This is not a prediction. It is an argument for managing risk and preserving strategic options under uncertainty.
Most national AI strategies describe an almost linear path. Governments invest in compute, data and people. Capabilities improve, adoption expands and growth follows.
That is a reasonable way to plan, but it is not the only path the technology may take.
Technological revolutions do not always move in a straight line. At times, their promise attracts investment faster than the economy can turn the technology into revenue and productivity. Companies are formed, infrastructure is built and valuations rise.
Then comes the correction. Valuations fall, companies disappear, investors lose money and the industry consolidates. When the correction is sharp and follows a period of exceptional gains, we often call it the bursting of a bubble.
National AI strategy should account for this possibility. Not because we know that an AI bubble exists today, and certainly not because we can predict when it might burst. The reason is more modest. If this pattern has appeared in earlier technological revolutions, governments should ask what they can do to soften the effects of a possible correction and preserve the capabilities they will need to enter the next phase of growth from a stronger position.
How a valuable technology can still produce a bubble
At the beginning of a technological wave, no one knows the eventual size of the market, how quickly organisations will adopt the technology or which companies will turn technical capability into a profitable business. Investors therefore price more than current revenue. They also price the possibility of much larger revenue in the future.
As the first successes appear, expectations rise. More money enters the market, more companies are established and more infrastructure is built. Investment expands what the technology can do and reinforces the belief that growth will continue. At this stage, it is difficult to know how much of the increase reflects newly created value and how much reflects expectations running ahead of reality.
Over time, the gap begins to show. Adoption may move more slowly than expected, competition may erode margins, and only a small number of companies may succeed in converting technical capability into revenue. The result can be a sharp market decline, company failures and lower investment. Yet a correction does not necessarily erase everything built during the boom. Infrastructure, knowledge and experience may remain and support broader adoption later.
Economist Carlota Perez, whose work examines the relationship between technological revolutions and investment cycles, identified a pattern across several technological revolutions. Rapid investment and rising valuations were followed by a correction and, later, by broader deployment of the technology across the economy.
Her framework is a useful way to think about the sequence, but it is not a forecasting tool. It does not imply that every technology will follow the same path, and it cannot tell us when a correction will arrive.
Other economic research helps explain what may happen within such a cycle. In an article published in the American Economic Review, economists Lubos Pástor and Pietro Veronesi showed how uncertainty surrounding a new technology can produce a sharp rise followed by a correction without assuming that investors are irrational. As evidence about the technology accumulates, some expectations prove excessive and valuations adjust.
A correction does not prove that the technology had no value. In a study published in the Journal of Financial Economics, economists Haddad, Ho and Loualiche examined more than one million patents and found that economically valuable innovation continues to emerge during periods identified as bubbles. Their finding strengthens the distinction between the value of a technology and the way markets price the companies developing it. A correction can change valuations and redistribute value among companies without negating the importance of the technology itself.
When investor losses become a national crisis
A sharp decline in share prices can cause heavy losses without becoming an economic crisis. As long as losses remain with investors who chose to bear the risk, the damage need not spread across the economy. It becomes a broader problem when affected companies cannot service their debt and losses pass to banks, public budgets or organisations whose continued operation matters to the wider economy.
In an article published in the Journal of Monetary Economics, economists Òscar Jordà, Moritz Schularick and Alan Taylor examined equity and housing bubbles in 17 countries over roughly 140 years. They found that not every burst bubble becomes a crisis. The risk rises when debt grows faster than the income available to repay it, making repayment dependent on continued growth. In such cases, corrections tended to lead to deeper recessions and slower recoveries.
The OECD Global Debt Report 2026, which reviews developments in 2025, shows that bond issuance by the largest cloud companies, or hyperscalers, and AI-related private credit transactions rose sharply that year. Data centre construction is also financed through bank loans, special-purpose vehicles and securitisation, and sometimes rests on leases and other long-term commitments. The question is therefore not only how much money is being invested in AI. It is also how much of that investment is financed by debt, who remains obligated to pay if demand disappoints and where the losses will ultimately sit.
