• Bank of America (BAC) warns that the AI investment boom is displaying classic capital-cycle risks: surging spending, intensifying competition, and falling prices.
  • Hyperscaler capex expectations have jumped to roughly $1 trillion, or about 3% of U.S. GDP, while AI token prices have plunged 50% since summer.
  • The key threat: AI technology may succeed, but investors could suffer from overinvestment and declining returns.

Bank of America is cautioning that the artificial intelligence investment boom is beginning to mirror the classic pattern of historic capital busts, where heavy spending and cutthroat competition erode returns even as the technology itself advances.

In a note published October 16, strategist Sebastian Raedler outlined a familiar sequence: high early margins, aggressive investment, excess capacity, and weakening pricing power. The warning is not a prediction that a bust is inevitable, but rather a risk assessment that the AI infrastructure buildout may be reaching a precarious phase.

The scale of spending is staggering. Consensus expectations for hyperscalers’ next-12-month capital expenditure have tripled from $330 billion in early 2025 to approximately $1 trillion, equivalent to roughly 3% of U.S. GDP. That figure, however, is a forward spending estimate—not money already spent, and not necessarily earmarked exclusively for AI.

Meanwhile, pricing is deteriorating. Average AI token prices have reportedly fallen about 50% since summer, partly due to competition from Chinese open-weight models. Tokens are the units of text processed or generated by language models. The report does not provide enough methodology to establish that every provider or workload experienced the same decline, but the trend is clear.

Usage and revenue are diverging. Raedler cites Ramp data showing U.S. business AI spending easing slightly since August despite sharply rising token volumes. That suggests weaker monetization within that dataset, though it is not proof that worldwide AI revenue is falling. Financing risk is also mounting. Reuters (TRI) reported on October 7 that major technology companies are increasingly turning to debt and equity markets to fund AI and cloud expansion, moving beyond the sector’s traditional reliance on internally generated cash.

The economic test is straightforward: lower prices must stimulate enough additional usage—or be accompanied by sufficiently lower production costs—to preserve profits. Growing adoption alone does not guarantee an adequate return on investment.

Market weakness is already visible. The October 16 account of BofA’s note reports South Korea’s KOSPI down 20% since early October and the Philadelphia Semiconductor Index 13% below its recent high. Those declines cannot be attributed solely to this warning, but they underscore investor anxiety.

Bank of America’s own third-quarter results, released October 14, provide context for the institution issuing the research. Net income fell 17% year over year to $6.93 billion, and earnings per share came in at $0.91 reported, or $1.08 excluding a regulatory penalty. Revenue rose 3.2% to $30.12 billion but missed expectations. A $1.2 billion regulatory penalty and weaker-than-expected trading results weighed on earnings—issues separate from the AI research warning. Shares fell almost 5% on the earnings day.

The broader implications are significant. The buildout supports demand for computing equipment, construction, power infrastructure, and related services. But BofA’s infrastructure research identifies electricity, labor, and supply chains as constraints, arguing that sustainable expansion depends on utilization, cash flow, and productivity gains rather than spending alone. Debt financing adds another pressure point. Reuters reported in July that Goldman Sachs (GS) expected approximately $750 billion of hyperscaler capex in 2026 against roughly $778 billion of operating cash flow—an earlier estimate not directly comparable with BofA’s latest rolling figure, but illustrative of how investment can absorb most operating cash generation.

Political and international developments add layers of complexity. U.S. federal regulators have ordered grid operators to consider new connection rules for large electricity users, including faster pathways for facilities offering flexible power consumption. The European Union’s AI Act became broadly applicable on August 2, 2026, adding compliance considerations. And on October 9, Reuters reported Chinese guidelines calling for tighter monitoring of technology risks and accountability for major losses from “blind” investment in AI and other frontier technologies. Concerns about overbuilding extend beyond the United States.

International competition cuts directly into the investment thesis. BofA identifies Chinese open-weight models as a source of pricing pressure. The competitive outcome may be wider access to capable AI but weaker pricing power for incumbent providers. The gains and risks fall on different stakeholders: businesses and consumers may benefit from cheaper AI services, while shareholders and lenders face the possibility that revenue growth will not cover infrastructure and financing costs.

Raedler compares today’s spending intensity with the 1870s railroad boom and the 1990s technology boom, citing investment equivalent to approximately 2.2% and 1.1% of U.S. GDP respectively, against today’s reported 3%. These are BofA’s comparisons, not independently reconstructed estimates. The underlying precedent is that useful infrastructure can outlast disappointing investments. Heavy spending creates capacity; competition then lowers prices, transferring benefits to customers while reducing investors’ returns. BofA contrasts this with the cloud industry of the 2010s, which it says avoided that outcome because high entry barriers restrained competition. Its concern is that competition at the AI model layer is much less constrained.

The central question in the near term is whether paid demand catches up with spending commitments. BofA’s bearish evidence is the combination of falling token prices and softer spending in Ramp’s business dataset. The bullish argument invokes Jevons paradox: cheaper computing could stimulate enough additional use to increase total spending despite lower unit prices. Neither outcome is established by the current evidence. Power availability is an immediate execution risk as well. A September 29 Reuters commentary reported potential delays of up to a year for an Oracle (ORCL) AI data-center project in New Mexico because of power constraints. Separately, Reuters reported October 8 that technology companies and utilities were pursuing flexible data-center power consumption to ease grid bottlenecks.

Longer term, the plausible outcomes are not limited to universal success or collapse. Strong utilization and productivity gains could validate much of the investment. Persistent price competition could produce useful AI services but poor returns for some suppliers. Uneven outcomes could reward efficient operators while forcing weaker projects to retrench. BofA’s broader infrastructure analysis offers an important counterweight to the headline: it describes a buildout with real economic foundations and remaining investment capacity, but increasingly demanding project economics. The central test is whether spending produces sustainable economic value—not whether the spending number is large by itself.

Editor’s note: This article was updated to clarify that the $1 trillion figure is a forward capital expenditure estimate for hyperscalers, not confirmed spending.