artificial intelligence Bubble at $2.1T Is 2025 the Breaking Point? What Happens Next

For now, markets are still willing to wager enormous sums on artificial intelligence. But beneath the confidence, questions are growing louder about whether the boom is sustainable and what happens if expectations finally collide with reality.

Some of the biggest names tied to the AI surge, including Nvidia, Oracle and CoreWeave, have seen their valuations slip from the peaks of mid 2025. Even so, the US stock market remains heavily concentrated around AI driven growth. Roughly 75 percent of the S&P 500’s gains are attributed to just 41 AI linked stocks. The so called magnificent seven, Nvidia, Microsoft, Amazon, Google, Meta, Apple and Tesla, account for 37 percent of the index’s overall performance.

That level of dominance, largely built on a single approach to artificial intelligence known as large language models, has intensified fears that the market is inflating an AI bubble.

Industry leaders dismiss the idea.

“We are long, long away from that,” said Jensen Huang, chief executive of Nvidia, last month.

Not everyone shares his confidence. For investors watching unprecedented spending deliver limited profits so far, patience is being tested.

The risk is not just to venture capitalists. According to Gary Marcus, AI scientist and emeritus professor at New York University, the fallout could be far wider.

“If a few venture capitalists get wiped out, nobody’s going to be that sad,” he said. “But when a large part of US economic growth is driven by AI investment, the blast radius is much bigger.”

“In the worst case,” Marcus warned, “the whole economy starts to unravel. Banks lose liquidity, bailouts follow and taxpayers pick up the bill.”

A trillion dollar gamble

The warning signs are not subtle. Microsoft, Amazon, Google, Meta and Oracle are expected to spend around one trillion dollars on AI by 2026. OpenAI alone plans to commit roughly 1.4 trillion dollars over the next three years.

So far, returns lag far behind spending. OpenAI, the company behind ChatGPT, is expected to generate just over 20 billion dollars in profit in 2025. That is a significant sum, but nowhere near enough to justify investment on the scale being deployed.

At the heart of the boom is how modern AI has been built.

Building computer cities

The AI revolution arrived in early 2023 with the release of ChatGPT 4. Its leap in language, coding and image generation ability stemmed from a single idea, scale.

GPT 4 required between 3,000 and 10,000 times more computing power than GPT 2. It was trained on vastly more data, jumping from 1.5 billion parameters to an estimated 1.8 trillion. That meant ingesting nearly all available text, image and video data online.

The improvement was so dramatic that many concluded artificial general intelligence would emerge simply by repeating the process. Demand for advanced graphics processing units surged and Nvidia’s valuation soared.

To house these machines, vast new data centres followed. Bulldozers moved quickly.

The Stargate project, announced in January by Donald Trump, OpenAI’s Sam Altman and partners, already has two massive data centre buildings operating in central Texas. By mid 2026, the site is expected to span an area comparable to Central Park in New York.

Even that scale is being eclipsed. Meta’s 27 billion dollar Hyperion facility in Louisiana is closer in size to Manhattan itself and is projected to consume twice the electricity of nearby New Orleans.

This explosive demand is straining the US power grid, with some data centres waiting years for connections. For firms like Microsoft, Meta and Google, the solution may be to build their own power generation. The question is whether these vast AI systems will ever generate profits proportionate to their cost.

The problem of ageing chips

Unlike roads or power networks, AI data centres face rapid obsolescence. Investors lack reliable depreciation models for facilities that barely existed five years ago.

Nvidia releases more powerful chips almost annually and claims its latest processors will last three to six years. Critics are sceptical.

Michael Burry, the investor made famous by The Big Short, has recently bet against AI stocks. His view is that chips may need replacing every three years, or even sooner as competition accelerates. Cooling systems, wiring and switching infrastructure also degrade, often requiring replacement within a decade.

The Economist estimated that if AI chips lose their edge every three years, the combined value of five major tech firms could fall by 780 billion dollars. If the cycle shortens to two years, losses could reach 1.6 trillion dollars.

Those figures widen the already vast gap between AI spending and plausible revenue. Some estimates suggest big tech would need to generate two trillion dollars in profit by 2030 to justify current investment levels.

Where are the profits

AI usage is rising. Social media is flooded with AI generated text, images and video. Students use it for homework, parents for research and professionals for drafting documents.

Yet beyond novelty and convenience, profits remain elusive.

There are early signs of promise in areas like software development, drug discovery, creative industries and online commerce. OpenAI reports 800 million weekly active users, double the figure from February.

Only five percent pay for the service.

Business adoption tells a similar story. At the start of 2025, just 8 to 12 percent of US companies reported using AI to produce goods or services. Among larger firms, adoption rose to 14 percent in June but has since slipped back to 12 percent. McKinsey analysis shows most companies are still running pilots or struggling to scale deployment.

That caution is understandable. Generative AI is still a young technology, even for its creators. The question is how long shareholders will tolerate losses while waiting for demand to catch up.

Is scaling hitting limits

Large language models continue to improve on technical benchmarks, with performance roughly doubling every six months as computing power increases. But real world gains are less convincing.

These models predict likely answers based on statistical patterns. They do not truly understand meaning or possess long term memory. As a result, they excel in narrow tasks yet repeatedly fail in others.

“I don’t think simply scaling by 100 times transforms everything,” said Ilya Sutskever, co founder of OpenAI, speaking on the Dwarkesh Podcast. He suggested the field is returning to a research driven phase rather than a guaranteed path to breakthroughs.

For critics like Marcus, that admission is crucial.

“It’s a scaling hypothesis, a guess that this would work,” he said. “You are spending trillions, profits are tiny and depreciation is huge. At some point, the market realises that does not add up.”

The question facing investors is not whether AI will improve, but whether modest progress will be enough to justify the biggest technology bet in history.