With the flow of life insurance and savings premiums of millions of citizens into the dark market of private credit, the bursting of the AI bubble targets people's daily lives.
In a corner of Texas or Louisiana, bulldozers are tearing through wasteland to erect the steel skeletons of massive datacenters that will one day host thousands of AI chips. A few hundred kilometers away, a retired teacher who may never have chatted with a chatbot in their entire life checks their life insurance account, unaware that their retirement money is tied up in the power cables and cooling pipes of the datacenter.
In the summer of 2026, for the first time since the end of World War II, the stock market's share of American households' wealth surpassed the value of their real estate, reaching a record 34 percent. About one-third of all these stocks are held in retirement accounts and workers' pension funds.
At the same time, the stock market itself is in a more concentrated state; so much so that by October 2026, the ten largest tech companies accounted for more than 46 percent of the total value of the S&P 500 index, and if we add smaller AI-related companies to them, the future of AI will determine the growth or collapse of more than half of the US stock market's value.
Of course, workers and retirees do not voluntarily put their money into the high-risk game of AI. You may remember that in June 2026, SpaceX, which had merged the AI startup xAI a few months earlier, made the largest IPO in history by offering $70 billion in stock.
The entry of emerging giants into the stock market brings families' money into the riskiest game of the century
Prior to their initial public offering (IPO), SpaceX executives pressured managers of major stock indices such as Nasdaq and Russell to change their rules and fast-track SpaceX shares into major indices outside the normal procedure.
Passive investment funds, which manage people's retirement accounts based on these indices, automatically became buyers of shares in a company that considers part of its commercial future to be the costly idea of building AI data centers in space. With Anthropic preparing to go public in the fall of 2026 and OpenAI having a similar plan for 2027, ordinary people's savings are increasingly being drawn into this arena.
In October 2026, Matthew Scherer and Maya Jenkins from the Open Markets Institute and Americans for Financial Reform Education Fund examined the potential costs of an AI bubble burst in an analytical report titled "Brace for Impact," warning that tech companies have turned to massive borrowing to bridge the enormous gap between their costs and revenues; consequently, their potential collapse will directly involve families and the real economy.
Billions of dollars circulating among a few companies
Perhaps when chatting with chatbots on your phone, you don't consider the fact that even the free plans of AI models are costly for the companies themselves. According to Goldman Sachs estimates, global capital expenditures for building AI infrastructure will surpass the $1 trillion mark in 2026.
Just four tech giants—Microsoft, Alphabet, Amazon, and Meta, known as the "hyperscalers"—planned to spend about $745 billion on developing their AI infrastructure in 2026, which is five times their capital expenditures in 2023, and their total spending is projected to reach $5.3 trillion by 2030.
The authors of the report note that the scale of these tech giants' investments relative to US GDP even surpasses the Manhattan Project, the Apollo space program, and the construction of the Interstate Highway System. Such expenditures are only economically justified if they yield a proportional return on investment; however, the money that actual customers pay for AI products falls short of covering the costs.
Capital expenditures for AI infrastructure even exceed the scale of the Manhattan Project and the Apollo space programs.
Gartner estimates that total worldwide end-user spending on AI models and platforms will be around $39 billion in 2025 and about $64 billion in 2026. Simply put, the actual revenue of this industry from external customers barely reaches 10 to 20 percent of the money spent on building its infrastructure.
Pioneer companies such as OpenAI and Anthropic, which according to Ramp data control over 98 percent of the commercial AI API market, recorded net losses of 38 and 42 billion dollars respectively in 2025, and xAI, prior to its merger with SpaceX, had a negative profit margin of 1100 percent. For comparison, Amazon, which was famous for operating at a loss in its first decade, accumulated a total of only 3 billion dollars in losses over 9 years.
Circular Economy; The Illusion of Exploding Demand
How do AI companies always present an image of market boom to their audiences in their financial reports?
Part of the boom seen in the financial reports of AI companies stems from "circular financing." Tech giants invest in model-building startups, and the startups return that same money to their investors to buy chips or rent servers.
