Many, many years ago, when I was an analyst covering the fiber optics industry, a bubble emerged in the telecommunications industry and burst spectacularly — leaving a global network of “dark fiber” infrastructure as a stranded asset, and destroying upwards of a trillion dollars in market cap and investments in what was then called the “information superhighway.”
Lately, I’ve been wondering: are we in a similar bubble today, but at the nexus of artificial intelligence and energy? If so, how will it play out? Is there any chance that utilities, yoked to the arcane processes of rate cases, feasibility studies, environmental reviews, interconnection requests, and compliance certifications, ironically provide the necessary friction to keep this bubble from expanding beyond its ability to hold?
Let’s have a look. There are many types of bubbles in history, but infrastructure bubbles — where a great deal of stuff is built and then stranded in the wake of a price collapse — provide the right analog. There are often three reasons for this type of crash.
- Assumption of exponential growth
This is the most familiar bubble in tech. A transformational technology comes along to rewrite the rules of business, allowing for exponential growth forecasts that are difficult to argue with because of the lure of a “paradigm shift.”
We have that today in spades. The promise of artificial general intelligence (AGI) leads to more superlatives than even Marc Andreessen can muster. The signals from the dominant players couldn’t be stronger: hundreds of billions in allocated capex, investments in the startup ecosystem, and appearances in Washington, D.C. to evangelize the strategic importance of AI to the nation’s competitiveness.
The signposts of imminent collapse are typically missed revenue targets, too many undifferentiated participants, and the lack of a truly paradigm shifting use case. And as for whether we see those today — well, it’s a mixed bag.
Revenue run rates are into the billions at OpenAI and Anthropic now, even as no player at this point has a truly differentiated approach to AGI and a defensible “moat.” And DeepSeek still raises questions about whether brute force scaling by adding expensive chips will sustain any leadership in the race to AGI.
Finally, Gary Marcus still isn’t buying any of this.
- Cheap credit and speculative asset pricing
This is a hallmark of real estate bubbles, where a crash comes about as cashflows fall short of debt obligations and asset-backed securities lose value. Bankruptcies follow, investors flee the space, and the market freezes. Stranded assets abound, as they did in recent housing bubbles, and consumers are often harmed by the socialized costs of bailouts.
Debt is not cheap today, but it’s increasingly playing a role in the AI data center market, so this is a point worth watching. JLL’s recent report on data centers in North America notes that more lenders are coming into the space, including private credit debt fund vehicles and insurance companies. The real estate services firm also sees a rapid growth in asset-backed securities in the space, which is generally a good contributor of liquidity in a market, but can also lead to bubbles.
What caught my eye here, and set my bubble alarm ringing, was a recent headline in The Information: “SoftBank’s Son Goes on a New Borrowing Binge to Fund AI.” Anyone who has followed tech over the last 25 years should shudder when they read this. Consider that as far back as 1999, Son was nicknamed “Mr. Internet” because he had invested so widely in the market, at incredible valuations. He was briefly the wealthiest person in the world, then as the dotcom bubble burst, he lost 97% of that wealth, and nearly took down SoftBank with him.
But he wasn’t done. In the 2010s, Son’s Vision Fund backed WeWork and Klarna, among others, and went on to lose $27.4 billion. Now he’s into AI, backing OpenAI and the wildly ambitious Stargate project — and is reportedly borrowing another $16 billion because SoftBank’s balance sheet is tapped out.
The JLL report adds another element of concern here, noting that utilities “have been unable to decipher legitimate data center projects from land speculators.” It goes on to say “speculators have been acquiring tens of thousands of acres and requesting massive amounts of power. Once power is secured, they have been looking to sell the land for up to 10 times their original investment.”
- Regulatory or geopolitical shock
Sometimes a market is on a good run, and then new regulations or a rapid shift in national priorities make it harder to scale at the same pace. As a result, many infrastructure investors are left holding unfinished or devalued assets and walk away from the market.
In AI, this could potentially take the form of much stricter environmental regulations making it difficult to site and complete data center builds, or else cybersecurity measures on AI models or infrastructure so stringent that they stall a market’s development, leaving assets stranded. For those dominant AI companies looking to expand internationally, geopolitical tensions or outright conflicts could limit options, capping the prospects for growth.
Today, despite general political uncertainty, this feels like the least likely cause of a crash at the AI-energy nexus. The Trump administration is loath to add any additional regulations on AI infrastructure and the current geopolitical tensions between China and the U.S. are actually fueling support for AI investment, and show no signs of abating.
Thank you, byzantine energy market
The energy market itself may provide the necessary friction to limit a bubble. Many infrastructure bubbles of the past had little standing in the way of building more and more infrastructure. But AI needs power, and power markets are slow.
Power markets make you get in line and follow a process to interconnect. They have rules (and price signals) at the national, regional, and local levels. And they often socialize the cost of new infrastructure, which means you have to reckon with utility commissions, ratepayer advocates, policymakers, legislators — and generally a lot of people who have heard about paradigm shifts before and meet your excitement with an eyeroll.
Already, many utilities are introducing new tariffs for data center operators, upfront capital requirements and cost assignment, demand management and load shedding, and take-or-pay contracts to protect consumers. At the local level, the sheer size of AI data centers is raising alarms in many areas, which can delay permitting and construction.
The AI companies may see the energy system as a frustrating obstacle standing in the way of necessary speed in the race against their competitors, but it’s worth instead thinking of it in the context of bubbles. In past collapses, absent any real friction, assets were built far ahead of actual demand. In today’s AI market, though, power constraints are forcing developers to align with grid expansion timelines, reducing the risk of overbuilt, stranded data centers.
If AI demand grows sustainably, power markets may help pace investment rationally. And if demand shrinks unexpectedly, the rigid structure of energy contracts may be what prevents a total wipeout.
A version of this story was published in the AI-Energy Nexus newsletter on March 12. Subscribe to get pieces like this — plus expert analysis, original reporting, and curated resources — in your inbox every Wednesday.
