A growing debate has taken hold among global executives: is there an AI capital spending bubble, and if so, when does it burst? As financial media keep lifting estimates for data center investment, with headline figures moving from the hundreds of billions into the trillions of dollars, concern has risen around excess compute capacity, weak or even negative returns, and the possibility of a broader downturn.
The article argues that this is the wrong question. Bubbles are hard to identify in real time, and even harder to time. Framing the discussion around the moment of collapse also sneaks in two assumptions: that the bubble will necessarily create a macroeconomic problem, and that executives, investors, and policymakers can and should avoid the risk in advance. The piece says neither assumption should be taken for granted.
A better question is what the current AI capex boom actually means, how it connects to the macroeconomy, what risks sit beneath it, and how prudent managers should operate while the trend is still strong.
How large is the AI capex boom?
The article says the exact size of current and future AI capital spending is difficult to pin down. One way to estimate it is to add together capital expenditures by large cloud providers such as Google and projected spending by private AI companies including AI labs, then subtract the share deployed outside the United States or unrelated to AI. On that basis, AI-related spending in 2026 comes to about $630 billion, slightly below 2% of U.S. GDP.
That number, however, overstates the direct macro impact on the U.S. economy. Roughly half or more of AI capex goes to imports, especially semiconductors, which do not directly boost domestic economic activity in the United States. After adjusting for that, the effective contribution in 2026 falls to roughly $315 billion, or around 1% of GDP. Bloomberg consensus expectations cited in the article indicate that this adjusted figure could rise to 1.5% of GDP by 2028.
The article also explains how an estimate around $315 billion can coexist with claims that the boom is worth $3 trillion to $4 trillion. Larger headline figures are designed to capture the total scale of the technological shift. They count multi-year spending plans, include global expenditures rather than only U.S.-based activity, and do not make the domestic macro adjustments described above. If the goal is to gauge the impact on the current U.S. economy, the article says $315 billion is more useful than $4 trillion.
Where the macro hit could show up
Big numbers create drama, but they do not by themselves tell us how severe a turning point would be. The article describes the current AI spending wave as still being in an early stage. One point is clear, though: once expansion slows or stops, what had been a tailwind for growth becomes a headwind. It breaks the risk into three transmission channels.
1. A direct drop in economic activity
Suppose excess capacity becomes impossible to ignore, or a policy backlash forces projects to pause, causing expected returns on data centers to collapse and the AI capex boom to stop abruptly. In the most extreme scenario, if all related projects were halted immediately, economic activity would fall by $315 billion, directly reducing GDP growth by about 1%.
On its own, the article says, that scale of drag would not be enough to end a healthy U.S. expansion. Combined with other shocks, it could contribute to a recession.
2. Wealth effects
A sell-off led by AI equities could pull the broader market into a deep correction. U.S. households, either directly or through funds, hold stocks worth nearly 30% of total household wealth, up 10 percentage points from a decade ago, according to the article. Falling stock prices reduce household wealth, raise the savings rate, and put pressure on consumer spending.
Still, the piece argues that household balance sheets have shown resilience, limiting the spread of stress. In the post-pandemic period, there were already two bear markets in 2022 and 2025, each involving declines of more than 20%, yet the effect on consumption and the real economy remained relatively limited. The article says the link people often assume between bear markets and recessions is weaker than common perception suggests. Only a deeper and more prolonged equity decline is likely to do material damage.
3. Credit tightening
If debt issued to support the AI boom starts to sour, investor balance sheets would take a hit. Creditors facing losses may reduce lending elsewhere, and that tightening in credit conditions could weigh on growth.
But the article stresses that what matters is not just how much money is lost. It also matters who absorbs the losses. Different holders have different capacities to withstand write-downs, and the broader macro amplification can vary sharply. If losses hit a critical part of the financial system, the damage can be much greater than in an ordinary downturn.
When a bubble becomes a real crisis
No two bubbles are the same, the article says, and neither are their macro consequences. It contrasts the relatively mild economic fallout from the late-1990s internet bubble with the long-lasting structural damage left by the housing bubble that burst in 2008. For a bubble to turn into a severe crisis, a tougher condition has to be met.
The key condition is damage to the banking system. For a burst bubble to leave permanent structural scars, bad debt has to hit banks hard enough that they shrink their balance sheets, market liquidity dries up, and credit tightening is rapidly magnified. Banks, households, and companies then spend years repairing their balance sheets, missing investment opportunities along the way. At that point, the problem is no longer a normal cyclical slowdown. The economy’s growth path shifts down on a lasting basis.
