The Generative AI Bubble Has Begun to Pop

Generative AI has dominated the tech news over the last two years. More than 50% of venture capital funding goes to generative AI, which I distinguish from predictive AI, the older and, at one time, the more researched form of AI. Announcements and news from the big consulting firms and top business schools are also dominated by generative AI. The profit figures suggest that they are the most profitable organizations within AI outside of Nvidia.

Nirvana, doomsday or — bubble

Much of the AI discourse in the media was previously split between nirvana, doomsday and bubble. Recently, talk of a bubble has increased as it has become clear that OpenAI, and likely other AI software companies, may never become profitable. The startup losses are bigger than we have ever seen and the fact that the startups are not required to release audited earnings like public companies do, means that the losses are likely much bigger than the announcements.  Analyses of Microsoft’s third quarter earnings in 2025 confirm this.

Nevertheless, some companies are profitable, particularly chip companies such as Nvidia. Thus they would like to squeeze the system for every dime they can get, even though they know that the situation is not sustainable.

Some cloud companies may also be profitable, but we don’t know because the big tech companies don’t separate out the numbers for their AI cloud businesses from their other businesses. Core Weave only does AI cloud services, so we know that it is highly unprofitable.

The big reason why the semiconductor companies are profitable is because OpenAI and other software companies pay the bills for cloud companies even though they must borrow or receive new investments to do so, often from Nvidia, AMD and other tech companies. The cloud companies then pay for chips from Nvidia and others.

Debt financing is an old and reasonable way to survive, but the problem is that the cost of debt financing has gone up because the risks have increased. Furthermore, the cost of debt may go much higher if the overall interest rates were to rise, which is possible because of the appetite for borrowing by the federal government.

Circular financing

OpenAI and other AI software companies are also receiving new investments, mostly through so-called circular financing. Once used by Enron, it usually means that one company invests in another in return for purchases of their product. Think of it as a low-cost loan, but in some cases no money may change hands. Some of this financing is for show.

The details of debt financing and circular financing are interesting and can be found in many places including X (formerly Twitter) through posts by investment bankers. I quote one of the investment bankers from X (@HedgieMarkets) here: For instance, “CoreWeave uses Nvidia’s money to buy Nvidia’s chips and rents them back to Nvidia while spending $20 billion against $5 billion in revenue. The company has $14 billion in debt due within a year and $34 billion in lease payments starting through 2028.” “Nvidia invested $100 billion in OpenAI. OpenAI has $300 billion in agreements with Oracle, $38 billion with Amazon, $22 billion with CoreWeave.”

Similarly, “Oracle has just over $100 billion in debt and has become a bellwether for sentiment toward AI as bubble concerns grow,” which is why its shares have been falling for weeks. Oracle’s credit default swap spreads — which are a measure of risk — have shot up to around 1.26%, the most expensive on record, after its latest earnings reignited worries about AI spending.

Similarly, “Meta built a $27 billion data center using special-purpose vehicles to keep debt off its balance sheet,” which is borderline legal. Enron and several banks used this technique to hide their losses before they were caught. 

Excessive risk

Some of these loans have exposed many investors to excessive risk that they didn’t realize because “many of these loans also have zero disclosure requirements. Private-equity firms have $450 billion in tech loans, and life insurers have $1 trillion tied up. If AI loans fail, private credit fails, which could bring down banks and insurers because they’re all connected.”

Investors have become more concerned about these risks in the last week.  AI infrastructure stocks sold off sharply after Oracle delayed OpenAI data center completion to 2028 from 2027 and Broadcom’s AI revenue forecast missed expectations.

This greatly impacted semiconductor chips, as news items showed: “The Philadelphia Semiconductor Index fell as much as 5%, on track for its worst decline in two months. Broadcom dropped as much as 12%, while Astera Labs and Coherent Corp. fell more than 10%. Nvidia dropped as much as 3.2%.”  “The selloff extended to power-related stocks including Constellation Energy, Vistra, GE Vernova, and Cummins.”

Furthermore, “recent weeks have brought home the reality that there are numerous bottlenecks on spinning up data centers, from power supply to local politics, and any delays in building them mean pushing back spending worth hundreds of billions of dollars.” “Investors are increasingly on alert for any hint of AI infrastructure projects taking longer than expected.”

Shares in CoreWeave have been the most impacted. They have fallen about 45% since October while bonds issued by so-called AI hyperscalers like Oracle, Microsoft, and Meta have also been severely impacted.