For example, Microsoft invests in OpenAI, and OpenAI buys Azure cloud services from Microsoft; Amazon and Alphabet have a similar relationship with Anthropic, and Nvidia, by financing startups and cloud companies, provides them with the funds to buy its own chips. Studies show that about 70 percent of Microsoft's AI revenue and 60 to 70 percent of Amazon's equivalent revenue come from OpenAI and Anthropic.
This intra-family money circulation creates an illusory picture of exploding demand. According to Ramp data analysis in August 2026, about 80 percent of the total enterprise revenue of OpenAI and Anthropic comes from just one percent of their customers, who are themselves mostly technology companies.
By purchasing services from one another, tech giants have created an illusory cycle of multi-billion-dollar demand and paper revenues.
On the other hand, in the second quarter of 2026, out of the total $285 billion in profits reported by the five companies Alphabet, Amazon, Meta, Microsoft, and Nvidia, $160 billion did not come from product sales, but stemmed purely from the rising paper value of their shares in companies such as Anthropic and SpaceX.
Alongside this money rotation, private startups that are not required to provide audited reports sometimes announce "annualized run-rate revenue" instead of actual annual revenue—meaning they multiply one month's revenue by twelve. This is exactly what Anthropic did at the end of 2025 by announcing a $9 billion revenue, but it was later revealed in court documents that the company's total actual revenue for the entire three-year period from 2023 to 2025 amounted to about $5 billion.
Datacenters Built on Debt
Until mid-2025, tech giants funded the construction of datacenters from the cash revenues of their other profitable divisions, such as internet advertising and online retail. However, project costs escalated so much that in 2026, the free cash flow of companies like Alphabet and Amazon turned negative for the first time in decades, and Microsoft and Meta also found themselves on the verge of a similar situation. It was at this point that the AI industry rushed to the debt market to keep going.
To make their official balance sheets look less risky to shareholders, technology companies keep a large portion of their debts off the books. Today, nine major AI companies—Alphabet, Amazon, Meta, Microsoft, Nvidia, Oracle, SpaceX, Broadcom, and AMD—have about $595 billion in long-term debt on their official balance sheets; however, Wall Street Journal calculations in August 2026 revealed that they have accumulated another $3.1 trillion in commitments and debt off their balance sheets.
Speculative debts of AI data centers are being sold in the form of new financial packages to pension funds and university endowment funds.
The Abilene Data Center Layer Cake
To see how hidden debts are engineered, we look at one of the most thoroughly documented cases in the report: the massive data center project in the city of Abilene, Texas, which the authors call a "corporate layer cake." At the top layer of the cake is OpenAI, which uses Nvidia chips at the Abilene data center to train its models, but does not own the building or the chips; rather, it leases them from Oracle using money obtained from investors like SoftBank and Nvidia itself.
In the next layer, it turns out that Oracle—having committed to buying billions of dollars worth of chips from Nvidia so as not to fall behind—does not own the Abilene data center either, leasing it instead from an emerging cloud company, or neocloud, called Crusoe Energy. The story gets even stranger when we realize that Crusoe itself does not fully own the project!
Crusoe, alongside private credit giant Blue Owl Capital and a digital infrastructure fund, has established a "paper company" or Special Purpose Vehicle (SPV). The three partners have put up $5 billion in cash, and the entity will also borrow $10 billion from JPMorgan Chase to build the data center on land owned by a sixth company named Lancium—the exact same tactic previously seen in the 2001 Enron bankruptcy scandal.
AI companies are keeping a large portion of their debt off-balance-sheet.
Similarly, Meta registered its $26 billion data center debt in Louisiana—in partnership with Blue Owl and PIMCO—under an SPV so it wouldn't appear on the company's official balance sheet.
A huge portion of off-balance-sheet liabilities are hidden in the form of long-term leases and purchase commitments in future years. The data center lease commitments of the five tech giants multiplied elevenfold from the release of ChatGPT in late 2022 to mid-2026, surpassing $1.1 trillion.
Alphabet alone has over $800 billion in future purchase commitments, and Nvidia, in addition to over $422 billion in supply chain commitments and off-balance-sheet contracts, has guaranteed $109 billion of its customers' debts and leases—a role that prompted The Economist to dub Nvidia the "central bank of AI."
