The AI infrastructure boom is usually discussed in terms of dollars spent and GPUs secured. Gartner projects global data center spending will hit $653 billion in 2026, a 31.7% jump over the prior year, while rising memory costs and supply constraints have pushed server prices up 15% to 20% in recent months. But according to commentary from Mitsubishi HC Capital America, the more consequential shift is happening beneath the headlines: how organizations own and finance the assets that make up a data center.
Different Assets, Different Clocks
A data center building can last decades. Power and cooling systems typically remain productive for 10 to 15 years. GPUs and servers follow entirely different refresh cycles. For years, many organizations treated all of these assets as if they depreciated on the same schedule, similar to financing a kitchen renovation on a 30 year mortgage. That approach simplified procurement and accounting but created inefficiencies as AI workloads scaled.
The result is a growing split in how buyers approach the market. Hyperscalers now place equipment orders 10 to 12 months ahead of need, using their scale to lock in pricing and supply. Newer AI cloud providers, by contrast, often prioritize speed over optimization, acquiring capacity as fast as possible to keep up with demand, even at the cost of efficiency.
Power Constraints Are Driving New Approaches
Utility and grid limitations remain a real bottleneck in many markets, but operators are increasingly turning to independent generation, geothermal technology, advanced turbines, and carbon capture initiatives to reduce dependence on traditional utility timelines. These strategies don’t eliminate power constraints, but they point to an industry finding more adaptable paths forward than expected.
GPUs Are Lasting Longer Than Assumed
Conventional wisdom held that GPUs would become obsolete quickly as new generations shipped. In practice, many organizations are finding the hardware remains productive well beyond initial expectations, with some extending GPU depreciation schedules to seven or eight years. That’s notable because it separates the investment narrative around AI, often focused on bubbles and disruption, from the lending narrative, which is grounded in actual utilization and repayment data. Lenders reviewing real performance numbers are not seeing the rapid obsolescence many market commentators predicted.
For hosting providers and site owners watching infrastructure costs ripple through the market, the takeaway is that data centers are no longer single assets on one economic timeline. Buildings, power systems, cooling, and compute now each carry distinct useful lives and financing considerations, and organizations that adapt their capital strategies accordingly will have more flexibility as AI infrastructure demand continues to evolve.