Regarding the AI bubble, these are the questions you should think about.
A new topic of discussion is trending among global corporate executives: Is there an AI capital expenditure bubble? If it exists, when will the bubble burst? Financial media have continuously raised their total investment estimates for data centers, with the scale climbing from hundreds of billions of dollars all the way to trillions of dollars recently. Market concerns about excess computing power, low or even negative return on investment, and economic recession continue to rise.
But asking the right question is paramount. "When will the AI capital expenditure bubble burst?" is not the right question. Bubbles cannot be accurately identified, and their bursting time is also unpredictable. Raising this question often implies an assumption that bubbles will bring macroeconomic challenges, and corporate managers, investors and policymakers can and should avoid the risks involved. Neither of these two points can be taken for granted.
A better direction for thinking is: What does the current AI capital expenditure boom mean? What are its connections with the macroeconomy and what hidden risks does it entail? How can prudent corporate managers move forward rationally in this strong trend?
The surge in economic activities such as the data center construction boom will indeed bring macro risks, but a destructive crash is not a destined outcome. The threshold for causing lasting systemic damage is actually very high. Although bubbles generally have a bad reputation, even if they burst, they may leave positive long-term legacies.
Estimating the Scale of the AI Capital Expenditure Boom
It is currently difficult to accurately grasp the exact current and future size of the AI capital expenditure boom. We can first count the capital expenditures of large cloud vendors such as Google, add the estimated investment of private AI enterprises such as AI labs, then exclude part of the capital expenditures that are implemented outside the United States or unrelated to AI, and estimate that the relevant investment in 2026 will be about 630 billion US dollars, slightly less than 2% of the US GDP.
But this figure overstates the actual impact of this round of AI capital expenditure boom on the macroeconomy. About half or more of AI capital expenditure is used for imports, especially semiconductor products, and this part will not drive domestic economic activities in the United States. After adjustment accordingly, the actual driving scale of this boom on US GDP in 2026 is close to 315 billion US dollars, equivalent to 1% of GDP. Bloomberg consensus forecasts show that this adjusted figure may rise to 1.5% of GDP by 2028.
Then how did the estimate of 315 billion US dollars turn into the claim of 3 to 4 trillion US dollars? The larger-caliber estimate is designed to measure the overall scale of this technological change, counting all expenditure plans over many years, covering global expenditure (not just the United States), and does not make the domestic macro-level adjustment deduction mentioned above. If you want to assess its actual impact on the current US economy, 315 billion US dollars has far more reference value than 4 trillion US dollars.
Potential Impact on the Economic Cycle
Huge numbers can easily create a sense of oppression, but it is difficult to judge how big the economic inflection point or potential loss is. In fact, this round of capital expenditure boom is still in its early stage. But it is certain that once the expansion slows down or stops, the tailwind that originally boosted the economy will turn into a headwind. It is easy to imagine that the economy will face catastrophic consequences, but a rational assessment needs to be carried out through analysis from three major transmission channels.
1. Shutdown of Economic Activities
Assuming that the problem of overcapacity is fully exposed, or policy pushback presses the pause button, the expected return of data centers collapses, and the AI capital expenditure boom suddenly ceases. In an extreme scenario, if all relevant projects stop immediately, economic activities will decrease by 315 billion US dollars, directly dragging down GDP growth by about 1%.
Looking at this magnitude of headwind alone, it is not enough to end the healthy economic expansion cycle of the United States, but if combined with other negative factors, it may trigger a recession.
2. Wealth Effect
The stock sell-off triggered by the AI sector may drive a deep correction in the broader market. Stocks held by US households directly or through funds account for nearly 30% of their total wealth, 10 percentage points higher than ten years ago. The decline in the stock market will directly impact household wealth, push up the savings rate, and in turn compress consumer spending.
However, the resilience of household balance sheets is strong, which will limit the spread of risks. In the post-pandemic era, two bear markets (with a drop of more than 20%) have already occurred in 2022 and 2025, but their impact on consumption and the real economy is relatively limited. People often link bear markets to economic recessions, but the correlation between the two is far less strong than the public impression. Only a deeper and longer-lasting stock market crash can cause substantial damage.
3. Credit Crunch
Once the debts issued to support the AI boom become bad debts, they will erode the balance sheets of investors. After creditors suffer losses, they may contract credit supply to other sectors, and credit crunch will in turn trigger an economic downturn.
But who bears the losses is crucial. Different entities have different capacities to absorb losses, and their amplification effects on the overall economy are also different. If capital losses occur in key links of the financial system, as will be explained in the next part, the damage caused will far exceed that of an ordinary recession.
How Bubbles Evolve into a Real Crisis
No two bubbles are exactly the same, and their macroeconomic consequences are also different. The Internet bubble in the late 1990s had a relatively mild impact on the economy; while the real estate bubble that burst in 2008 left long-term structural trauma. What conditions need to be met for a bubble to evolve into a severe crisis?
For a bubble burst to cause permanent structural damage, bad debts must hit the banking system hard. Banks are forced to shrink their balance sheets, market liquidity dries up, and the credit crunch effect is instantly amplified. The long process of balance sheet repair for banks, households and enterprises will miss a large number of investment opportunities. This is no longer an ordinary economic downturn, and the economic growth path will be permanently lowered.
