Generative Artificial Intelligence: Catching The Bubble On The Bounce

The moment of truth is approaching.

I happened to live through the rise and fall of the dot-com bubble from the inside: I was running communications for Alcatel Optics, Alcatel’s global optical networking division, which at the time was the worldwide leader in building the “information superhighways,” the physical infrastructure of the new digital universe¹.

I remember as if it were yesterday the successive executive committee meetings at Alcatel Optics where we dealt with the wave of bankruptcies hitting our customers around the globe. That was followed by endless internal and external crisis communications around the closure or sale of seventeen of our sites and the elimination of 16,000 jobs (excluding the teams at divested sites) across several continents. The bursting of the dot-com bubble was anything but a minor episode.

Like the technology and consumer sectors, the telecom market² was shaken up by one bold startup after another. The most iconic was 360networks, a Canadian company that aimed to build³ and operate, by mid-2002, a high-capacity optical network connecting the world’s top 100 cities. When it went public in April 2000, 360networks was hailed as a star among a sector that already included long-established giants. Fifteen months later⁴, it filed for bankruptcy. That said, without the massive investments that went into building the physical Internet network, we wouldn’t enjoy the online experience we have today⁵.

This personal experience is what led me, long ago on Superception, to point out that we are once again stuck in a bubble — this time driven by generative artificial intelligence.

The similarities and differences between the dot-com bubble and the generative AI bubble

Generative AI shares many elements with the dot-com bubble, starting with the “irrational exuberance” – Alan Greenspan’s famous expression as Fed Chair during the digital revolution – reflected in startup valuations. Today, this exuberance extends both to companies like OpenAI, which do generate revenue, and to others like Safe Superintelligence (Ilya Sutskever) or Thinking Machines Lab (Mira Murati), which have neither products nor customers. That hasn’t stopped Mira Murati from being a paper billionaire.

The trajectory of the Nasdaq Composite underscores the parallel: It’s up 106% from its December 2022 low (ChatGPT launched on November 30, 2022), almost identical to the 111% gain over a comparable period starting in 1995 (see chart below). As for the S&P 500, now heavily influenced by tech stocks, it trades at 24 times forward 12-month earnings, a premium that’s only been reached 0.8% of the time since 2010.

The astronomical valuation of Nvidia – the first company to cross the $4 trillion market-cap threshold on July 10 – doesn’t strike me as reassuring. Nvidia’s chips are at the heart of today’s infrastructure spending spree, making its valuation part and parcel of the bubble. It tells us nothing about the real-world value and profitability of generative AI applications. Treating Nvidia as proof of the sector’s financial health would be like claiming that a surge in robot sales to an automaker guarantees the profitable sale of its cars to consumers.

Another striking similarity between the dot-com bubble and today’s AI bubble is the market logic: The more money companies lose, the brighter their future seems to investors. The theory is that today’s losses fuel unstoppable future growth. Once again, OpenAI is the poster child of this collective – pun intended – hallucination.

The two biggest differences between the current bubble and the late 1990s are the very low number of IPOs, for multiple reasons, and the scale of the talent wars. Top AI experts are being courted with contracts more typical of professional sports stars.

How bubbles work — and why they have long-term benefits

The reference book on financial bubbles is Technological Revolutions and Financial Capital by Venezuelan-British economist Carlota Perez. Published in 2002, just after the dot-com crash, it predicted the coming golden age of the Internet – at a time when many observers were dismissing it. Carlota Perez based her argument on the observation that nearly every major transformation – the Industrial Revolution (1771), steam and railways (1829), steel, electricity and engineering (1875), oil, autos and mass production (1908), information and telecoms (1971) – was accompanied by a financial bubble.

Carlota Perez identifies four phases in each roughly 50-year innovation cycle:

• A period of foundational investment in new technologies.
• A speculative frenzy where investors chase ever-higher returns in a volatile, consolidating market.
• A crash, followed by broader consolidation and market correction (often including regulatory responses to excesses), leading to a golden age of integration of the new technologies at the core of society.
• Finally, once those technologies mature, the emergence of a new revolution, financed by investors, which restarts the cycle.

The companies that survive the crash and consolidation have sound business models that secure their future without extravagant funding rounds.

The key value of every bubble is that it funds the over-investment needed to build out the infrastructure for new technologies. Those systems can’t be profitable until enough customers adopt them, so bubbles give investors the prospect of capital gains to offset the temporary absence of operating profits and dividends. That’s how the dot-com bubble gave rise to the telecom networks that have proven indispensable ever since.

