The wrong lesson from the dot-com bubble

Every major technology wave tends to invite comparisons with the a previous one. Today, the obvious parallel is between the artificial intelligence boom and the dot-com era. The comparison is useful, but the lesson today is not that investors should avoid transformative technologies. It is that technological success does not necessarily determine which companies will capture the returns.

Concentration in stock markets often reflects a genuine transformative theme in economies and corporates. But while the theme itself may endure, market concentration tends to wane over time. The internet, for example, transformed communication, commerce, and the global economy and yet many of the companies that financed and built its early infrastructure ultimately proved poor investments. Technological success and investment success are not the same thing, and by the end of the following decade only a few of those internet titans could still be counted amongst the world's largest companies (see Figure 1).

During the telecom boom, enormous sums were invested in fiber-optic networks on the broadly correct assumptions that internet adoption would grow strongly. What proved much harder to forecast was the timing of demand, the pace of technological change, and where future profits would accrue. Rapid advances in fiber technology increased the capacity of existing networks, even as long-haul fiber deployment accelerated. By 2001, it was clear that substantial overinvestment had occurred. The infrastructure endured; many of the original financing assumptions did not.

What makes the AI buildout different?

The AI buildout is frequently portrayed as another example of infrastructure being constructed ahead of demand. There is, however, an important difference. Data centers are not being constructed solely in anticipation of a hypothetical future application—cloud computing, digital services, and AI workloads already consume substantial capacity.

That distinction does not mean every investment will succeed. Efficiency gains in chips, models, and software could change the amount and type of computing capacity required. Power constraints may delay projects, while new capacity could eventually weaken pricing. It should also be noted that utilization alone does not guarantee a strong return on invested capital.

Where might AI value accrue?

For investors, the question is no longer whether AI will matter, but which companies can convert its adoption into robust returns for shareholders. Markets have so far rewarded businesses where AI-related growth is already visible. Semiconductors and physical infrastructure have benefited most directly from the initial buildout, while parts of the software industry have been treated more cautiously despite resilient earnings. This divergence is evident in the returns across the AI ecosystem as investors appear to be distinguishing between today’s suppliers of capacity and tomorrow’s owners of economic value. (See Figure 2).

History suggests that this becomes harder as a technology matures. Much of the value created by the internet accrued to businesses that exploited cheaper and more abundant connectivity, not necessarily to those that first financed it. AI could follow a similar path. The infrastructure may be essential and well used, while returns still vary sharply according to competition, pricing power, adoption, and capital discipline.

Why portfolio construction matters

This is precisely where disciplined, risk-controlled portfolio construction matters. When the range of plausible outcomes is wide, confidence in a single forecast is a weak substitute for portfolio resilience.

The objective is to participate in the important AI-related theme, but without allowing it to dominate the portfolio. That means broadening the opportunity set across infrastructure providers, enablers, adopters, and potential disruptors; combining complementary stock-level insights; and constraining unintended exposures to companies, sectors, factors, and narratives.

A systematic process can compound many modest views while continuously reallocating risk as the evidence changes. Importantly, diversification begins long before securities enter the portfolio. A broad alpha model can bring together multiple complementary perspectives, from fundamentals andvaluation to earnings, sentiment, catalysts and macro context, reducing reliance on any single dataset, narrative, or analytical technique. The portfolio then extends that diversification through explicit risk controls and position sizing.

The enduring lesson from the dot-com era is not to avoid transformative infrastructure. It is to separate useful capacity from profitable capacity, and technological conviction from portfolio concentration. Investors do not need to predict one definitive AI winner. In uncertain markets, diversification of insight, diversification of holdings, and explicit risk controls are not a retreat from active investing. They are what make active insight investable. Our investment process is designed to participate in value creation across several possible paths, while aiming to limit the cost of being wrong about any one of them.