Fears of an AI bubble persist this week, spurred on by more “circular” deals between OpenAI and AMD, a stock market debut for Fermi America that lavishly rewarded its hubris, and the skyrocketing fortunes of nuclear companies who, to their credit, can deliver clean firm power — but, importantly, can’t deliver that power this decade. The exuberance surely feels irrational, but every bubble is a snowflake.
Market watchers all have their preferred signals of a bubble, and many are feeling justifiably nervous of late. Jeff Bezos and Sam Altman themselves have used the word “bubble” — not in a pejorative sense, but simply to point out that any transformational technology will have its casualties from the hype.
Our concerns here at Latitude Media are not so much the fortunes of LPs at venture capital firms (who all understand they are investing risk capital), but rather the utilities, power system operators, and ratepayers who risk building out and paying for a landscape of worthless data centers and energy infrastructure whose costs have been socialized, with none of the benefits.
One can embark on a deeply rigorous analysis of the underlying fundamentals of the AI market, as Azeem Azhar has recently done. However, if you’re primarily in need of a glib set of talking points at an industry happy hour, a checklist could suffice. Here’s mine:
- How big is the role of debt in the market, and are the debt structures increasingly complex? A year ago it was easy to dismiss this question. Most infrastructure capex was from the hyperscalers, straight from their balance sheets. Their free cash flow was the envy of the world, and building data centers to power their AI ambitions seemed like the obvious best use of it. Today, though, the picture is getting cloudy. Debt is playing a larger and larger role in financing data center builds, and that debt is often tied to special purpose vehicles that do not appear on the hyperscalers’ balance sheets. Much of that debt is private capital, and therefore even harder to see. Debt is always a part of infrastructure builds, but its growing role is understandably raising alarms, as it echoes previous bubbles.
- Are vendors financing their customers’ deployments? Nvidia is definitely up to this already, and this week’s OpenAI/AMD deal has a whiff of “circularity” as well, where opaque agreements are forged between suppliers and customers. Back in the telecom bubble, vendor financing fueled the crash. It demonstrated that once customers stop placing orders, the debt on the books of vendors becomes a huge liability, and can destroy the value of their equity and start a cascade of failures throughout the market.
- Is there widespread call for deregulation, or a “light touch” from governments to let these entrepreneurs innovate and markets flourish? Of course! And President Trump himself is one of the loudest voices in this call.
- Are there IPOs from companies with no product, revenues, or much more than LOIs and artist renderings? Yes indeed.
- Is Masayoshi Son from SoftBank investing at sky high valuations? Why, yes he is. He is, to some, a leading indicator of a bubble, since he’s been party to each of them for the last 30 years.
- Are there accounting scandals among the major players as they look to keep their stock prices inflated in the face of difficult economics? Fortunately, not yet. In the telecom bubble, accounting irregularities led to the downfall of WorldCom and to some extent Nortel, further destroying faith in the market and the willingness of investors to stay in the market.
- Finally, is infrastructure overbuilt, well ahead of customers and revenues? The answer here is not so clear. In the telecom bubble companies like Global Crossing and WorldCom deployed massive amounts of fiber optic cabling and networking gear ahead of demand, assuming a new crop of competitive carriers would take it all up as internet usage exploded. That was definitely not the case, and they were left with giant networks of stranded assets. In the case of AI infrastructure, we’re in a much more supply-constrained market, not demand-constrained. Frontier AI labs aren’t looking for workloads to put on their GPU clusters; they’re looking for more chips and more power, and are able to pay for it. Energy in particular is a bottleneck, and access to power is actually putting the brakes on this market’s rapid rise. As we’ve covered in this newsletter previously, that may actually be a good thing, if it allows time for profitable use cases to develop and end user markets to mature.
There’s no escaping the fact that the AI market remains both driven by forward expectations of growth and profit, and tied to digital infrastructure (particularly GPUs) that depreciates rapidly. If demand doesn’t sustain these investments, well, we’ll have a problem.
But in the power sector, the boom may come with some silver linings. This wave has kickstarted necessary conversations around ratemaking reform, how to value and employ distributed energy resources, how to improve the design of wholesale power markets, and whether a reexamination of the values of vertically integrated utilities is in order.
What I’ve heard at recent conferences and climate weeks has generally been encouraging, and not what I would call “bubble talk.” Our industry wants to serve this new demand with clean power while protecting ratepayers. The risks around a collapse are very real — but the AI boom is already improving the outlooks for clean energy options that have been on the sidelines. So to the extent that AI doesn’t usher in a social apocalypse, there may still be reasons to root for the technology’s success.
A version of this story was published in the AI-Energy Nexus newsletter on October 8. Subscribe to get pieces like this — plus expert analysis, original reporting, and curated resources — in your inbox every Wednesday.