The stock market is finally noticing the problems with AI, but they are focusing on the companies on the periphery. They can’t influence OpenAI, Anthropic, or other AI software whose unprofitability constitutes the bubble because those companies are privately held. But they can go after the public companies on the periphery whose costs and revenues are unsustainable.

AI stocks remain crucial

Some analysts said the rapid reversal demonstrates how crucial to markets the AI trade remains, able to drag indexes lower even while investors try to widen their bets beyond the small group of tech giants that powered stock records earlier this year. “I think it’s going to be a big question for the market: How patient are we going to be for these companies to get past the enthusiasm of the build-out of AI to the timeframe when we start expecting a return?” said Steve Wyett, chief investment strategist at BOK Financial.

But the market’s response soured as investors zoomed in on questions about the margins of its custom AI chips, the timing of a massive commitment from OpenAI and the company’s ability to see over the horizon into 2027. But recent weeks have brought home the reality that there are numerous bottlenecks on spinning up data centers, from power supply to local politics. Any delays in building them mean pushing back spending worth hundreds of billions of dollars. Investors are increasingly on alert for any hint of AI infrastructure projects taking longer than expected. Shares in CoreWeave, which builds data centers with Nvidia chips, are down 26% after it revealed a delay last month. Bonds issued by so-called AI hyperscalers like Oracle, Microsoft and Meta traded at unusually high volumes, as traders rushed to reduce their exposure to the sector.

Oracle shares tumbled on Thursday after signing up OpenAI and others to big contracts for its advanced cloud-computing services. But that business model required Oracle to spend tens of billions of dollars on Nvidia chips and networking equipment for data centers it leases. Those outlays come well before it books the revenue from those long-term contracts, and before it could demonstrate that renting access to AI computing was a durable or profitable business.

How Oracle plans to finance those commitments is an open question. Analysts expect the company will need to issue tens of billions of dollars of additional debt, plus tens of billions more it will owe on leases. As it was put in a headline at Futurism, “The Amount of Money OpenAI Lost Last Quarter Will Make You Choke on Your Slurpee.”

Why ads won’t save OpenAI

Listed, the problems are:

1. Financial problems;

2. Too expensive for processing;

3. Users aren’t buying anything;

4. Look at what people do on OpenAi. Jensen Huang: “In 2-3 years, 90% of the world’s knowledge will be generated by AI.” whether we learn from books written by unknown people or AI that is combining assimilating knowledge, it makes little difference.

AI adoption at work has fallen and now sits at 11% of Americans, according to Census Bureau data. Adoption has fallen sharply at the largest businesses employing over 250 people. Other surveys show stagnation too: one found AI use at work dropped from 46% in June to 37% in September.

Even the Wall Street Journal now warns “It Really Is Possible to Spend Too Much on AI” (November 26, 2025) And Futurism reports, “Insurance Companies Are Terrified to Cover AI, Which Should Probably Tell You Something  Why AI Workers Won’t Let Bots Do the Most Basic Tasks”

America’s startup system has become a black hole for losses

America’s startup system was truly once great. My analysis of startups found 24 startups founded between 1975 and 2004 have made the world’s top 100 public companies in terms of market capitalization for at least two years. Twenty-two of those startups, one of which was Amazon, were profitable by their tenth year. Ten of them had made the top 100 by their tenth year.

These companies also mostly provided value to users.

How things have changed. Only one startup founded since 2005 has made this list (Uber) despite record funding of startups by venture capital firms over the last ten years. Furthermore, Uber became profitable in its 14^th year largely by raising prices, which were enabled by the decimation of the public transport system during the pandemic lockdown.

Other startups came close before faltering. Zoom.

Readers may be thinking, wait a minute, what about OpenAI and other AI startups? They are hugely valuable. Yes, they are. But they aren’t publicly traded and perhaps more importantly we went through the same story over the last ten years.

Countless startups have been awarded with huge valuations (e.g., WeWork) and IPOs only to never escape the losses. My analysis of Unicorn startups (valued at $1 billion or more before doing IPOs) over the last ten years found that 85% of them were still unprofitable and this doesn’t count the dozens that went bankrupt or were acquired for low prices.

The story is being played out again:

The tech bros are descending into la-la land with their illogic about startup losses. For instance, when I pointed out the losses, one bro, Robert Scoble, who has 548k X followers, said there are many ways out of the hole. Someone will soon figure it out.

Optimism is commendable but the logic in this case is terrible.