TS Lombard speculates the AI bubble may burst in early 2028, with U.S. tech stocks potentially plunging 40%

Currently, global markets are concentrating their bets to an extremely high degree on an unprecedented infrastructure boom—the AI investment frenzy. However, the underlying logic of this boom is becoming extremely fragile: current profits depend mainly on "capital expenditure recycling (Capex recycling)" within the tech giants and on debt financing, rather than on real payments from external end users.

According to news from the Zhuifeng Trading Desk, Dario Perkins, chief macro strategist at TS Lombard, said in an October 22 research report that, according to an ultimate projection, this AI capital expenditure frenzy is highly likely to come to an end in early 2028 and trigger a severe market reset.

The report points out that the S&P 500 will fall 30%, tech stocks will plunge 40%, about $20 trillion in market value will be wiped out, and the tech weighting of the U.S. stock market will return overnight to the level before ChatGPT was released. Although the tech giants can survive thanks to strong balance sheets, weaker players upstream in the AI ecosystem (such as emerging cloud service providers and independent model developers) will face a wave of defaults, expected to generate $125 billion to $190 billion in credit losses.

TS Lombard believes the macroeconomy will fall into a mild recession: the sharp drop in AI capital expenditure will cause U.S. GDP to fall by about 2 percentage points and lead to the loss of about 800,000 data center construction-related jobs. But because private-sector balance sheets are healthy and the Federal Reserve has ample room to cut rates, this will not turn into a systemic financial crisis like that of 2008.

AI capital expenditure has become one of the largest infrastructure projects in human history. U.S. tech companies invest about $1 trillion a year in AI construction, and cumulative investment in super data centers has surpassed the Manhattan Project, the Apollo moon landing program, and even the Marshall Plan. The share of hyperscalers' capital expenditure in U.S. GDP has approached a historical peak.

However, the foundation of this boom is not solid. Revenue growth in the current AI ecosystem is still, to a large extent, a product of "capex recycling"—one company's capital expenditure is another company's revenue. As long as hyperscalers continue to invest on a large scale, revenue within the ecosystem will continue to grow, but this does not represent true sustainability.

The real sign of sustainability should be a significant increase in the share of revenue coming from outside the ecosystem—ordinary consumers and non-tech enterprise users. But as of the report's publication, this proportion had barely improved. Monetization of AI remains the biggest question mark: about 80% of OpenAI's and Anthropic's revenue comes from just 1% of customers, and most of these customers are loss-making AI startups that rely on continued venture capital funding.

The report says the current AI boom is essentially a "capex recycling"—as long as the tech giants keep injecting capital into the system, revenue will grow. But as debt becomes the main source of marginal financing, this model is extremely vulnerable to a fatal blow from a slowdown in capital expenditure.

Six crash triggers: any one of them could set off a chain reaction

The report sorts out six potential risk points and points out that an AI crash is already in an "overdetermined" state—any one of the following six triggers could be the blade that pierces the bubble.

Trigger one: revenue falls severely short of expectations

Multiple studies show that for current capital expenditure to earn a reasonable return, the AI industry's annual revenue would need to reach $2.5 trillion to $3 trillion by 2030, about 14 times the current level. Such growth is almost impossible to achieve in such a short time, especially against the backdrop of Chinese open-source (open-weight) models rapidly seizing market share. Ordinary consumers are generally resistant to paid AI services, and AI services face the risk of commoditization.

Trigger two: credit risk rises sharply

Hyperscalers previously relied mainly on their own cash flow for investment, but debt has become the marginal source of financing for AI construction. In 2027 alone, AI companies are expected to issue about $700 billion in new bonds, of which hyperscalers account for $400 billion and the rest comes from other participants in the AI ecosystem—a figure that exceeds total government bond issuance in the euro area.

Even more worrying is the rapid expansion of "off-balance-sheet liabilities." Take Meta's Hyperion data center in Louisiana as an example: it is a $30 billion investment project, and Meta sold 80% of its equity to Brown Owl, setting up a joint venture called Beignet, which raised $27 billion in new debt guaranteed by Meta (committing to lease the data center for 20 years). Because the lease was split into five four-year leases, only a very small portion of the leverage appears on Meta's balance sheet. Hidden liabilities across the entire AI ecosystem are estimated to have reached $1.5 trillion to $2 trillion. Financing rates for some emerging cloud service providers (such as CoreWeave) have reached as high as 15%.

