A new parlor game has taken hold in boardrooms and investment committees, and BCG economists Philipp Carlsson-Szlezak and Paul Swartz are calling it out in a September 2026 piece for Harvard Business Review1. "There's a new parlor game circulating global C-suites: Is there an AI capex bubble -- and if so, when will it pop?" Their answer, however, is not what most expect. The question itself is the problem.
Reframing the Risk
Carlsson-Szlezak and Swartz are direct: "Asking the right question matters, and 'When will the AI capex bubble pop?' is not it." Bubbles cannot be timed with certainty, and posing the question implies both that a bubble poses a macroeconomic problem and that investors and policymakers could sidestep the risk. Neither assumption holds.
The more productive frame begins with accurate sizing. Taking hyperscaler capex, adding estimates for private AI companies, and adjusting for imports and non-AI spend, the authors arrive at approximately $630 billion in 2026, just under 2% of U.S. GDP. The critical adjustment: around 50% or more of AI capex is spent on imports, particularly semiconductors, which do not drive domestic activity. The true domestic footprint is closer to $315 billion, or 1% of U.S. GDP. Bloomberg consensus estimates suggest that figure could rise to 1.5% of GDP by 2028.
The gap between $315 billion and the $3 to $4 trillion figures circulating in financial media is not a rounding error. It is a framing error with real consequences for how risk is assessed.
Three Channels, Three Constraints
Carlsson-Szlezak and Swartz map three transmission channels through which an AI capex bust could damage the broader economy. An abrupt activity stop would produce a direct drag of roughly 1% of GDP, meaningful but, by itself, not recession-defining on its own. A wealth-effect-driven equity sell-off carries weight given that equities now represent nearly 30% of household wealth, up 10 percentage points from a decade ago, yet two post-Covid bear markets in 2022 and 2025 produced only modest real-economy impact. The third channel, a credit crunch, is the most consequential, and it is here the authors draw their most important distinction.
Why This Is Not 2008
"To be structurally scarring, a bursting bubble must leave crippling losses in the banking system," Carlsson-Szlezak and Swartz state plainly. The contrast with 2008 is instructive. AI capex is financed primarily through hyperscaler free cash flow, corporate debt, and direct equity issuance. Private credit firms such as KKR and Apollo operate entirely outside the banking system. The authors conclude that any souring of AI investments "is not likely to weigh on their balance sheets the way housing did in the middle of the 2000s." Even in a bust scenario, hyperscalers hold some of the strongest balance sheets in corporate history. Private credit losses would fall on institutional investors and high-net-worth individuals, not on banks whose systemic role amplifies fallout across the economy.
Bubbles Can Be Generative
Perhaps the most counterintuitive argument in the piece: "If systemic risks are contained, it may be more sensible to cheer a bubble on than fear it." The dot-com era is invoked not as a cautionary tale but as a structural lesson. Amazon lost nearly 95% of its value. Global Crossing went bankrupt. But the fiber optic cables those investments funded became critical infrastructure for two decades of economic growth. "Think of a bubble as solving a collective-action problem," the authors write. Infrastructure built during a surge of exuberance can outlast the investment thesis that funded it.
"The only sure thing in a bubble is that someone will lose money -- and sometimes a great deal of money," Carlsson-Szlezak and Swartz note. "However, idiosyncratic losses are not a good guide to macroeconomic impact which is often contained." The conclusion is calibrated: the AI capex surge is "a manageable macroeconomic risk to be balanced against a welcome tailwind to growth." Changing that view would require credible evidence of material impact across all three transmission channels. That is not today's base case.
Five Key Takeaways for Advisors and Investors
- Calibrate to the right number. The $315 billion domestic AI capex footprint is the relevant figure for U.S. economic risk, not the $3 to $4 trillion headline. Scale without context misleads.
- The banking system is not the weak link this time. AI capex is financed outside traditional banking channels. Without bank-level stress, a bust would be painful for some investors but not structurally scarring for the economy.
- Bear markets alone do not make recessions. Two post-Covid drawdowns tested household balance sheets without triggering recession. A deeper, more sustained equity rout would be required to materially impair consumer spending.
- Bubbles can leave productive legacies. Investor losses and economic damage are not the same thing. The infrastructure being built today may deliver long-term productivity gains that outlast the hype cycle entirely.
- Watch banking credit spreads, not capex headlines. The key systemic signpost is whether AI-related losses migrate into the banking system. That is the early-warning indicator worth monitoring closely.
Footnote:
1 Carlsson-Szlezak, Philipp, and Paul Swartz. "The Questions You Should Be Asking About the AI Bubble." Harvard Business Review, 11 Sept. 2026, https://hbr.org/2026/09/the-questions-you-should-be-asking-about-the-ai-bubble.
