Bank of America has issued a stark warning that the current surge in artificial intelligence investment is beginning to look like the classic capital cycles that have historically ended in disappointment.
In a recently released note, investment strategist Sebastian Raedler highlighted that 12-month forward consensus expectations for hyperscaler capital expenditure have climbed from roughly $330 billion in early 2025 to $1 trillion.
That figure now equates to about 3% of U.S. GDP, higher than the 2.2% reached during the 1870s railroad boom and well above the 1.1% seen in the 1990s technology boom.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Raedler noted that most major capital cycles of the past concluded poorly for investors, even when the underlying technology ultimately proved transformative.
The typical pattern begins with high early profit margins that attract heavy investment. This leads to rapid capacity expansion, intense competition, falling prices, and eventually weaker returns as excess supply overtakes demand.
Signs of that sequence are already visible in AI. Aggressive competition among model providers has driven token prices down by approximately 50% since the summer.
The decline reflects an increasingly competitive market in which AI companies are under pressure to deliver more powerful models at lower prices.
The development is partly driven by the growing number of companies competing for a share of the AI market. Established players such as OpenAI, Google, and Anthropic face mounting competition from emerging developers, including Chinese AI companies that have introduced relatively inexpensive models capable of performing increasingly sophisticated tasks.
As providers compete for developers, enterprise customers, and AI-powered applications, pricing has become an important weapon for attracting users and expanding market share.
The introduction of more efficient models has further accelerated the decline in prices. Advances in model architecture, inference optimisation, and hardware utilisation have enabled providers to reduce the computational resources required to generate responses.
Some companies have also introduced smaller, specialised models that can perform specific tasks at a fraction of the cost of larger systems. These improvements allow providers to offer lower prices without necessarily sacrificing performance across every application.
However, the decline in token prices also raises questions about the long-term economics of the AI industry. When prices fall rapidly, companies must process significantly more requests to maintain or increase revenue, particularly if their operating costs do not decline at the same rate.
This creates pressure to improve efficiency, reduce infrastructure expenses, and develop new revenue streams. Providers with substantial financial resources, proprietary infrastructure, and access to advanced chips may be better positioned to withstand prolonged price competition than smaller companies operating with limited capital.
For businesses adopting AI, cheaper tokens represent an opportunity to integrate the technology into more products and services. Applications that were previously too expensive to operate at scale could become commercially viable as inference costs decline.
According to the Bank of America analysis, this kind of monetization pressure in the face of rising competition is precisely the failure mode that has defined previous capital cycles.
Market reaction has been swift. AI-related assets have come under pressure, with South Korea’s KOSPI falling around 20% since early October and the Philadelphia Semiconductor Index trading roughly 13% below its recent peak.
While artificial intelligence is widely expected to deliver lasting productivity gains, the scale and speed of current spending have raised questions about whether returns will keep pace with the capital being deployed.
Bank of America’s message is clear, the cycle is following a familiar script, and history suggests the later stages rarely deliver the same easy gains as the early ones.