Data center infrastructure is long-lived. The 20 to 40 year depreciation schedules applied to data center facilities reflect a real historical truth: the buildings last for decades. The assumption underpinning $7.6 trillion in projected AI infrastructure investment through 2031 is that the facilities being built now will be productive through the depreciation period.
The assumption is not obviously wrong for the building itself. It is almost certainly wrong for a significant portion of the infrastructure inside.
What the Data Actually Shows
Goldman Sachs' March 2026 analysis identified this as the central risk in the entire AI infrastructure investment thesis:
"Today, developers are realizing that data centers designed less than two years ago may be insufficiently provisioned for the next generation of cutting-edge AI chips, given their significant power and cooling demands. When the design requirements of a facility may shift materially within a few years of commissioning — and when ambitious new concepts like novel cooling architectures emerge — the rate of architectural change itself becomes a risk factor."
The specific numbers define the problem precisely. A standard server rack drew 5 to 10 kilowatts five years ago. AI-class racks drew 30 to 50 kilowatts in 2023. Current-generation Blackwell systems draw 120 to 132 kilowatts per rack. Next-generation designs are pushing toward 200 kilowatts per rack. Each generation roughly doubles the power density requirement.
Facilities designed for 15 to 30 kilowatt racks cannot support 120 kilowatt racks without fundamental redesign of the power delivery and cooling infrastructure. The building stays. The power distribution, cooling systems, and structural supports for the dramatically heavier liquid cooling infrastructure all need replacement.
The capital cost of that replacement is not trivial — liquid cooling infrastructure adds $500,000 to $2 million per megawatt of capacity in capital costs — and it is not in the depreciation models for the original facilities.
The data center world is now explicitly distinguishing between "transitional AI data centers" — facilities designed in 2022 to 2023 that are already insufficient for 2025 to 2026 hardware — and purpose-built AI facilities designed from the ground up for current power densities.
The transition period was two years. Data centers that completed construction at the beginning of the AI buildout wave are already technically obsolete for cutting-edge AI training workloads.
Separately, the question of whether AI data centers built for training workloads will retain economic value as inference displaces training as the dominant workload is the one Goldman Sachs identifies as having the greatest sensitivity in their model.
Training workloads are dense, sustained, batch-oriented, and centralized. Inference workloads are variable, latency-sensitive, and geographically distributed. The same facility design cannot optimally serve both.
The Specific Mechanism of Failure
The failure has a name in the investment world: stranded asset risk.
An asset becomes stranded when its economic value is destroyed by an external development — regulatory change, technology shift, or demand substitution — before the end of its expected financial life.
The stranded asset mechanism in AI infrastructure is technological obsolescence at the cooling and power delivery layer, combined with workload shift at the compute layer.
The timeline is compressed relative to all prior technology infrastructure cycles because the rate of architectural change in AI hardware is faster than in any prior computing era.
Nvidia's annual product cycle, combined with 10x performance improvements per generation, means that a 2-year-old facility is designed around hardware assumptions that have been superseded by two generations of improvement.
The accounting treatment amplifies the risk. Facilities are being financed with long-term debt (10 to 15 year maturities) and depreciated over 20 to 40 years.
If the economic life of the AI-specific components within them is 3 to 5 years, the financing structure outlasts the productive asset. Interest payments on the debt continue long after the infrastructure inside has been replaced or abandoned.
CoreWeave's $7.5 billion in interest payments through the end of 2026 is the most visible version of this math, but the dynamic is present at every scale.
The Industry Cost
The estimate that AI data centers to be built in 2025 will suffer $40 billion in annual depreciation while generating $15 to $20 billion in revenue is the clearest statement of the stranded asset problem.
The asset base is being built faster than the revenue to service it is arriving. And because the economic life of the AI-specific components is shorter than the accounting life, the depreciation understates the actual rate of value destruction.
The Goldman Sachs sensitivity analysis makes the magnitude visible: shortening average chip useful life from 5 years to 3 years in their model increases annual depreciation costs by nearly $1 trillion.
The infrastructure is not aging physically. It is aging architecturally.
That is not a scenario analysis. That is a description of what the evidence suggests is actually happening, valued at $1 trillion in understated annual cost.
For the broader economy, the implications extend to energy planning. Data centers consuming US electricity at rates projected to reach 9 to 12% of national demand by 2028 are signing long-term power purchase agreements and grid interconnection agreements based on demand forecasts that assume current AI architectures scale linearly.
If architectural efficiency improves at the rate AI hardware improvements have historically delivered, the power consumption forecasts could be wrong by orders of magnitude in either direction.
The grid is being built for a demand curve that the technology itself may or may not produce.
What Needs to Exist
The industry lacks an independent AI infrastructure obsolescence assessment framework.
Such a framework would evaluate deployed data center assets against current and projected hardware requirements, scoring them on:
- Power density compatibility
- Cooling architecture adequacy
- Networking capability
- Workload profile fit
- Remaining economic life
This assessment would value infrastructure on its actual remaining economic life, not its accounting depreciation schedule.
It does not exist.
Every infrastructure investor, every data center operator, and every hyperscaler finance team is making this assessment privately, with different assumptions and different incentives.
A standardized, independent framework would immediately be used for infrastructure investment decisions, refinancing negotiations, asset acquisition due diligence, and regulatory disclosure.