Last month, Nvidia—the maker of the go-to chips for artificial intelligence (AI) developers—crossed $5 trillion in market capitalization and followed up with a strong earnings report. Yet nagging fears of an AI bust have not gone away. The fundamentals of the AI industry—unprecedented levels of investment with little evidence of revenues in the foreseeable future—remain. By November’s end, even Nvidia’s stock fell despite strong earnings, reflecting that the market is still worried.
This begs the questions: Would an AI bust be an unmitigated disaster? With storm clouds continuing to gather, might there be a silver lining to the sky falling? And what if there were such a thing as a “good” AI bust?
Let’s get the bad news out of the way. As we ponder the possibilities of an AI bust, the inevitable parallel is to the dot-com bubble of the late 1990s. Between 1995 and 2000, the Nasdaq stock composite had climbed 600 percent, and the bubble burst in March of 2000. The Economist estimates that a dot-com-level pop of today’s AI bubble would cut U.S. consumption by roughly $500 billion (1.6 percent of GDP) and wipe out 8 percent of household wealth. With more than 60 percent of Americans invested in the stock market, the impact would be felt more widely than was the case during the dot-com collapse, when a smaller proportion of the public was invested in public equities.
The so-called “Magnificent Seven” tech companies that comprise 37 percent of the S&P 500 index would see their stock prices and market capitalizations shrink, and they would be forced to scale back their AI initiatives. Many AI start-ups would struggle to stay afloat; some of the more promising ones may get acquired. Operators of AI infrastructure would have stranded assets on their hands.
Aside from the industry’s dominance of the U.S. stock market, some idiosyncratic features of the way that the AI ecosystem has been funded could cause added pain in the event of a bust.
One such feature is that of “circular financing.” Many AI firms are funded by large tech companies, such as Nvidia or Microsoft, and those AI firms in turn provide revenue to Nvidia and Microsoft by using their chips and cloud services respectively. If demand drops, companies such as Nvidia are hit twice: once by a loss of value in their equity investments in the AI firms and secondly through a loss of revenue as chip demand falls. Additionally, debt-financed “special purpose vehicles” that fund several data centers where long-term leases from Big Tech companies are offered as collateral could default—leaving creditors exposed to huge losses.
Beyond tech, utilities and energy companies are exposed as well, since they have been increasing their generating capacity to meet the energy demand of the data centers. In an AI bust, this capacity could be underutilized, and the companies will have to take financial write-downs while energy customers possibly foot the bill.
For both individual and enterprise users of AI, shrinking research and development budgets would mean that they may have to work with AI models that are good but still imperfect. The promise of productivity gains from AI may be delayed. Just as concerning are the security and safety implications. As companies scale back, security updates, safety testing, and maintenance for deployed systems could lapse, creating vulnerabilities. We may witness a wave of layoffs in the tech industry, adding to the massive layoffs that have taken place already.
Meanwhile, unlike the 2008 financial crisis, when governments could deploy massive stimuli, today’s higher debt-to-GDP ratios would constrain policymakers’ ability to bail out failing companies or support affected communities. This could force more painful adjustments, including higher unemployment and a wider recessionary fallout. A bust might also lead to consolidation and greater concentration of the AI industry, giving antitrust authorities new headaches.
Much of the bubble is concentrated in the United States, and therefore, a bust would put the United States at a disadvantage relative to its primary competitor in the AI race: China. The European Union—while its investments in AI are more limited—has more regulatory checks and balances in place. In other words, an AI bust could shift the global power balance in profound ways, with the United States as the net loser.
Next, though, let’s consider five ways in which a bust may bring unexpected benefits.
First, even with the varied forms of financing of the AI boom, much of the build-out has been equity funded by tech companies with healthy balance sheets. This means that a bust is less likely to set off a widespread financial collapse.
That said, the AI economy is fragile. Harvard University economist Jason Furman has estimated that without AI investments, the U.S. economy would have been barely growing at 0.1 percent during the first half of the year, possibly indicating that the AI boom is masking an economy at a standstill. To compound the problems, the economy is also highly unequal because of AI’s predominance. The benefits of the boom have been disproportionately enjoyed by a few: high- and middle-income households heavily invested in the stock market, workers in the few economically vibrant tech hubs, a handful of Big Tech players along with a select group of AI start-ups.
