big AI labs and other companies: “It’s ‘I spend this amount, and I get this amount of improvement.’ However, after ChatGPT-4, it has flattened out.”

Even as investment continues in the most advanced models (the so-called “frontier” ones), Kersmanc says there is growing evidence that many AI developers and businesses are going in the other direction. Rather than using large language models like Claude, they are gravitating toward small language models. These are generally trained on limited sets of data for specific texts, and they therefore need less compute power. For example, “if you want live translation, you can do that through an SLM and translate on your phone, with no token generation, and you don’t have to ping a data center.”

If businesses prefer SLMs for their solutions, “the broader implication is that the market may not need nearly as much frontier-scale compute as investors currently assume,” Kersmanc says. “If AI development is moving in that direction, then the case for massive data center and GPU spending weakens.”

At the same time, developers are increasingly using open-source Chinese models that are cheaper to train and come with significantly lower infrastructure capex. “China has 500 data centers planned, and the US has 5,500,” Kersmanc says. This comes against a backdrop of companies looking to restrain ballooning AI usage costs.

Kersmanc explains that another piece of the puzzle has been how hyperscalers have been accounting for their data center buildouts. He points to Meta Platforms’ META joint venture with Blue Owl Capital OBDC announced last year, which enabled the company to transfer a $30 billion data center project in Louisiana off its balance sheet. Meta has also been classifying some substantial infrastructure assets as “construction in progress.” That line item doubled between 2024 and 2025.

These accounting questions tie back to the mismatch between spending on chips and other hardware, Kersmanc says. “If large amounts of hardware are sitting in construction-in-progress or similar balance sheet categories, investors may not be able to tell how much of that spending is actually deployed and earning a return. That matters because companies can spend huge amounts of cash upfront while their income statements reflect the cost only gradually. So the concern is not just accounting presentation; it is that the accounting may be masking overbuilding, underutilization, or weaker returns.”

Then there are the uncertainties around the ability to build the data centers being ordered. Backlash against data center infrastructure is growing, thanks to these buildings’ heavy use of electricity and water, as well as their other impacts on communities. “Roughly half that were supposed to be completed haven’t even started or were cancelled,“ Kersmanc says.

Putting it all together, “the AI buildout shows many classic signs of a bubble,” Kersmanc says. “Investors are still extrapolating demand much further than the underlying economics seem to justify.”