At present, are banks that provide financing for the AI boom exposed to the risk of a crash? For now, the answer is basically no. Major cloud giants rely on their own cash flow, corporate bond issuance, and additional equity issuance directly to investors to support capital expenditures. Private credit institutions such as KKR and Apollo lend directly through their own funds, and are themselves outside the banking system. Some banks may have direct exposure, and there are inevitably some indirect risks, but losses from AI investment are unlikely to hit bank balance sheets as hard as the real estate crisis in the 2000s.
Even if AI investment cools down, cloud giants will be impacted, but it will not trigger a systemic existential crisis. These companies hold some of the strongest balance sheets in corporate history. Related investments can be written down, and the companies will most likely survive. Even if the AI boom does not meet expectations, their original main businesses still have strong value.
At the same time, private credit institutions face two types of related risks: the powerful code capabilities of AI will make the previously issued credit to software enterprises face uncertainty; while data center construction financing may hedge the headwind brought by the software business, and once the capital expenditure cycle reverses, it will further amplify losses. But in this adverse scenario, the core issue is still who bears the losses. Most of the investors in private credit are institutional investors and high-net-worth individuals, and generally do not include entities such as banks that play a systemic amplification role in credit intermediation.
There are other hidden concerns in the capital expenditure boom, such as the problem of circular financing between chip manufacturers, cloud giants and data centers, and new risks may emerge in the future. These hidden dangers cannot be ignored, but their potential impact can still be assessed using the three major transmission channels mentioned above.
Bubbles Can Also Bring Positive Value
Historical systemic trauma cases such as the 2008 crisis have given bubbles a completely negative reputation, and this view is too short-sighted. Risks certainly need to be taken seriously, but most bubbles do not trigger such systemic disasters.
Bubbles can even leave considerable positive macro legacies. Building new technology infrastructure requires a certain degree of optimistic fanaticism to mobilize massive resources and talents. Without the expectation of creating huge wealth, many emerging technologies cannot be implemented at all. Bubbles can be understood as a mechanism to solve collective action problems.
People tend to focus on the staggering losses of investors after the bubble bursts: when the Internet bubble burst, Amazon's stock price plummeted by nearly 95%, and Global Crossing went bankrupt directly. But this cannot measure the long-term macro and social value brought by the bubble. Although investors of Global Crossing lost all their money, the terrestrial and submarine fiber optic cables laid by this company became the key foundation for economic growth in the following two decades; Amazon completely reshaped the retail industry.
As long as systemic risks are controllable, instead of fearing bubbles, it is better to embrace them rationally.
Adapt to the Environment and Lead Development in the AI Capital Expenditure Bubble
For better or worse, bubbles are a normal state of capitalism. Learning to adapt to bubbles and manage businesses in them may be uncomfortable for corporate managers, but it is unavoidable. Here are five suggestions:
1. Participate actively to survive and grow
Many managers believe that their responsibility is to avoid bubbles. But paradoxically, this choice itself will bring survival risks. Imagine a major memory chip manufacturer that determines this is a bubble and chooses to wait and see while its peers are investing heavily. Even if the bubble bursts later and proves its judgment is correct, it will face a market pattern of overcapacity, with competitors holding higher production capacity, and the enterprise will fall into a passive position.
2. Accurately identify the type of risk
Grand figures can easily bring shock and panic, but you need to calibrate the risks according to the assessment objectives and interpret them in the overall context of economic impact. You can first clarify which channels the risks are transmitted through, and then assess the strength of the headwind they bring. For example, focus on the investment flows that affect real GDP and think: How large is its impact on growth?
3. Distinguish between investor losses and macro risks
The only certain thing in a bubble is that someone will lose money, sometimes a huge amount. But individual investment losses do not equal a macroeconomic crisis, and such losses are often limited to a local scope.
4. Understand the sources of systemic risks
When a bubble bursts, losses are inevitable, but the impact of different losses varies greatly. Only when losses fall within the banking system, eroding the capital foundation and triggering a credit crisis, will they cause long-term macroeconomic trauma. This situation is difficult to predict, but bank credit spreads are signals worthy of continuous observation.
5. Treat policymakers as firefighters
Monetary policymakers have the responsibility to protect the economy, but it is difficult to identify bubbles in advance. Actively pricking bubbles may interrupt the economic cycle that brings employment and wage growth. Unless systemic risks in the banking sector are found, policymakers had better wait and see and prepare for post-incident disposal.
So, do we need to fear the AI capital expenditure bubble? The view of this article is: At present, it is a controllable macro risk, and it should be weighed comprehensively against its positive driving effect on economic growth. To change this judgment, credible evidence is needed to prove that it will substantially impact one or more of the transmission channels of economic activities, wealth, and credit. This possibility exists, but it is not a foregone conclusion at present.
Philipp Carlsson-Szlezak, Paul Swartz | Article
Philipp Carlsson-Szlezak is a Managing Director and Partner of the Boston Consulting Group's New York office, and also serves as the firm's Global Chief Economist. Paul Swartz is Executive Director and Senior Economist of the BCG Institute.
Zhou Qiang | Editor
This article is from the WeChat official account "Harvard Business Review" (ID: hbrchinese), written by HBR-China, and republished with authorization from 36Kr.