Trigger three: the "teaser rate" trap and the depreciation cliff

The report draws an analogy between the current AI financial structure and the "teaser rate" of the 2000s subprime crisis. Data centers do not need to be depreciated during the construction period, but once they are put into use, depreciation pressure will suddenly intensify. At the same time, the multi-year data center leases signed by AI labs are not counted as costs before the data centers officially begin operations. Once more data centers come online one after another, the financial condition of AI companies may undergo a fundamental transformation within 12 months.

Trigger four: the IPO wave reveals the truth

Both Anthropic and OpenAI plan to go public (the latter has been postponed to 2027). Historical experience shows that IPO waves often mark cycle tops. More importantly, IPOs will for the first time provide the outside world with a window into the true economic condition of the AI industry—the metrics currently disclosed by AI companies (such as "annual recurring revenue" ARR) are clearly embellished, and there are extensive problems of information opacity.

Trigger five: semiconductor cycle and supply chain risk

NVIDIA is the "picks and shovels" of AI construction and the barometer of the entire AI boom. NVIDIA is carrying out "vendor financing" on a large scale, providing loans to customers to buy more chips, which the industry calls "balance sheet as a service." According to the latest research by the Bank for International Settlements (BIS), more than 55% of the investment flowing into AI companies comes from other AI companies, and the circular financing characteristic is extremely obvious. In addition, semiconductor demand is highly cyclical, and there is a "bullwhip effect"—once demand slows, the current hardware shortage may quickly turn into severe oversupply.

Trigger six: regulatory and security risks

The public image of AI technology continues to deteriorate. About 55% of Americans want to slow the pace of AI development, while only 13% oppose it, and this conclusion holds across all major political groups. According to the 2026 IBM/Ponemon Cost of a Data Breach Report, one quarter of malicious corporate data breaches worldwide involve some form of AI-assisted attack. Once a major security incident occurs, it could trigger a strong political backlash and tighter regulation.

The "second derivative problem": a slowdown in growth alone is enough to set off a crisis

The report particularly emphasizes the "second derivative problem" of the AI boom: AI capital expenditure does not need to decline in absolute terms; it only needs to slow in growth to be enough to trigger financial stress.

This is exactly the same logic as the 2000s subprime crisis—what triggered the crash at that time was not falling house prices, but house prices ceasing to rise rapidly. In AI construction, frontier labs sign contracts committing to pay data center operators billions of dollars, and the latter borrow against this to build data centers. But these labs themselves have no profits, and their ability to keep paying depends entirely on continuous financing. The capital expenditure of hyperscalers is like a loan ledger: once it slows, the entire boom will collapse.

Quantifying the market shock: $20 trillion in wealth evaporates, credit losses manageable

Tech stocks currently account for as much as 40% of U.S. stock market capitalization, with the semiconductor sector alone accounting for nearly 20%. The U.S. tech sector currently trades at a price-to-earnings ratio of about 30 times (40 times a year ago), far below the peak of 70-75 times during the internet bubble. In an AI crash scenario, the report predicts:

The S&P 500 falls 30% overall; tech stocks fall 40%; other sectors fall about 20%, for a combined market value loss of about $20 trillion.

The report says these estimates would pull the tech sector's share of the U.S. market back to the level before ChatGPT was launched.

By the end of 2027, the AI ecosystem outside the hyperscalers is expected to have about $545 billion in outstanding bonds, $920 billion in bank loans and about $650 billion in private credit. Total AI-related debt is about $2.9 trillion, of which about 25% is exposed to risk.

The report assumes a cumulative default rate of 35% between 2028 and 2030 and a recovery rate of 25%-50%, from which it estimates overall credit losses of $125 billion to $190 billion. Although this scale is considerable, it does not constitute a systemic threat to the overall U.S. financial system. Equity holders will bear the largest losses (about $20 trillion), but there will be no 2008-style cascade of credit losses and wave of full-scale bankruptcies.

The most vulnerable links include: neoclouds, Oracle, independent model developers (OpenAI, Anthropic) and construction/equipment suppliers.

