It feels like every time I check the news or open Instagram, I am hit with a new AI-related headline, each more terrifying than the last.
It’s everything from Grok’s creation and distribution of child pornography, data centers’ water use exacerbating droughts across the Southwest or a potential literacy crisis. Then, of course, there are the nefarious public-private partnerships that raise grave concerns about AI target generation in warfare and ubiquitous surveillance infrastructure. For those of us ideologically opposed to using generative AI in our daily lives, it can be easy to ignore the panic and say we are not causing the problem, so it will not affect us. This is where another prominent issue arises: the so-called AI bubble.
OpenAI plans to go public to stockholders by 2027 and Anthropic’s initial October Initial Public Offering (IPO) deadline was pushed to November. An IPO is when a private company sells its shares to the public on the stock market for the first time. IPOs can be highly beneficial for companies, as they provide new cash sources that can expand production or pay off debt and allow early investors to sell their private shares for a lot of money.
This public buy-in doesn’t only impact people who consciously choose to buy AI stock, though. Anyone with a retirement fund, for instance, will likely have their 401(k) tied up in these companies, whether they are willing and aware or not. If expert economists are correct, the public could end up paying the price for private speculation; for people who lived through the dot-com bubble and the 2008 financial crisis, massive overinvestment in AI products with much lower profit returns is ringing alarm bells.
Five of the biggest technology giants — Amazon, Meta, Alphabet, Microsoft and Oracle — had capital expenditures equaling almost 1.4% of American gross domestic product in the most recent quarter. According to The Guardian, 41 AI-related stocks now account for half the S&P 500’s market value, and, as one Harvard economist argued, “investment in information processing equipment & software” accounted for 92% of the total GDP growth in the first half last year. An expert from the Columbia Business Review argues that without the AI infrastructure buildout the U.S. would be in a recession.
There are also concerns that AI models cannot advance further in any meaningful way — that is, they have met their technological ceiling. The dead internet theory posits that most of the content we now see on the internet is AI-generated, or the product of bot activity. This means LLMs trained with the current internet are being fed data points from nonhuman sources, hence why we saw Anthropic buying and shredding books to access a training dataset without having to navigate copyright law.
By one estimate, more than $475 billion was spent on constructing data centers in 2025 alone. However, total AI revenue this year will be about $60 billion, while capital expenditures from the aforementioned five companies are on track for around $700 billion and $900 billion. In other words, estimated profit makes up less than seven percent of estimated spending. If you have seen the Bloomberg graphic representing what industry analyst Dylan Patel called an “infinite money glitch,” you understand how circular funding in tech circles is.
This circular financing was known as roundtripping during the dot-com boom. Essentially, through a series of financial gymnastics, companies can make an industry look more robust and solid than it actually is; Company A gives Company B money to buy products made by Company A, sometimes through a third party. In September, tech hyperscaler Nvidia announced an investment of up to $100 billion in OpenAI for a 10-gigawatt increase in data center capacity — OpenAI, in return, will use chips produced by Nvidia. By shifting money around, these companies create the appearance of new revenue that does not exist. This is what people mean when they talk about an “AI bubble.”
Like past market crashes, the bursting of this bubble will have wide-reaching consequences. Big tech companies are shifting AI-related spending off their balance sheets, instead electing to partner with private capital firms in joint ventures known as special purpose vehicles (SPVs) to finance data center construction. CoreWeave is one AI company that has relied on SPVs to fuel its expansion, transferring risk onto private-capital partners, like pension funds and insurance companies. If CoreWeave is unable to meet wildly speculative profit projections, the burden will not fall exclusively on tech shareholders, as retirees and insurance policyholders will be impacted too.
The Wall Street Journal reports that Oracle’s “adjusted debt, a measure that includes what it owes on leases in addition to what it owes creditors, is forecast to more than double to roughly $300 billion by 2028, according to credit analysts at Morgan Stanley.” The $18 billion the company borrowed to finance a new data center caused it to declare ‘force majeure’ last week, a measure protecting it from financial liability and delayed payments if they do not meet its 2028 online target.
Former Treasury Secretaries Robert Rubin and Hank Paulson wrote in a recent opinion piece that “The world has never seen so much capital chasing one opportunity.” Economic incentives for developers, they argue, are so alluring that the risks have been largely ignored. Rubin and Paulson advocate for meaningful government oversight, stating that self-regulation in competitive markets has never been rewarded and cannot be expected to solve potential crises. Without regulation, spending will continue to go unchecked in the absence of comparable profit margins and the possibility that private folly could become a public burden is all too real.
If the government’s approach to AI mirrors how it treated banks in the 2008 financial crisis — that is, as too big to fail — taxpayers will once again be funding private bailouts to avoid a complete market crash. The AI boom is much bigger than the subprime mortgages boom that set off the 2008 crisis, and companies like OpenAI and Anthropic would rather die trying than admit failure and lose the hundreds of billions invested in beating the competition.
In an interesting turn of events, former employees and even tech leaders themselves have recently called for regulation or a development slowdown, though they are still going forward with plans to go public. As the economist Owen Lamont joked, it’s “like Robert Oppenheimer doing an I.P.O. for the Manhattan Project in 1945.” He posits that, in making the implicit argument that AI is a potentially civilization-ending tool, CEOs encourage further investment and cash injections; after all, the more money the Amodeis have at their disposal, the better, safer and less existentially threatening their AI tools will be, right? Of course, while some developers may be calling for a slowdown, none will be the first to do it. Despite broader concerns about human society as we know it, none want to sacrifice their advantage or potential for profitable discoveries and developments. This is the danger of the so-called AI race: if no one can afford to slow down for fear of letting a competitor win, the question becomes what we will all lose along the way.