Thanks to that bubble, the digital revolution triggered an unprecedented wave of innovation. Intel engineers once calculated what Moore’s Law progress would have meant for cars: By 2015, a Volkswagen Beetle would have cost four cents, hit 300,000 mph, and traveled 440,000 miles on a single liter of gas⁶.

Today’s increasingly obvious generative AI bubble

This year alone, Amazon, Google, Meta, and Microsoft are expected to invest more than $340 billion in AI. Banks and private-equity firms are competing to fund projects. As analyst Ben Evans has pointed out, about half of this spending is financed by ad revenues⁷.

Looking at Morgan Stanley’s projections, AI leaders are on track to spend $2.9 trillion building data centers by 2029. They’ll cover $1.4 trillion themselves, leaving $1.5 trillion to be financed by others.

Meanwhile, the drumbeat of disappointing AI results, relative to the size of these investments, has grown louder in recent weeks.

S&P Global Market Intelligence found that 42% of U.S. and European firms it surveyed have abandoned most of their AI initiatives, up from 17% a year ago.
McKinsey reports that while nearly 80% of companies claim to use generative AI, they see no impact on profits. And MIT research shows that only 5% of pilot AI projects generate millions in added value, while the vast majority stall out with no measurable bottom-line impact.

Sam Altman, ever the skilled operator, didn’t miss the moment. Just days ago, at a dinner with reporters, he admitted the AI sector is indeed in a bubble. The confession is especially telling coming from the one person who has most fueled the bubble – through relentless hype and through OpenAI’s own services. The timing is revealing too: It came right after the launch of GPT-5, billed as a major leap over GPT-4 (otherwise it could have been called GPT-4.6), but which left observers underwhelmed.

At the same time, Altman has walked back earlier claims about the imminence of artificial general intelligence, now saying we don’t even know how to define that threshold – which is true, but not something he used to emphasize when hyping its near-term arrival.

Not to worry, though: In the same breath, Altman said he expects OpenAI to invest trillions of dollars in AI infrastructure. His talk of a bubble seems designed more to discredit competitors and kick off the coming consolidation phase than to admit a misjudgment.

Generative AI’s benefits today: reserved for a privileged few

Whatever happens for corporate users, generative AI has already paid off for one group: Entrepreneurs and some of their employees.

We are witnessing the largest wave of personal wealth creation in history. There are now 498 AI “unicorns” worth a combined $2.7 trillion, including 100 founded since 2023. If current funding rounds close, 29 startup founders will split $71 billion this year alone – a record.

The latest example: Current and former OpenAI employees are reportedly preparing to sell about $6 billion worth of shares in the privately held company – valuing it at $500 billion – to a group of investors led by SoftBank and Thrive Capital. That would place OpenAI among the world’s 20 largest companies by market cap, despite its massive losses. And just recently, Leopold Aschenbrenner, a 23-year-old former OpenAI researcher, raised more than $1.5 billion for his hedge fund Situational Awareness, despite having no investment track record and having been fired by Altman for allegedly leaking sensitive information.

One can hope – and maybe believe – that today’s over-investment in AI infrastructure will prove as useful over the coming decades as the digital revolution’s over-investment has over the past twenty years.

The question is which companies will emerge from the coming consolidation as winners. As with the digital revolution, we should expect some long-term surprises from the players able to catch the bubble on the bounce.

—

¹ It’s worth noting that, today, Meta owns a significant share of the undersea cables that make up the global optical network – without which the digital world couldn’t function.

² The telecom sector experienced its own bubble, which was essentially a subcategory of the dot-com bubble.

³ By laying its own fiber and swapping capacity with other operators.

⁴ In June 2001.

⁵ In telecommunications, the legacy of the Internet bubble is very positive: The heavy over-investment in the 1990s to build terrestrial and undersea optical networks ultimately provided an international infrastructure that enabled affordable broadband communications. That said, performance still needs improvement – especially reducing latency, which prevents several future applications from reaching their full potential.

⁶ Anecdote told by Brian Krzanich, then Intel CEO, during the event celebrating the 50^th anniversary of Moore’s Law – the empirical projection that the number of transistors on a microprocessor, and thus its computing power, doubles every two years.

⁷ At a high level, the global advertising market is worth about $1 trillion, with Alphabet accounting for 20-25%, and Meta and Amazon together for roughly another quarter.