The event that starts a downturn is not necessarily its underlying cause. Economist Pascal Paul of the Federal Reserve Bank of San Francisco showed how a long period of growth, increasingly dependent on new credit and the assumption of continued income growth, can make a system fragile. Under those conditions, even a moderate shock can trigger asset sales and a contraction in credit. The shock exposes fragility built during the boom. It did not necessarily create it.
For AI, the implication is not that a simple policy tool can prevent a bubble. It is that governments should understand how the current wave is being financed. Who is paying for data centres, electricity, cloud contracts and chips? Who must meet those commitments if demand grows more slowly than expected? Which losses will remain with investors, and which could pass to banks, public budgets or the public?
This mapping does not require the state to decide whether AI companies are overvalued or to predict what will trigger a correction. Its purpose is to assess whether the system being built around the investment wave could absorb lower revenue, valuations and financing without transmitting the damage to the rest of the economy.
What has helped countries limit the damage
The global financial crisis of 2008 showed how falling asset prices can spread through banks to the broader economy. Yet the severity of the damage differed across countries, partly because some banking systems entered the crisis better able to absorb losses and continue lending.
In its Annual Survey of Israel's Banking System for 2008, the Bank of Israel examined how the domestic system weathered the crisis. Israeli banks suffered from declining asset values and exposure to financial institutions abroad, but remained relatively resilient compared with banks in many other countries. The Bank of Israel attributed this in part to conservative banking practices, close supervision, limited exposure to the assets at the centre of the crisis and relatively low dependence on capital markets for funding.
The International Monetary Fund reached a similar conclusion about Canada. In its 2009 report on Canada, it linked bank resilience to strong capital, stable sources of funding and conservative regulation and supervision. These cases are not controlled experiments, so they cannot tell us how much weight to assign to each factor. They do suggest that the structure of the system before the crisis affected its ability to absorb the shock.
Spain allows a closer look at one specific tool. Beginning in 2000, the Spanish central bank required banks to build provisions when credit losses were low so that they could draw on them when conditions deteriorated. In a study published in the Journal of Political Economy, economists Gabriel Jiménez, Steven Ongena, José-Luis Peydró and Jesús Saurina compared banks that were affected differently by changes in the provisioning rules. This allowed them to examine how provisions accumulated during the boom affected banks' ability to keep lending during the crisis.
The researchers found that the accumulated provisions helped banks continue supplying credit and supported company activity during the crisis, but were not sufficient to prevent Spain's banking crisis.
The lesson for AI is the value of preparing in advance. The tools, however, must reflect the way today's market is financed.
Preparation is not an attempt to prevent market fluctuations or corrections. It is an attempt to reduce the risk that they spread through the financial system and the wider economy. Building resilience, however, also carries a cost.
Resilience is not free
Any measure that reduces fragility imposes some cost during normal times.
An IMF study examined the effect of tools designed to reduce financial risk using data from roughly 900,000 companies in 48 countries. The researchers found that the use of these tools was associated with slower growth in credit, investment and sales, particularly among smaller and younger companies.
This finding matters for AI. Young companies can be a source of innovation and technological breakthroughs. Restricting credit may curb speculative investment, but it may also make it harder to finance companies developing valuable technology.
When governments consider measures intended to build resilience, they cannot ask only how much protection those measures provide. They must also account for what they will cost.
The following table gives examples from another IMF study, which compared the contribution of several tools to reducing crisis risk with their cost to economic activity.
| Policy measure | How it may help during a crisis | Cost during normal times |
|---|---|---|
| Require banks to keep a larger share of funds in reserve | Slows rapid credit expansion and reduces banks' exposure | Leaves less money available for loans and investment |
| Require banks to hold more capital against risky loans | Creates a layer of protection that can absorb losses | Credit may become more expensive and available to fewer borrowers |
| Limit loan size relative to income or asset value | Reduces the risk that borrowers accumulate debt they cannot repay | Less credit is available and economic activity may slow |
What will allow the economy to grow after a correction
Even if a country succeeds in softening the downturn, another question remains. What will allow its economy to renew itself and return to growth?