Smaller cloud companies known as neoclouds, such as CoreWeave, which lack the creditworthiness of the tech giants, borrow at double-digit interest rates and pledge the very AI chips inside the data centers as collateral for their loans.
The Heavy Shadow of 2008 and Loans Sold Before Invoicing
When experts look at the $3.7 trillion volume of official and hidden debts held by nine tech companies, they are reminded of the bitter memory of the 2008 financial crisis; because the entire subprime mortgage market at the peak of the 2007 bubble, adjusted for inflation and in 2026 dollar values, was about $2.3 trillion. The similarity between the two eras is not limited to the magnitude of the numbers or the use of complex financial instruments and special-purpose entities, but also manifests tangibly in the transfer of risk from corporations and lenders to ordinary people.
The use of shell companies to build data centers bears a striking resemblance to the financial tricks of the Enron scandal.
Many of today's data center construction loans are disbursed based on "booked, but unbilled contracts"—meaning the bank or financial institution releases a massive loan solely based on an AI company's written promise to lease the building in the future.
These loans are then sliced up and securitized into bonds, exactly like the mortgages of 2008, and sold to institutional investors such as pension funds, university endowments, and insurance companies, shifting the risk of loan default away from the original issuer and onto retirees.
When Retirees' Savings Back High-Risk Loans
Part of the debt securities of the loans ends up in the $3 trillion market of private credit funds and life insurance companies. Operating outside of strict banking regulations, private credit funds have become the main lenders for high-risk projects over the past decade, and are projected to hold about $800 billion of datacenter debt by 2028.
The risk of datacenter loans, much like the 2008 mortgages, is being transferred to ordinary people.
Private equity giants like Apollo and KKR have acquired certain life insurance companies to secure the necessary liquidity—institutions that 90 million American families rely on for retirement pensions and financial security after the death of a household head.
Private funds take the savings and insurance premiums of ordinary people and invest them in high-risk, high-interest loans for building datacenters or purchasing consumer debt from "Buy Now, Pay Later" companies. More than half of the fixed-income asset portfolio in many of these insurance companies is now allocated to private credit and opaque securities.
To conceal the high risk, companies also assign asset reinsurance to their own subsidiary companies in offshore tax havens; meaning that if a crisis occurs, the insurance company and its guarantor company will collapse simultaneously.
Just like in 2008 when commercial banks claimed they were insulated from risky mortgages, yet in practice were connected to the shadow banking system through credit lines, today traditional commercial banks have also provided about $2.7 trillion in credit lines and loans to non-bank financial institutions through the first quarter of 2026, and have themselves allocated $450 billion in direct credit to artificial intelligence companies.
The collapse in datacenter valuations through private funds and insurance companies will spill over into the banking system
When you see all the steps of the AI financial system in the big picture, you encounter a chain in which a collapse in datacenter valuations could spread from private equity funds and life insurance companies to the country's banking system.
Since late 2025, amid growing doubts, many investors attempted to withdraw their money from private credit funds; however, fund managers blocked withdrawals to prevent a meltdown, creating a quiet and silent crisis in this market.
Cracking the Illusion of Productivity; Why Revenues Fall Short of Costs?
The industry's borrowing system operates on the assumption that companies worldwide will soon realize the productivity leap of AI and pay trillions of dollars to purchase subscriptions for it; but field evidence demonstrates the fragile foundation of such an assumption.
Numerous studies in 2025 and 2026 have shown that between 80 and 95 percent of corporate AI pilot projects have failed to yield a positive return on investment (ROI), and the rate of AI usage in the workplace among American adults dropped from 46 percent in mid-2025 to 33 percent in mid-2026.
The majority of enterprise AI projects have not generated a positive return on investment.
Klarna, the financial services company that had grandly replaced its support staff with AI chatbots, quietly brought hundreds of the laid-off employees back to work after realizing the algorithms could not handle empathy and manage angry customers.
In the spring of 2026, when Anthropic decided to shift its pricing model from a flat subscription to usage-based billing, Microsoft and Coinbase turned to in-house or open-source models to cut costs, and companies like Walmart and Uber set strict caps on their employees' AI usage.
