Concerns about an artificial intelligence ( AI ) bubble have intensified since the start of 2026, as investors and policymakers focus on whether, and when, it might burst. But the real question is not whether current valuations are inflated; it is whether AI’s emerging business model differs from those of earlier technological revolutions.

For decades, scale has been the primary driver of tech companies’ performance and valuations. As apps, websites, online retailers and social media platforms expanded their user bases, marginal costs fell, network effects took hold and pricing power increased. Valuations came to reflect long-term growth potential rather than short-term profitability.

The forces that defined past tech winners are unlikely to dominate AI’s rollout, because the competitive dynamics differ across six critical dimensions. First, capital expenditure is no longer a shallow moat; it is a formidable barrier. In earlier technological waves, capital requirements were largely confined to the start-up phase and relatively modest. Facebook, for example, initially raised just US$500,000 in seed funding.

But those earlier innovations were built on top of existing infrastructure, such as Linux, Apache, MySQL and PHP ( the so-called Lamp stack ), which dramatically lowered upfront costs. AI, by contrast, is extraordinarily capital-intensive. Industry-wide capital investment is projected to exceed US$7 trillion by 2030 as companies build data centres, expand computing capacity, and invest in specialized hardware. Unlike previous tech cycles, these investment requirements will not fade as the industry matures and may even intensify.

Moreover, those costs may never decline meaningfully, since the lifespan of data centres is often measured in years, not decades. While cloud computing also required massive investment in general-purpose servers, AI demands entirely new infrastructure, including graphics- and tensor-processing units, to handle the vast number of simultaneous calculations involved in training and running AI models.

Such systems are expensive and energy intensive. A single large-scale AI training run is expected to cost over US$1 billion by 2027. Only firms that can afford the entry price will survive, giving today’s tech giants – with their enormous cash flows, robust balance sheets, and access to capital markets – a decisive advantage.

Second, AI’s operating-cost structure undermines traditional economies of scale. In earlier tech cycles, marginal costs per user collapsed as platforms grew. Whether it was social media, software, or ride-sharing apps like Uber, costs were spread across an expanding customer base, enabling platforms to sustain high margins as they scaled up.

Those models were also marked by low operating expenses. Once Facebook reached sufficient scale, the marginal cost of adding users became negligible. As a result, companies paid little attention to the cost of serving each user, as it rarely threatened financial viability.

AI flips these dynamics. Controlling marginal costs is no longer optional, since large language models and other AI systems incur significant costs with every interaction, which requires billions of calculations. This is why AI firms focus on reducing per-query costs through custom hardware like TPUs and by developing leaner, more efficient models such as China’s DeepSeek.

Scale is not enough

The third area where AI departs from previous tech revolutions is in the weakness and fragility of network effects. Legacy tech platforms benefited from self-reinforcing growth. Buyers and sellers were drawn to Amazon’s marketplace precisely because activity was already concentrated there.

AI users can switch easily between models, use several at once – one for text, another for images, a third for coding – or even build their own. Switching costs are low and loyalty is weak, making network effects far less influential in determining long-term winners.

For legacy tech companies, the combination of falling marginal costs and network effects amplified the benefits of scale, fuelling a race to capture as many eyeballs as possible. That strategy made sense for companies like Facebook, which created value by monetizing consumer attention through advertising.

AI companies confront a different cost structure. Each new iteration of their product requires additional capital investment. Every additional user increases costs, particularly inference costs. While training expenses can be amortized across a larger user base and some economies of scale may emerge, usage growth still leads to higher operating costs.

The fourth difference lies in the shift from market fragmentation to instant saturation. Earlier tech platforms grew within largely siloed markets: Google dominated search; Amazon focused on retail. By seeking distinct niches like college students ( Facebook ) and professionals ( LinkedIn ), companies had time to mature before competition intensified.

AI, by contrast, is a general-purpose technology that cuts across industries. With users able to gain access to it instantly through apps or application programming interfaces, companies no longer have the luxury of reaching maturity before competitors emerge. This dynamic gives AI the potential to disrupt not just individual sectors but every existing technology business model.

Fifth, political influence now matters as much as market power. Earlier innovation waves did not require companies to engage with governments and regulators to the extent AI must. While social media platforms eventually faced scrutiny over their addictive effects, the perceived risks posed by today’s emerging technologies are deeper, and, in many ways, existential, given AI’s potential to cause job displacement, exacerbate inequality and undermine democratic governance. With AI companies confronting both market forces and political pressures, firms that can shape regulation, influence public opinion and absorb reputational risk are better positioned to succeed.

Microsoft is a prime example of such a firm. In a clear effort to gain political and social legitimacy, the company recently pledged to cover the electricity costs of its data centres, so that higher prices would not be passed on to consumers.

End of winner take all?

Lastly, AI may be less susceptible to winner-take-all dynamics. Scale, near-zero marginal costs, and strong network effects enabled companies like Facebook, Google, Amazon and Apple to dominate social media, search, e-commerce and smartphones, respectively. The AI sector, at least initially, is unlikely to follow that pattern. Rather than converging on a single monopolistic winner, it could support multiple dominant players, each controlling its own niche.

To be sure, an AI company could reach a point at which its technological lead becomes self-reinforcing and effectively insurmountable. Through continuous self-improvement and overwhelming product superiority, or even the development of artificial general intelligence, such a firm could achieve lasting market power, allowing it to dominate the field.

Until then, investors must recognize that AI follows a new strategic logic. Applying legacy technology metrics to this rapidly evolving landscape is not only counterproductive but potentially costly. Investors who rely on past playbooks risk becoming the losers in today’s AI-driven market.

Consider stock-based compensation. Historically, equity incentives enabled tech companies to hire and retain talent, acquire intellectual property, and expand through mergers and acquisitions. But stock options cannot pay for data centres, computing power or energy infrastructure. To meet these needs, AI companies require real investment, established cash flows and reliable access to capital markets.

Similarly, investors once tolerated negative margins so long as user growth was robust and advertising revenues were growing. But the uncertainty surrounding AI and the scale of required capital expenditure limit their ability to assess when these investments will break even or how AI-driven transformations will ultimately increase margins. The result is a growing emphasis on strong balance sheets and demonstrable financial resilience.

So, the race for AI leadership will not be won by the companies with the most users or fastest growth rates. Instead, the victors will be firms that can combine superior products with financial strength and political influence.

In this sense, AI more closely resembles the capital-intensive industries of the mid-20th century than the asset-light tech models of recent years. With operating costs rising and consumers moving easily between models, profitability will depend on capturing elastic demand while translating political capital and regulatory influence into lasting competitive advantage.

Dambisa Moyo is an international economist.

Copyright: Project Syndicate