Why the AI Bubble Is Poised to Burst, According to GQG’s Rajiv Jain

Is the artificial intelligence boom just getting started, or is it an overinflated balloon poised to burst? One top stock fund manager is positioned for an AI stock collapse that will make the 1999 dot-com bust pale in comparison.

“One of the lessons of navigating these kinds of bubbles is you’re either early or late. If you’re late, the losses are horrendous,” says Rajiv Jain, chair and chief investment officer at GQG Partners. GQG runs an array of Gold-rated funds, including GQG Partners Global Quality Equity GQRIX, GQG Partners Emerging Markets Equity GQGIX, GQG Partners US Select Quality Eq GQEIX, and Goldman Sachs GQG Ptnrs Intl Opps GSIMX.

Jain came early to the view that today’s stock market is “much worse” than the dot-com bubble, which he acknowledges has “already cost us performance in the short run.” But he maintains that AI is a bubble that will burst, and that when this happens, it will hit a range of sectors across the stock market—including industrials, like steel companies buoyed by data center buildouts.

Jain shared his thesis in a white paper. In a recent interview with Morningstar, he discusses why he sees warning signs in profit margins for the most advanced semiconductor chips and the growing use of debt to finance the AI boom, along with what could cause the bubble to burst.

Read the following excerpts from our conversation for more. And for the opposing view that this is a boom and not a bubble, check out our interview with Alger Funds CEO Dan Chung.

Leslie Norton: You published a white paper about why you thought an AI bubble was forming and poised to burst. Are you anti-tech?

Rajiv Jain: No. In the first quarter of 2023, we were the single largest institutional buyer of Nvidia NVDA stock. It was at 60 times forward earnings. Until the summer of 2024, some clients would say, ‘Don’t you have too much in tech?’ We started getting worried. This year, almost all the data began pointing in the other direction. We grew concerned that the data doesn’t jibe with what people are saying.

I lived through the dot-com bubble, the Hong Kong property bubble, and the commodity supercycle. There’s always a narrative that takes hold. During the late 1990s, people said those who lived through the 1973-74 bear market didn’t know what was going on and were luddites. It was 25 years between that and the dot-com bubble, so a generation had to pass. It’s been 25 years since the dot-com peak.

We’ve cut aggressively over the last nine months. We’ve been early. It’s cost us some significant underperformance.

Norton: First, let’s talk about how you do what you do.

Jain: We look quantitatively for high-quality businesses where we have visibility three to five years out at sensible prices. A team of traditional analysts looks at the names, then we have non-traditional analysts like former journalists who talk to different sources, such as former employees, resellers, distributors, and others in the ecosystem. Sometimes you find nuggets. If you talk to the company and sellside, you get the party line, which is fine most times, but at key inflection points, things begin to differ.

During the housing bubble, although the FBI warned about mortgage fraud in 2005, Countrywide and Washington Mutual weren’t going to tell you there was a housing problem. [Countrywide was acquired by Bank of America during the 2008 financial crisis, and Washington Mutual was seized by regulators.] Because of the investment banker relationship, they’re very reluctant to say anything negative. Wall Street is an echo chamber. The sellside talks to management, management talks to the buyside, the buyside talks to the sellside, and it’s happily ever after.

Why Today’s Market Is Worse Than the Dot-Com Bubble

Norton: Tell us why you think this bubble is poised to burst.

Jain: We compared sellside notes back then to today. The narrative is that those companies were not cash-rich, and today they are. Both assertions aren’t true. I’m astonished at the gap between what the ground reality seems to be and what Wall Street is saying. Whoever’s making money off it has zero incentive to say anything but that it’s fantastic, revolutionary, or as Sundar Pichai at Google said, that it’s more important than fire or electricity. Large tech is obviously making money off it, but the multiple has expanded.

During the dot-com era, the big infrastructure builders were incumbent telecommunication businesses and global long-haul telcos and equipment names whose regulated local and long-distance franchises generated stable, utility-like cash flows that underwrote the internet and fiber infrastructure capex binge. Today’s infrastructure arms race is driven by the hyperscalers—Microsoft MSFT, Alphabet GOOG, Amazon AMZN, Meta Platforms META, Oracle ORCL—whose reported free cash is increasingly strained by data center and GPU capex, and whose overall earnings quality is propped up by extending server/chip “useful lives” to five to six years, versus the two to three we believe it will be. And they’re treating very large and increasing stock-based compensation as a non-cash item that is being added back to cash flow metrics.

This is actually worse than dot-com. In early 1999, the S&P 500 had a low-20s multiple. S&P 500 EPS growth for the five years until 1999 was over 20%. For the five years ending 2026, it’s expected to be 8%. Back then, Microsoft’s growing EPS north of 40% and revenue north of 30%. I don’t know of any company that is growing EPS near these rates, except Nvidia, and we know that is cyclical. More than half the S&P 500 is indirectly or directly linked to the AI trade. The S&P 500 itself is a growth stock.

GPU Pricing Shows Why Margins Will Come Under Pressure

Norton: What recent grassroots data supports your view?