On the current AI investment cycle, the article says the answer is largely no when asking whether banks are heavily exposed to a potential collapse. Large cloud companies are funding capital expenditures mainly through internal cash flow, corporate bond issuance, and direct equity issuance to investors. Private credit firms such as KKR and Apollo are lending through funds that sit outside the banking system. Some banks may have direct exposure, and there are bound to be indirect risks, but losses on AI investments are not expected to hit bank balance sheets in the same way that housing-related losses did in the 2000s.
Even if enthusiasm for AI investment cools, the article says the major cloud providers would come under pressure without facing a systemic threat to survival. These firms hold some of the strongest corporate balance sheets on record. They may need to recognize impairments on some investments, but they would most likely survive. Even if AI falls short of expectations, their core businesses would still retain significant value.
Private credit firms, however, face two related risks. First, stronger AI coding capabilities may create uncertainty around legacy loans made to software companies. Second, financing for data center construction may offset weakness tied to software exposures, but it could also deepen losses if the capex cycle turns. Even in that scenario, the article returns to the same question: who ultimately bears the losses? Investors in private credit are mostly institutions and high-net-worth individuals, not typically banks, which are the type of intermediaries that can amplify losses through the system.
The article also notes other areas of concern, including circular financing relationships among chipmakers, cloud giants, and data centers, as well as the possibility that new risks could emerge later. Those issues should not be ignored. Even so, it says they can still be evaluated through the same three channels: economic activity, wealth effects, and credit.
Bubbles can leave something valuable behind
Historical episodes such as the 2008 crisis have given bubbles an almost entirely negative reputation. The article says that view is too narrow. Risk needs to be taken seriously, but most bubbles do not end in systemic disaster.
They can also leave behind meaningful positive macro legacies. Building infrastructure for new technologies often requires a period of optimism intense enough to mobilize large amounts of capital and talent. Without the expectation of very large wealth creation, many emerging technologies would never be deployed at scale. In that sense, the article suggests, bubbles can function as a mechanism for solving a collective action problem.
It points to the internet bubble as an example. Investor losses were enormous after the crash. Amazon shares fell by nearly 95%, and Global Crossing went bankrupt. But those losses do not capture the longer-term social and macroeconomic value created during the boom. Global Crossing’s investors were wiped out, yet the terrestrial and submarine fiber it built became a critical foundation for growth over the following two decades. Amazon, for its part, reshaped retail.
As long as systemic risk remains contained, the article argues, fear is not the most useful response.
Five suggestions for managers operating inside the AI boom
The article says bubbles, for better or worse, are a recurring feature of capitalism. Learning how to operate through them may be uncomfortable for executives, but it is unavoidable. It offers five suggestions.
1. Participate to preserve room for survival and growth
Many executives think their job is to avoid bubbles. The article warns that this choice can itself create existential risk. It gives the example of a large memory-chip maker that decides the current cycle is a bubble and waits while rivals invest aggressively. Even if the bubble later bursts and the diagnosis proves right, the company would still face an oversupplied market in which competitors have much greater capacity.
2. Identify the risk precisely
Large figures can shock and unsettle, but risk has to be measured against the specific question being asked and set in the broader economic context. The article recommends first mapping the transmission channels, then estimating the size of the headwind. If the focus is real GDP, the relevant question is how much investment flow actually affects real GDP growth.
3. Separate investor losses from macro risk
One thing is almost certain in any bubble: someone will lose money, sometimes a lot of it. But investor losses do not automatically amount to a macroeconomic crisis. In many cases, the damage stays local.
4. Understand the source of systemic risk
Losses are unavoidable when bubbles deflate, but their macro significance differs sharply. The article says long-term macro damage appears only when losses enter the banking system, erode capital, and trigger a credit crisis. That outcome is hard to predict in advance, but bank credit spreads are described as a signal worth watching.
5. Treat policymakers as firefighters
Monetary policymakers are responsible for protecting the economy, yet they are poorly placed to identify bubbles ahead of time, according to the article. Trying to pop a bubble preemptively can interrupt an expansion that is still producing jobs and wage growth. Unless there is evidence of systemic banking risk, policymakers are better off standing by and preparing to respond after the fact.
The article’s bottom line
So should the market fear an AI capex bubble? The article’s answer is that, for now, it remains a manageable macro risk, one that has to be weighed against its positive contribution to economic growth.
That judgment would need to change only if credible evidence emerges showing that the AI spending boom is materially damaging one or more of the three main channels: economic activity, wealth, or credit. The possibility exists, the article says, but it is not the base case today.
The article is by Philipp Carlsson-Szlezak and Paul Swartz. Carlsson-Szlezak is managing director and partner in Boston Consulting Group’s New York office and also serves as the firm’s global chief economist. Swartz is executive director and senior economist at the BCG Henderson Institute. Zhou Qiang is listed as editor. The piece was published via the WeChat account Harvard Business Review (ID: hbrchinese), with authorship attributed to HBR-China.