For the past two years, the conversation around AI has focused largely on software. We debate models, benchmarks, agents and the race between technology companies. The story is usually told as one of innovation.
But another story is unfolding beneath the headlines.
The AI boom is not merely creating digital products. It is driving a wave of physical investment. Data centres are rising across regions. Utilities are expanding capacity. Water systems are being drawn into planning discussions that would once have seemed unrelated to software. AI increasingly looks less like an application layer and more like infrastructure.
That distinction matters.
When people compare today’s enthusiasm around AI to the dot-com era, they often focus on valuations. They ask whether technology stocks have become detached from reality. It is a reasonable question. Yet it may be the wrong one.
The more interesting comparison concerns where the money is going.
The dot-com boom generated extraordinary optimism. It created companies, websites and fortunes. But much of its value existed in expectations about a digital future that had not yet arrived.
Every major technological revolution eventually encounters physical constraints.
When the correction came, it was painful. Many investors suffered significant losses. Yet much of the physical economy remained relatively insulated from the damage.
Today’s AI boom appears different.
The investment is no longer confined to software companies and venture capital portfolios. It reaches into energy grids, construction projects, semiconductor manufacturing, cooling systems and industrial supply chains. In effect, AI is becoming a customer of the real economy.
When physical systems set the pace
History offers a useful reminder here. Every major technological revolution eventually encounters physical constraints. Railways needed steel and land. Electrification required generation capacity and transmission networks. The internet ultimately depended on fibre, servers and logistics systems.
And the AI boom is beginning to run into older systems that are harder to accelerate.
Energy is one. Every breakthrough model ultimately runs on electricity. The more capable AI becomes, the more dependent it becomes on power generation and distribution systems that move at a slower pace than software.
Water is another. Data centres need cooling, land and local approval. These are not abstract inputs. They are civic and environmental questions.
Like many technologies before it, AI generates externalities that extend beyond its users. Energy demand, water consumption and emissions are not borne solely by the companies that build models. These costs are increasingly shared by utilities, communities and, ultimately, society.
Capital may be the most important of all. Much of the current growth depends on continued spending. Data centres, chips and network infrastructure require enormous upfront investment.
The assumption is that future demand will justify those expenditures. Perhaps it will. But assumptions about future demand are often where technological booms become vulnerable.
Venture capital has always depended on a degree of overconfidence. Entrepreneurs believe they can achieve what established firms cannot. Investors believe they can identify the winners before everyone else. The challenge is that confidence can remain high long after the economics have become uncertain.
Ultimately, technologies do not become durable businesses because they attract attention. They become durable businesses because they generate sustainable cash flows. That distinction is easy to overlook during periods of rapid growth, when user adoption and market share often receive more attention than profitability.
There is another cost question, too. Many AI tools are still priced as if the full cost of running them can be deferred. Low subscription prices encourage adoption, but they may not reflect the true cost of compute, energy, chips and infrastructure. In that sense, consumers may be subsidised by investors still willing to fund the gap.
Part of the challenge is that the industry’s preferred pricing mechanism may not align neatly with customer value. Many frontier AI services are also billed through token consumption. That works well when experimentation is the goal. It becomes more difficult when organisations begin scrutinising costs at scale.
Token-based pricing effectively charges customers for computation. Yet most customers are not buying computation. They are buying outcomes. The more capable models become, the more pressure there may be to disconnect pricing from tokens and connect it to measurable business value.
That transition is not guaranteed to be easy. If providers continue charging less than the underlying cost of compute, investors may end up subsidising usage. If prices rise to reflect the true cost of infrastructure, demand may prove less elastic than today's growth projections assume.
Why the possible fallout could travel further
For years, the technology sector presented itself as an escape from the constraints of the physical world. Software was supposed to be infinitely scalable. Digital products seemed liberated from geography and material limits.
Yet the AI revolution is pulling technology back into the world of power grids, construction projects and resource allocation.
The AI economy is becoming more dependent on physical systems.
None of this means the AI boom is destined to fail. Many transformative technologies experienced periods of over-investment before eventually reshaping society. Railways did. The internet did.
Some infrastructure built during speculative periods later became genuinely valuable.
The challenge is that timing matters as much as direction. Markets often struggle to distinguish between a technology that will eventually change the world and an investment cycle that has temporarily outrun reality.
There is also a political economy to this boom. Big technology firms have become deeply entangled with governments, infrastructure planning and national competitiveness narratives. If the bet pays off, the rewards will be concentrated. If it disappoints, the costs may be more widely distributed.
That is the uncomfortable part. A bursting AI bubble would not only hurt venture capitalists or technology executives. It could affect ordinary pension fund holders whose savings are exposed through index funds, retirement portfolios and the wider market. The people least involved in making the bet may still share in the losses.
That may be the central question facing investors and policymakers today.
The risk is not simply that AI companies become overvalued. The risk is that entire sectors begin to organise themselves around expectations that prove to be slower, smaller, or more complicated than anticipated.
When that happens, corrections may no longer remain confined to technology stocks. They may ripple outward through the systems that were built to support them.
AI technology may be moving quickly. The systems beneath it are not. Power grids, water infrastructure and major data centre projects tend to evolve over years rather than months.
Perhaps that is why we keep talking about AI as software. We notice the models because they are visible. We overlook the infrastructure because it is not.
Yet all intelligence ultimately runs on physical systems.
Share
