AI is often portrayed as a software revolution. We picture algorithms, machine learning models and clouds operating somewhere in the digital ether. Yet behind every AI model sits an expanding network of highly physical infrastructure, hyperscale data centers, substations, transformers, cooling plants and power generation. As investment in AI accelerates, insurers, asset owners and valuation professionals are being forced to confront an increasingly important reality: AI may be digital in function, but its risk profile is fundamentally physical.
Few stories better illustrate the dependence of digital wealth on physical assets than the tale of the Bitcoin fortune reportedly worth GBP 695 million that was lost in a Welsh landfill. Bitcoin itself has no physical form, there are no banknotes or coins, but the private keys stored on a discarded hard drive were essential to accessing that value. Once the hard drive disappeared, so too did access to the fortune. The episode serves as a striking reminder that even the most intangible digital assets ultimately rely on physical infrastructure.
The same principle applies to AI, albeit on an entirely different scale. Rather than a single hard drive, modern AI depends on vast hyperscale data centers that can individually exceed 100,000 square meters, housing thousands of servers, networking equipment, transformers, switchgear, backup power systems and increasingly sophisticated cooling systems. Individual facilities now cost billions of dollars to construct, dramatically increasing capital intensity per square meter and creating new challenges for insurers managing concentrated, high-value risks. This rapid escalation in asset values is placing increasing pressure on insurance markets. Zurich Insurance recently warned that average data center project values in its portfolio have increased from around USD 150 million to USD 3 billion in just five years, while the industry’s ability to provide sufficient insurance capacity is becoming increasingly constrained as AI-driven projects continue to grow in size and complexity.

