Most conversations about AI in insurance still start in the same place: claims automation, underwriting efficiency, chatbots. These are real applications, and they matter. But they describe insurance as a consumer of AI. The far more consequential question is what role insurance plays in deciding whether AI gets built at all.
That question only makes sense once you understand what AI has become. It is no longer a software problem, it is an infrastructure problem: arguably the largest civil engineering challenge of the century.
The physics of intelligence
The frontier AI systems being developed today are not clever algorithms running on ordinary servers. They are industrial-scale operations that consume electricity equivalent to mid-sized cities. According to the International Energy Agency, data centers are already consuming disproportionate shares of regional grid capacity in major technology hubs, and demand is projected to more than double by 2030.
The binding constraint is not software, but instead power. A data center shell can be built in under two years, but the electrical infrastructure to feed it face lead times of up to four years, according to Wood Mackenzie’s supply chain analytics. That mismatch is why Microsoft signed a multi-billion-dollar deal to restart the Three Mile Island nuclear plant, and why Amazon acquired a nuclear-powered data center campus from Talen Energy. These are not PR stunts, but rather moves to secure a key input that AI itself cannot generate: reliable electricity at scale.
McKinsey estimates that global data center capital expenditure will approach $7 trillion by 2030, with over $5 trillion driven by AI alone. This is no longer a technology sector, but a heavy infrastructure sector, with all the project finance, risk transfer, and capital structure complexity that implies.
Why efficiency won’t save us
The obvious pushback is that software will get more efficient, reducing the need for all this hardware. That is technologically true and economically irrelevant. Every time AI systems become more efficient, the companies operating them do not reduce power consumption. They use the same power budget to train larger, more capable models.
The economist William Stanley Jevons identified this dynamic in 1865: when you make a resource cheaper to use, people use more of it, not less. In AI, demand for intelligence is essentially infinite, so every efficiency gain gets reinvested at the frontier. The physical ceiling does not move.
The result is a market splitting in two. At the bottom, lightweight AI models are becoming free: they run on laptops and handle everyday tasks. At the top, a small number of frontier systems require tens of billions in ongoing capital expenditure to operate. The gap between these tiers is growing, and the value is currently accumulating overwhelmingly at the top.
Underwriting the Infrastructure
The technology sector cannot finance this buildout alone. Hyperscalers are collectively spending over $200 billion a year in AI capital expenditure, but still rely on project finance and debt syndication for the pace and geographic spread of the buildout. And project finance requires risk transfer.
As Aon has documented, data center project values now routinely reach $10–50 billion for a single campus. These values often exceed what the standard insurance market can cover in a single placement. Lenders require insurance coverage up to the full project value before making a final investment decision. If the project cannot secure coverage for construction risk, delay-in-startup, or business interruption, the project will not close, the financing stalls, and the data center does not get built.
This is not theoretical leverage, but a structural gating mechanism that is already operating. Insurance capacity is quietly determining which AI infrastructure projects move forward and which ones wait. The question is whether insurance leaders recognize this leverage and deploy it strategically, or continue treating data center policies as just another line of business.
Your data is worth more than you think
There is a second, less obvious source of leverage. The AI industry is running out of training data. Researchers at Epoch AI have projected that frontier laboratories will exhaust the stock of high-quality public text within the next few years. And training AI on synthetic data has been shown in a landmark Nature study to cause model collapse, where outputs degrade progressively with each generation.
What these labs need is real-world, empirical data: the kind that is verified by actual outcomes, tested by financial losses, and continuously refreshed. Insurers sit on exactly this: decades of catastrophe data, health outcomes, claims histories, supply chain exposures, and financial behavioral patterns, none of which is available on the public internet.
No major insurer has yet traded this data for equity in a frontier AI lab or for dedicated computing capacity, but the technical means exist. Through privacy-preserving techniques like federated learning, AI models can be trained on insurer data without that data ever leaving the insurer’s environment. The strategic question is not whether this is possible, but when someone will be the first to do it.
The insurance industry’s traditional self-image as a passive buyer of technology is dangerously outdated. In the physical AI era, insurers hold two assets the technology sector cannot replicate: the capital capacity to underwrite the infrastructure buildout, and the empirical data to keep frontier models grounded in reality.
The executives who recognize this will not spend the next five years optimizing claims workflows. They will be negotiating the terms on which the AI economy is physically constructed. That is a fundamentally different strategic position, and it is available right now, to those willing to see it.
Brandon Nuttall is Chief Digital & AI Officer at Xceedance, a global provider of technology-driven business solutions for the insurance industry, where he helps clients safely and securely incubate GenAI into their daily operations, enabling them to do more, faster. He has almost 20 years of experience in the insurance sector and a proven track record of curating ecosystems that combine the best of industry professionals and digital solutions.