Among potential AI adopter companies, a tiny minority have found tangible benefits from deploying the technology, while the rewards of AI-driven productivity have not been enjoyed uniformly. Moreover, unregulated and accelerated AI development is compounding many risks, from potential job losses to spiking energy prices that have given rise to multiple forms of societal inequality. A bust-induced slowdown may help with putting corrective measures in place as well as presenting an opportunity for new forms of innovation that help mitigate the negative impact of AI.
Second, the dot-com-era experience suggests that the overbuilding of telecommunications infrastructure proved to be a catalyst for growth after the dot-com bust. The underutilized facilities and fiber-optic networks ultimately made internet access more affordable and prepared the ground for the thriving digital ecosystem that emerged. In the current environment, infrastructure such as AI chips and data centers may have a shorter lifespan since they are likely to become obsolete more quickly than the dot-com-era infrastructure.
However, in the event of a bust, data center operators could get creative about redeploying excess capacity. Data centers can be reused for many purposes: facilitating traditional enterprise applications, such as file storage and sharing or processing transactions; high-performance computing for a variety of specialized uses, including genomic sequencing, drug discovery, and numerous other forms of scientific research as well as weather forecasting, climate modeling, and financial risk analysis; or even cryptocurrency mining.
Adversity can also stimulate creative repurposing, such as capturing waste heat from data centers and transforming them into energy sources. Beyond these uses, there is an established resale market for used data center equipment.
Ultimately, the uses of excess capacity will depend on the flexibility and currency of the technology of the infrastructure as well as its location. Overall, while a portion of the AI infrastructure will, no doubt, be truly stranded, there are many ways in which a significant portion can be reused.
Third, the energy infrastructure powering the AI industry can also be put to productive reuse—most immediately to meet existing needs and to ease the rising electricity prices that have been growing at twice the rate of inflation across the United States over the past year. Currently, the energy infrastructure buildup must contend with several inefficiencies.
For example, utilities are caught in a speculative exercise where data center developers submit power requests to multiple jurisdictions simultaneously, causing multiple utilities to build new capacity in parallel with consumers paying for additional power plants through rate increases. A bust will dampen this “fishing expedition” approach, allowing utilities to right-size their demand forecasts and invest in a more rational manner.
As for other utilities, as the data center build-outs slow down, there will be a deceleration in demand for water to cool them. This could lead to an easing of the water crisis in many communities.
Fourth, an AI bust will enforce more discipline on spending and eagerness to embrace cost-efficient alternatives. Chinese AI developers have demonstrated that this is possible with their release of DeepSeek earlier this year. DeepSeek showed that at a fraction of the cost and resources, you can produce AI models that perform almost as well as the U.S. models in every task. AI developers may look to these cheaper alternatives while also paying attention to innovations that prioritize efficient use of resources, such as chips that use light rather than electricity for power-intensive uses of AI, thereby introducing 10 times to 100 times the efficiency of the chips currently in use.
Plus, the United States lags behind China in terms of open-source downloads. In the event of a bust, model providers may offer open-source capabilities to maintain relevance and boost adoption, thereby democratizing access and catalyzing innovation and new application areas.
Finally, an AI bust would underscore the risks of unregulated AI growth and permit the creation of better governance systems and regulation as the hype recedes. It could make space for policymakers to debate and agree on implementable standards, ethical frameworks, risk mitigation protocols, and so on.
While a bubble makes it hard for competitors to step away from the treadmill of continuous acceleration, a bust creates conditions that favor resource efficiency, sustainability, and prioritization of AI that produces actual value and does so with more guardrails in place.
During earlier technological revolutions, busts have led to value destruction in the immediate term, but over the longer haul, they provided the time and space needed for institutions to catch up and for innovators to consider longer-term value.
The electricity revolution, for example, began with intense competition between a handful of private players—including Thomas Edison, George Westinghouse, and Nikola Tesla, among others—until a slowdown allowed regulatory bodies to form and act as stabilizers and guardians of the public interest.
Closer to home, AI’s own journey has been marked by at least two “winters,” in 1974-1980 and 1987-1993. As a result of earlier winters, researchers shifted their focus to developing more robust approaches, such as machine learning, which allowed computers to learn from data rather than relying on explicit rules. Ironically, this development gave rise to the present AI surge.
A bursting of the AI bubble will hurt, no doubt. But if history is a guide, then it is likely to accelerate progress toward a more beneficial and more societally sustainable AI. Even as we fear a bursting of the AI bubble, a bust may end up producing AI that is best for the long haul.
