LONDON – Concerns about an 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 startup phase and relatively modest. Facebook, for example, initially raised just $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 $7 trillion by 2030 as companies build data centers, 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 centers 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 processing units and tensor processing units (TPUs), 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 $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, fueling 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 centers, so that higher prices would not be passed on to consumers.
The End of Winner-Take-All?
Lastly, AI may be less susceptible to winner-take-all dynamics.