A mild economic recession, similar to 1990 or 2001

By early 2028, U.S. tech-related capital expenditure will be about 3 percentage points higher than in 2023 (as a share of real GDP). The report assumes a complete reversal, but because data center investment is highly dependent on imports, the decline in imports will partly offset the impact of falling capital expenditure on GDP. Taking all effects together, the end of the AI data center boom will reduce the level of U.S. GDP by about 2 percentage points (the IMF's latest World Economic Outlook report reached almost the same estimate).

The report says that overall, the severity of this recession will be comparable to 1990 or 2001, a mild recession by historical standards, and gives the following three reasons:

First, private sector balance sheets are healthy. Private credit growth is less than 10 percentage points of GDP, more similar to the late 1980s and far from the level before the subprime crisis, so there is no risk of a systemic credit event.

Second, monetary policy ammunition is ample. The Fed has rebuilt ample monetary policy space for the first time since the global financial crisis, has ample room to cut rates, and does not need to rely on large-scale fiscal stimulus.

Third, rate cuts will activate the bottom of the "K-shaped economy." Low-income groups, demand for consumer durables, and the residential real estate market (where there is a large amount of pent-up demand), which have long been suppressed by high inflation and tight monetary policy, will benefit directly from rate cuts. Historically, after the dot-com bubble burst, the Fed's easing policy quickly spawned a new bubble in the housing market, and a similar scenario may repeat itself. Just as the Fed spawned a real estate boom after the dot-com bubble burst, rate cuts will quickly stimulate growth in non-AI sectors, which is also the core reason the recession will remain mild.

In addition, regarding the job market, the report points out that the AI construction boom has currently created about 500,000 jobs and is expected to increase to 800,000 by the end of 2027. If all these jobs disappear, the U.S. unemployment rate will rise by about 50 basis points, without considering the impact on broader business confidence.

The report points out that the biggest tail risk is that an AI crash may instead accelerate corporate adoption of AI automation to cut costs. History shows that the large-scale advance of internet-related automation occurred during the 2009 recession, not during the boom. If an AI crash triggers a large-scale automation wave, the shock to the labor market may exceed expectations, leading to a slow "jobless recovery."

How to preserve assets in a bubble

The report says an AI crash is a typical demand shock, not a 1970s supply shock (a stagflation-type recession). Long-term interest rates will fall significantly, the correlation between bonds and stocks will turn negative, which is favorable to the safe-haven properties of government bonds. But because the recession will be mild, there will be no return to zero interest rates or permanent quantitative easing.

The key difference between the AI boom and the dot-com bubble is that this tech boom is highly concentrated in a small number of U.S. and Asian companies, whereas the dot-com bubble involved broad participation from developed markets including Europe. South Korea, as the core of the AI supply chain, saw tech exports rise by 10 to 30 percentage points as a share of GDP over the past 12 months, and will be hit hard in a crash.

But markets such as Europe that did not deeply participate in the AI boom will significantly outperform the U.S. market after the initial panic selling. These economies have borne the negative effects of the AI boom (high inflation, high interest rates) but have not fully enjoyed its dividends, and the end of the AI boom may instead bring relief.

The report believes that the impact of an AI crash on the dollar will be far greater than that of the dot-com bubble and the subprime crisis. Because the root of the crisis lies in the U.S. financial system, funds will flow to other currencies such as the euro, yen, and even the pound. Global investors do not need to rush into U.S. Treasuries for safe haven; their own countries' bonds can play an equivalent role.

The optimal strategy refers to the "easy bubble" trading framework proposed by GMO's Ben Inker: hold a diversified global equity portfolio, benefit equally if the boom continues, and outperform concentrated bets on U.S. tech stocks in a crash. Only when clear stress signals appear in the AI ecosystem should government bonds be gradually increased.

This scenario report from TS Lombard conveys a sober rather than doomsday signal: the AI bubble bursting will not be 2008, but more like 2001. The U.S. economy will experience a mild recession, the financial system will not collapse, and the Fed has ample tools to respond.

But for investors, the evaporation of $20 trillion in wealth, a 40% drop in tech stocks, and the severe shock to countries and regions in the AI supply chain are enough to deal a devastating blow to any portfolio concentrated in AI.

~~~~~~~~~~~~~~~~~~~~~~~~

The above exciting content comes from Zhuifeng Trading Desk.

For more detailed interpretation, including real-time interpretation and frontline research, please join [Zhuifeng Trading Desk ▪ Annual Membership] Risk Disclosure and Disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the special investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investing based on this is at your own risk.