By Kennedy Maize
Just as the White House is proclaiming a national artificial intelligence industrial policy to challenge China, the frenzy for AI may be waning. Proliferating predictions of a bursting AI bubble are entirely speculative, such as a recent warning from the Bank of England. Real evidence is now increasing that the inflated expectations for AI by the Trump administration may already be deflating.
Last month (Nov. 6), Ohio-based giant utility holding company American Electric Power asked the Federal Energy Regulatory Commission to help it solve a problem. Four of its operating utilities have 750 MW of excess capacity they want to offer into a special PJM Interconnection auction in February. AEP pitches the offer as designed to help PJM meet its chronic and well-known capacity limitations that have emerged in the regional transmission organization’s regular capacity market. PJM supports the AEP move.
Not so fast, says PJM’s independent market monitor, Monitoring Analytics. According to the monitor, “While the concern of the AEP FRR Entities for the PJM Capacity Market is laudable, that is not the reason for the requested waiver.” The AEP companies — Appalachian Power, Indiana Michigan Power, Kentucky Power, and Wheeling Power — misjudged the impending demand from AI data centers by 750 MW, as the data centers turned out to be ghosts. Monitoring Analytics charges that “AEP over forecast large load additions by 751 MW, purchased capacity to serve that load and now wants to dump the resultant excess 750 MW in the PJM Capacity Market while asserting it is for the benefit of the market.” PJM’s capacity market prices are sky-high and AEP could take advantage of that as a short-term play.
While parent AEP participates in the PJM capacity market, the four separate regional utilities are defined as Fixed Resource Requirement (“FRR”) companies. They participate in all the PJM markets except the capacity market. AEP is asking FERC to approve a waiver of the PJM FRR rules to allow the four to bid into the February auction. “The waiver,” argues the market monitor, “would allow generators with fully guaranteed cost recovery paid for by their customers to undercut the capacity market in which investors take the risks.”
Phantom data centers are a growing and pernicious problem for the U.S. electric system. The Financial Times reported recently, “US data centre developers are flooding utilities with inflated growth plans, muddying efforts to plan for future power needs with projects that may never materialise.” Data center developers are playing a dangerous game. They are offering the same AI project to multiple utilities, seeking the lowest power price. Says the FT, “The plans are being left in the queue of projects waiting to connect to the grid even after they are no longer viable, leading to bloated demand forecasts.”
The result is stranded costs for generation and transmission to serve the fake loads, driving up retail customer rates. Josh Price of the D.C. research firm Capstone said, “Even the folks who benefit the most from sky-high projections are realising, ‘what happens if the pendulum swings back the other way and rates get unaffordable?’”
Pacific Gas & Electric, according to the FT, has cut 400 MW from its data center estimates, about 25 data centers. Veteran utilities analyst Julien Dumoulin-Smith said, “There’s an ongoing trend of whittling down. How many projects are real at this point? That’s what I want to know.”
The fates of many of the major players in the AI market are also in doubt, according to in-depth analyses by engineer and veteran technology journalist Will Lockett in his online publication Planet Earth & Beyond and the Medium platform. He recently published a lengthy and devastating analysis of the financial cesspool facing the major AI companies.
The major AI players began life with up-front venture capital financing from massive tech moguls such as Microsoft, Google, Amazon, and Meta. As AI development costs soared “exponentially,” Lockett notes, the companies turned to the stock market for operating cash.
Costs were still outpacing revenues. The equity play, he writes, “quickly devolved into the infamous circular financing deals between OpenAI, Nvidia, Oracle, CoreWeave and Microsoft, which, in the end, wasn’t even enough cash to cover the spiralling costs of AI.”
Where to go next as outflow continues to exceed inflow? This year, “Big Tech has turned more and more towards debt financing, or taking out loans, to cover their mounting AI costs.” According to Lockett, the AI firms are predicted to spend $400 billion this year and have borrowed $200 billion so far, half of that in September.
Says Lockett, “AI costs are really spiralling out of control. AI expenditure is set to double next year.” Morgan Stanley is predicting a 2025 spend of “$2.9 trillion, the vast majority of which will require debt financing.” J.P. Morgan says AI will need $1.5 trillion in debt financing just to cover planned data centers.
Lockett focuses on CoreWeave, which specializes in providing cloud-based graphics processing unit (GPU) infrastructure to AI developers and enterprises. “When the shit inevitably hits the fan at CoreWeave, its collapse will serve as a trigger point,” he writes. “This is why I say the bubble has already burst. The mechanisms are in place, and the catastrophic ending is set in stone. It is just that the market hasn’t noticed yet.”
Somewhere further out, an existential threat may be developing for today’s AI firms: Qubits. Former, and ousted, computer chip maker Intel CEO Pat Gelsinger told the FT recently that quantum computing is going to quickly eclipse conventional AI technology, based on graphic processing unit chips. Quantum computing is based on using polarized photons, which can represent both O and 1 at the same time, dubbed “Qubits.” Computing will become much faster with smaller electricity demands, flushing the legacy AI players down the technological toilet.
The FT reported that “Nvidia’s Jensen Huang has said it will take two decades for quantum to go mainstream.” Gelsinger disagreed, suggesting it will only take two years, noting that in either case, “we’re heading into the most thrilling decade or two for technologists.”
The FT recounted that Gelsinger “doesn’t see the AI bubble popping for another couple of years, but thinks a quantum breakthrough could trip it.”