AI Gold Rush or Bubble?

AI Infrastructure Boom: Déjà Vu?

Last week, The Wall Street Journal (WSJ) published an article discussing the current artificial intelligence (AI) infrastructure surge. As an investor, it's hard not to feel both amazed and apprehensive, along with a strong sense of déjà vu. The WSJ reports that Ellendale, North Dakota, a small town of only 1,100 residents, is now home to an AI data center under construction that will be bigger than 10 Home Depots. This single data center, costing over US$15 billion, accounts for a quarter of North Dakota's annual Gross Domestic Product (GDP).

While questions remain whether the investment justifies the returns, the sheer scale of the AI expansion causes reflection. AI advocates call it the Fourth Industrial Revolution – the continuing transformation of industries and societies through AI, robotics, the Internet of Things (IoT), blockchain, quantum computing, and biotechnology, all built upon the digital groundwork of the Third Industrial Revolution, which focused on computers and the internet.

Recouping Investments: A Big Question

The crucial question, as the article rightly points out, is: How will investors ever recover these massive investments, and when? The scale of this technological endeavor is like building offices 'on spec,' and it is truly astonishing. Over the last several years, major tech companies have committed to data centers, chips, and energy at levels never seen before. Microsoft's Satya Nadella hopes adoption won't take as long as electricity's 50-year rollout. Meta's Mark Zuckerberg suggests they're investing as if it won't, highlighting Meta's potential US$600 billion U.S. investment through 2028. But, in the end, it's a risky bet that AI will advance quickly enough to transform the economy and produce profits.

The WSJ draws parallels to the dot-com boom of the late 1990s, when telecoms invested over US$100 billion in fiber optics across the U.S., only to fail spectacularly. Giants like WorldCom and Global Crossing collapsed due to excessive building. Investors today should remember that chips and data centers, now a nearly trillion-dollar industry, were once considered unexciting.

CoreWeave: A Gold Rush Story

The WSJ describes CoreWeave's Ellendale project as representative of the current gold rush. Only six years ago, it was a small crypto mining operation with less than two dozen employees. Now, fueled by Wall Street funding, it's valued higher than General Motors or Target. After ChatGPT's release, the company shifted from crypto to AI cloud computing, embracing a culture of “YOLO” (You Only Live Once) and “GSD” (Get Stuff Done). It has secured over US$42 billion in contracts, including an expanded agreement with OpenAI. However, this growth is fueled by debt – US$15 billion in loans with rates above eight percent, plus US$56 billion in long-term data center leases.

CoreWeave's tech contracts are shorter-term (2-5 years). Aligning short-term revenue with long-term liabilities creates risk if demand decreases. Overbuilding or tenant departures could turn CoreWeave's offerings into the dark fiber cables of the 2020s. While a new AI version (2.0) might emerge, AI 1.0 needs to be dismantled first.

Sequoia’s David Cahn estimates that AI infrastructure investments in 2023-2024 alone will need US$800 billion in product sales over the chips’ 3-5 year lifespan to yield solid returns. Bain & Co. projects US$2 trillion in yearly AI revenue by 2030 – exceeding the combined sales of Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia today, and five times the global subscription software market. Where will this revenue originate? Which customers can afford it after spending with Amazon, Apple, Alphabet, Microsoft, and Meta?

Morgan Stanley estimated AI revenue last year at just US$45 billion, primarily from chatbot subscriptions and cloud access. Consumers enjoy free AI tools, but businesses are reluctant to spend more than US$30 per user monthly for services like Microsoft’s Copilot. Investor Roger McNamee believes this bubble surpasses all previous tech bubbles. He argues that even success won't justify the current expenditures.

Not the Dot-Com Bubble, But...

This situation differs from the dot-com bubble. Today’s hyperscalers (Alphabet, Microsoft, Amazon, Meta) generate substantial cash flow, unlike the fiber firms of the 1990s. Additionally, OpenAI’s ChatGPT boasts 700 million weekly users (up from 500 million in March), with revenue tripling in 2024 to US$13 billion. However, that’s small compared to Oracle’s US$60 billion annual commitment and over US$1 trillion in data center plans.

There's a contradiction: some suggest that AI's displacement of white-collar jobs could justify the spending through savings. But if numerous white-collar jobs are lost, who will have the means to purchase AI-driven products and services? Henry Ford raised his factory workers' wages because they were also his customers. If AI eliminates jobs, who will become the customers?

Debt and the Tech Cycle

The financial structure involves intricate webs of debt. Hyperscalers are reportedly considering US$400 billion in 2025 capital expenditures, surpassing the Apollo program's cost adjusted for inflation. Some build their own data centers, while others lease from intermediaries like CoreWeave, who equip them with Nvidia chips. In Ellendale, CoreWeave leased from Applied Digital, another former crypto company now focused on AI.

One cause of bubbles is the recurring investor belief that technologies changing history must be profitable for investors. History is filled with technologies that benefited consumers more than investors, including canals, railroads, electricity, air travel, automobiles, and television. The excitement around potential fuels overbuilding, followed by losses. Consumer demand allows technology to persist, but most investors in related providers do not profit. In fact, historical data indicates that a majority of investors will lose money.

The WSJ references Andrew Odlyzko’s concept of “collective hallucinations” during manias, where hype outweighs risks – similar to the dot-com fiber rush, where traffic growth was overestimated, leading to collapse. Executives at companies like Level 3 spent lavishly, only to see their stocks plummet 95 percent and fiber remain unused until streaming emerged. Currently, Applied Digital has started construction on another data center without a tenant. An MIT study we previously covered found that 95% of companies have zero ROI from their AI projects. A Chicago study also revealed that chatbots do not increase earnings in Danish workplaces. The release of ChatGPT-5 in August was underwhelming, despite significantly higher training expenses. Moreover, chips become obsolete faster than resilient fiber. The AI race presses on, fueled by a continuous flow of capital. Ellendale’s Mayor Don Flaherty stated that “we’re on the wave right now, and we’ve just got to keep riding it.”