S Yash Kalash is a Senior Fellow at CIGI and an expert in strategy, public policy, digital technology and financial services
The artificial intelligence boom is rapidly becoming an infrastructure boom. Across the United States, Europe, Asia and the Middle East, extraordinary amounts of capital are flowing into data centres, power generation, transmission networks and specialised real estate. Global real estate strategy firm JLL estimates nearly 100 GW of new data centres will be added between 2026 and 2030, with the sector experiencing what they call an investment supercycle requiring up to USD 3 trillion by 2030.
Much of this investment may prove necessary and AI could become one of the defining general purpose technologies of this century. But regulators should consider another possibility: what happens if technological change moves faster than the investment cycle financing it?
A data centre is financed on assumptions measured in years or decades, while AI technology, model economics and customer demand can change within months. Early expansion was largely funded by cash-rich hyperscalers, whereas AI labs relied on external capital. That balance is changing, for example infrastructure providers such as Oracle are taking on long-term investment obligations against contracted AI demand that possibly exceeds their near-term cash generation.
The risk is that debt and long-lived assets are being underwritten against technologies, customers and economics that may change far faster than the infrastructure itself. The Bank for International Settlements has warned that the scale of anticipated AI investment will increasingly require companies to move from internal cash flows toward debt, with private credit playing a growing role.
As the financing ecosystem expands to include banks, private-credit funds, institutional investors and increasingly sophisticated financing structures, technology risk can become financial risk. This creates a familiar equation reminiscent of the 2008 global financial crisis: optimistic demand forecasts, rapidly appreciating assets, leverage and confidence that the underlying asset will retain its value.
That said, the comparison with the 2008 global financial crisis should not be overstated. Data centres are not subprime mortgages and hyperscalers are not highly leveraged households. Still, the important lesson from 2008 was not simply that American homeowners borrowed too much, but rather that financial markets transformed a reasonable proposition — that housing would remain valuable — into a leveraged financial ecosystem whose stability depended on that proposition continuing to hold.
The dangerous assumption is not that AI will succeed, but rather that today’s architecture for delivering it will last long enough to repay everything being built around it
The relevant question today is therefore not whether AI resembles the housing crisis of the past, but whether AI infrastructure is beginning to acquire similar financial transmission mechanisms. The BIS has already identified an AI correction as a potential financial-stability risk, warning that a sharp reduction in hyperscaler capital expenditure could affect borrowers throughout the AI supply chain and ultimately transmit losses into corporate and private-credit markets.
The risk markets may be underpricing is technological obsolescence and evidence is already emerging. Deepseek in 2025 reported that their V3 model achieved performance comparable to leading closed models while activating only 37 billion of its 671 billion parameters.
Similarly, Moonshot’s Kimi K3 activates 104 billion of its 2.8 trillion parameters and claims approximately 2.5 times better scaling efficiency than Kimi K2. And finally, Apple now runs a roughly 3-billion-parameter model on-device, reducing reliance on centralised compute. Data centres are being built around assumptions concerning today’s chips, models, energy requirements and computing architectures.
Those assumptions could change dramatically as models may become substantially more efficient and smaller, move inference onto the edge and new semiconductors, and cooling technologies alter infrastructure requirements, while the geography of computing could increasingly follow abundant and inexpensive energy.
None of this means terrestrial data centres will disappear. The more important possibility is that specialized AI facilities built today may not retain the economic value assumed by their financing models. This deserves particular attention in emerging markets where data centres often depend on scarce long-term finance and substantial supporting investment in electricity grids.
In Malaysia, for example, IFC provided early-stage financing for a 300MW Johor campus because commercial lenders would not. If utilisation disappoints, the consequences may therefore extend beyond investors to lenders and electricity systems.
Technology companies and infrastructure investors increasingly promote cheaper land, construction costs, expanding electricity supply and skilled workforces as reasons for governments to facilitate data-centre investment.
Some projects will undoubtedly create genuine economic capabilities and governments may offer tax concessions, subsidised infrastructure or scarce electricity based on projections of future AI demand largely originating with the companies seeking capacity. The danger is that host economies socialise infrastructure costs while technology companies retain the flexibility to shift future workloads elsewhere.
Governments should require credible projections of utilisation, technological depreciation, employment, local value creation and long-term energy requirements before committing public resources. At the same time, financial regulators must start looking beyond conventional bank exposure.
They should map who ultimately owns data-centre risk across banks, private-credit funds, insurers, pension funds and securitised products. Stress tests should examine what happens when hyperscalers reduce capital expenditure, utilisation assumptions disappoint and specialised real-estate valuations decline simultaneously.
Crucially, scenarios should also include a world in which AI demand continues growing while the amount or type of infrastructure required to provide it changes dramatically — AI does not have to fail for an AI infrastructure bubble to burst.
History provides ample precedent. Railways transformed nineteenth-century economies while repeatedly bankrupting railway investors. The internet transformed the world despite destroying enormous amounts of capital after the dot-com bubble. During the Great Financial Crisis housing still remained indispensable even as securities constructed around assumptions about its value destabilised the global financial system.
A transformative technology and a speculative investment cycle can often coexist and thus regulators should begin treating the gap between the speed of technological change and the duration of infrastructure investment as a potential source of macro-financial risk. The dangerous assumption is not that AI will succeed, but rather that today’s architecture for delivering it will last long enough to repay everything being built around it.