Jain: This is where the journalistic work comes in. Let’s take GPU pricing. You can call up distributors, which have publicly listed phone numbers, and some are authorized Nvidia distributors. My question: Why is the Nvidia H200, which was released late last year, selling at a 50%-60% discount if there’s such a shortage? On Nvidia’s website, they’re selling at $40,000-plus. NetworkOutlet.com quoted $25,900 just a couple of days ago. If there’s such a shortage, why are there tens of thousands available? Nvidia’s latest and most powerful AI chips, Blackwell, are also offered at a discount.

Next, GPU rentals. Why are they in free fall? We’ve gotten quotes at under $4 per hour for Nvidia’s Blackwell GPU rentals. Would you let a $50,000 car rent for $4 if the car only has a three-to-four year life? Meanwhile, [Amazon Web Services] charges around $12-$13 for Blackwell. The bulk of new cloud growth is coming from AI startups. If they’re paying $4 per hour versus $12, then AWS can’t compete. Margins for AWS are already coming under pressure. Revenue growth was OK. Why? Because large tech is also investing in Anthropic, OpenAI, and so on. They go back and buy compute [computational resources] from these guys. Nvidia has invested in over 50 startups, which then go back and buy Nvidia chips.

Leverage Is Concerning

Norton: Tell us your concerns about leverage.

Jain: Meta did a $27 billion bond offering. It wasn’t on the balance sheet. They paid 100 basis points over what would cost to put it on their balance sheet. Special purpose vehicles happen at the tail end of the cycle, not the early part of the cycle.

Supermicro, a major partner for Nvidia, reported a revenue decline of 15% last week. I’m sorry, aren’t we supposed to be growing here? Oracle is growing revenue at 12%, and EPS declined by 2%, and this is supposed to be the best of times. And 90%-plus of Oracle’s order book is coming from Open AI. If you look at the Microsoft 10-Q earnings charge, it implies a $11.5 billion loss  at OpenAI. OpenAI says it’s investing $1.5 trillion in building computational resources, but its annualized revenue is under $25 billion.

Digital Ad Growth Is Slowing

Norton: Please elaborate on two of your other key points: slowing digital ad growth and a deteriorating competitive landscape.

Jain: These companies are fantastic. We owned them for the longest time. But they depend on advertising. Digital advertising revenue is now three-quarters of the overall advertising pie, which is growing at normal GDP. These companies dominate. About 38% of the market is digital advertising, Alphabet. Their market share has been flat for the last five years. Why? Because Walmart WMT, Netflix NFLX, Spotify SPOT, and Amazon have all come to digital advertising. Even if Alphabet’s market share remains flat, they are closer to saturation. Revenue growth on the ad side is down to 11%, without a recession. This isn’t a 15% grower anymore. To grow, capital intensity has to go up. So this business is growing in the high single digits, but capital intensity is growing north of 30%. Amazon’s capex is almost in line with AWS revenue.

We know AI data centers are margin-dilutive. Now you have a new player called Oracle that is willing to undercut everybody. The CEO of CoreWeave CRWV, another competitor, specifically said they want to gain market share against everybody. There are 200-plus cloud providers. So now GPU pricing is collapsing with rentals. Hence, we believe these companies are forced to invest in these startups so that they can sell them compute at a better margin.

Norton: What might catalyze the decline?

Jain: A lot of these names are very obese. You can’t have Nvidia at 15% of the GDP. At the dot-com peak, Microsoft was 5% of the GDP. Throughout that era, people were making wild claims about various end markets, like e-commerce growing into the trillions in only a few years. People throw out random meaningless numbers.

So, #1: Credit. A significant part of the growth in data centers is not coming from hyperscalers, but from other players. Private credit is a big part of it. You’re seeing cracks appear. Look at Oracle’s CDS. If the credit market seizes up—look at the First Brands and Tricolor bankruptcies—a lot of GPUs are being packaged as asset-backed securities. They need to get funding from the markets. Bond issuance is very heavy from the tech side. The whole narrative of these being cash-rich companies will be shown to be a myth. Based on our numbers, data center by data center, with the help of S&P, around 60% of the data centers aren’t through hyperscalers.

#2: Power shortages.

#3: If the economy slows down, technology is always cyclical.

#4: Geopolitics. Semiconductors are far more dependent on China than Wall Street believes. So if China cracks down on the use of Nvidia chips, that could impact demand meaningfully. By the way, China has an AI data center glut. Their capacity is running at 30%.

Norton: Let’s talk about the competitive threat from China. Earlier this year was the DeepSeek scare.

Jain: It’s an important factor. It tells you the whole LLM model may not work because these things would be a lot less compute-intensive. LLMs have a real hallucination problem; they sometimes generate information that sounds plausible but isn’t true, which can mislead users, erode trust in their responses, and make it hard for enterprises to pervasively adopt LLMs for streamlining high-valued business workflows.

Small language models, probably on your desktop or phone, might be the way things develop. DeepSeek showed these things can be done in a far more frugal manner and that open source is the way to go. Andreessen Horowitz has a data point that 80% of the startups now use open source, and more are coming from China. In fact, the vast majority of ChatGPT users are consumer, not enterprise. So the issues around DeepSeek haven’t gone away.

Norton: How did you deploy the money you removed from this sector?

Jain: We like regulated utilities a lot, because the real crunch is in power. Memory chips are never in a shortage forever. Demand is overstated every semiconductor cycle. Chips are being stockpiled, which is why GPU pricing has declined so precipitously. Power generation will take time. So we like power generation, utilities, P&C insurance, and healthcare.