Fear of customer churn drove Silicon Valley startups into a price war in the summer of 2026, offering 75% discounts and millions of free tokens to businesses—a move that further erodes their profit margins, especially as new reasoning models and agents consume significantly more tokens and electricity, driving up inference costs.
Simultaneously, with the Federal Reserve raising interest rates in September 2026, if investors realize that the multi-trillion-dollar revenues are never going to materialize, the chain of collapse will begin from the weakest links.
Where does the chain of collapse begin?
As demand slows, the pressure will first hit startups like OpenAI and Anthropic, and then Oracle—a company that has taken on $120 billion in long-term debt and $260 billion in off-balance-sheet lease commitments to supply the infrastructure required by OpenAI.
The cancellation of contracts and the failure of neo-clouds to repay their loans will also bankrupt paper companies, and their multi-hundred-billion-dollar commitments—much like Citigroup’s hidden debts in 2007—will return to the official balance sheets of tech giants.
With the halt in demand, the collapse begins from the weakest links of the chain and spreads to the balance sheets of the tech giants.
The major difference between the AI bubble scenario and the dot-com bubble of 2000 or the nineteenth-century railway bubble lies in the nature of the remaining assets. After the dot-com bubble burst, a network of fiber-optic cables remained with a useful life of several decades, which built the infrastructure of the modern internet. However, data centers—due to their massive power consumption and highly specialized cooling systems—are not economically viable for ordinary cloud computing, and their chips under heavy workloads last only one to three years.
Tech giants have recorded the lifespan of their chips as five to six years on their balance sheets to downplay depreciation costs; a gimmick that, according to calculations by famous investor Michael Burry, will leave Amazon, Google, Meta, and Microsoft with over $207 billion in hidden losses by 2028.
In the stock market as well—where the volume of borrowed money used to buy stocks (Margin Debt) reached a historical record of $1.5 trillion in the summer of 2026, roughly four times the peak of the housing bubble and five times the peak of the dot-com bubble—the decline in the value of tech stocks triggers margin calls and forced liquidations.
The report calculates that if the price-to-earnings ratio of the U.S. stock market merely returns to its historical average (based on the CAPE index, which has now reached 41, nearing the record of the dot-com bubble), $45 trillion of paper wealth will be wiped out; a shock that, through the "negative wealth effect," sharp declines in consumption, drops in investment, and the bankruptcy of pension funds and life insurers, will drag the real economy into a deep recession.
The Cost Ordinary Citizens Pay on Their Bills
Ordinary citizens are already paying the price for the AI boom. The thirst of data centers for energy and resources is increasing electricity bills for customers of the largest U.S. power grid by up to 76 percent, consuming water reserves in the arid regions of the Southwest, and laying the groundwork for a new wave of inflation by exacerbating shortages of semiconductors and construction materials.
In September 2026, the U.S. economy crossed a strange threshold: the amount of money spent on building data centers and processing hardware officially surpassed total investment in residential housing construction. Furthermore, in 2025 alone, the federal government granted nearly $27 billion in tax breaks for data center equipment depreciation to five companies: Alphabet, Amazon, Meta, Microsoft, and Oracle.
Corporate Bankruptcies, Not Bailouts with Public Funds
To prevent the repetition of crises like those of 1929 and 2008, the report's authors propose immediate actions to policymakers, such as reviving regulatory frameworks, requiring large private companies to publish audited financial statements, banning circular transactions and the use of shell companies to hide debt, setting realistic standards for calculating the useful life of chips, and cutting subsidies and tax breaks for data centers.
The authors also warn that the government and Congress must announce that if the bubble bursts, no taxpayer-funded bailout package will await Silicon Valley billionaires and Wall Street bankers.
Unlike banks, none of the tech giants hold everyday people's deposits, and in the event of default on their debts, they must go through the legal bankruptcy process, their executives must be dismissed, and pre-bankruptcy bonuses must be clawed back.
If the continued operation of part of the cloud infrastructure is deemed essential to the economy, the government can bring them under public management at market value during a crash. Public resources should also be directed straight toward supporting workers, policyholders, and retirees who have been involuntarily exposed to the risk of these debts, rather than bailing out speculators.
Comments