In a classic article published in the American Economic Review, economists Ricardo Caballero and Mohamad Hammour examined the claim that recessions cleanse the market of less efficient companies and activities. Their model describes an economy in which new production units embody more advanced technologies while obsolete units gradually leave the market. A recession may accelerate exit, but it can also slow the creation of the new activities meant to replace what was lost. Company closures alone do not guarantee renewal.
A later article, also published in the American Economic Review, connected reallocation directly to innovation. Daron Acemoglu and his co-authors used a model informed by US data on companies, research and development, and patents to show how the exit of less productive firms can release skilled workers for R&D in companies with greater innovative capacity. The study is not about crises or AI, but it illustrates how movement of workers between companies can support technological progress.
During a crisis, however, this process can also move in the opposite direction. In an article published in American Economic Journal: Macroeconomics, Diego Anzoategui and his co-authors showed how the fall in R&D and technology adoption during the Great Recession contributed to the persistence of the productivity slowdown. When financing and demand contract, companies and activities with real potential may also cut investment.
The opportunity created by a crisis therefore depends on whether the economy can move skilled workers into innovative activities and whether businesses can continue financing R&D and adopting technology. Without these conditions, a crisis may weaken innovation rather than accelerate it.
An OECD report published in 2009 examined how countries incorporated investment in innovation into their responses to economic crises. Finland maintained public investment in innovation during its early 1990s crisis, while Korea increased the government R&D budget after the Asian financial crisis of the late 1990s to offset the decline in business R&D. The report assigns these measures a role in recovery and in strengthening the two countries' innovation capacity, while noting that policy alone cannot explain the outcome.
What should national AI strategy include?
Governments do not need to determine whether AI is a bubble. The literature provides enough evidence of a possible sequence of boom, correction and renewed growth to justify including the scenario in national AI strategy.
The first question is which mechanisms could soften the downturn if the boom ends in a sharp correction. Where is fragility accumulating, and what could stop investor losses from spreading into credit, employment, public budgets and services that depend on the technology?
But resilience is not free. Governments should also ask how they can reduce the risk of a crisis without imposing excessive costs on financing, young companies and the momentum of technological development.
Finally, if a bubble does burst, governments should consider which capabilities the country will still have at the bottom. Will it retain infrastructure, human capital, R&D and a financial system able to support the next phase? Could a global decline in investment, for example, make it possible to bring experienced professionals back to Israel or redirect them into other innovative activities? These capabilities cannot guarantee that a crisis becomes an opportunity, but they may improve the relative position of a country that prepared in advance.
We do not need to predict the bursting of an AI bubble to prepare for the possibility of a sharp correction. Governments can already examine what might turn such a correction into a crisis, what could soften the damage and which capabilities they need to preserve in order to return to growth.
The research behind the argument
The following sources provide the main building blocks of the argument and clarify what each one contributes.
Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages, Carlota Perez, Edward Elgar, 2002
Perez presents a historical framework in which technological revolutions move from installation and financial boom through a turning point and into broader deployment. It is a comparative framework for interpreting technological revolutions, not a tool for predicting the timing of a correction.
Technological Revolutions and Stock Prices, Pástor and Veronesi, American Economic Review, 2009
The article develops a model in which uncertainty and learning about a new technology can produce a rise and subsequent decline in the value of innovative companies. It shows that a bubble-like pattern does not require irrational investors.
Bubbles and the Value of Innovation, Haddad, Ho and Loualiche, Journal of Financial Economics, 2022
Using more than one million patents, the study shows that during bubbles the market valuation of innovation can become disconnected from its economic effects. It supports the distinction between the value of a technology and the valuation of the companies developing it.
Leveraged Bubbles, Jordà, Schularick and Taylor, Journal of Monetary Economics, 2015
The study examines equity and housing bubbles in 17 countries over roughly 140 years. It shows that bubbles accompanied by credit expansion are more dangerous and tend to end in deeper recessions and slower recoveries. Its relevance is not limited to technology bubbles. It identifies the mechanism through which leverage can turn a correction into a crisis.
Global Debt Report 2026: Sustaining Debt Market Resilience Under Growing Pressure, OECD, 2026
The report describes the shift from financing AI investment mainly through internal cash flow toward greater use of bonds, bank loans, private credit and financing structures that sit outside the main corporate balance sheet.
A Macroeconomic Model with Occasional Financial Crises, Pascal Paul, Journal of Economic Dynamics and Control, 2020
The article presents a model in which credit and leverage accumulate during a boom and create financial fragility. Under those conditions, a moderate shock can trigger a deep crisis, which means the event that starts the downturn is not necessarily the main cause of its severity.
Israel's Banking System, Annual Survey 2008, Bank of Israel, 2008
The survey describes the impact of the global crisis on Israeli banks and their relative resilience. It attributes this in part to conservative banking practices, close supervision, limited exposure to the affected assets and relatively low dependence on capital markets for funding.
Canada: 2009 Article IV Consultation, IMF, 2009
The IMF report describes the relative resilience of Canadian banks during the crisis and links it to limits on leverage, strong capital ratios, stable funding and effective supervision. This is an institutional account of the case, not an experiment that isolates the contribution of each factor.
Macroprudential Policy, Countercyclical Bank Capital Buffers, and Credit Supply: Evidence from the Spanish Dynamic Provisioning Experiments, Jiménez and co-authors, Journal of Political Economy, 2017
The study uses changes in Spanish provisioning rules that affected banks differently. It finds that provisions accumulated during the boom helped banks continue lending and supported company activity during the crisis, but did not prevent the wider Spanish banking crisis.
The Micro Impact of Macroprudential Policies, IMF Working Paper, 2018
The study combines data on roughly 900,000 companies in 48 countries and finds an association between tools designed to reduce financial risk and slower growth in credit, investment and sales. The relationship was particularly pronounced among small and young companies.
Evaluating the Net Benefits of Macroprudential Policy: A Cookbook, Arregui and co-authors, IMF Working Paper, 2013
The paper provides an analytical framework for comparing the benefits of tools that reduce the probability and cost of a crisis with the price they impose on economic activity. The examples in the table illustrate the kinds of costs included in that framework rather than the results of a single experiment.
The Cleansing Effect of Recessions, Caballero and Hammour, American Economic Review, 1994
The article examines how recessions affect the exit of obsolete production units and the creation of new ones. A recession may accelerate exit while slowing the creation of new activity, so company closures do not guarantee renewal.
Innovation, Reallocation, and Growth, Acemoglu and co-authors, American Economic Review, 2018
The article develops and estimates a model of innovation and reallocation using US company, R&D and patent data. In the model, the exit of less productive companies can release skilled workers for research and development in companies with stronger innovative capacity.
Endogenous Technology Adoption and R&D as Sources of Business Cycle Persistence, Anzoategui and co-authors, American Economic Journal: Macroeconomics, 2019
The study shows how declining demand can reduce R&D and technology adoption, and how this contributes to a persistent productivity slowdown after a crisis. It explains why a downturn can also weaken the growth phase expected to follow it.
Policy Responses to the Economic Crisis: Investing in Innovation for Long-Term Growth, OECD, 2009
The report discusses Finland's and Korea's responses to earlier crises. Finland maintained public investment in innovation during the early 1990s crisis, while Korea increased its public R&D budget after the Asian financial crisis of the late 1990s. The report presents these cases as policies that supported recovery, but does not establish that policy alone produced the outcome.