The debate around depreciation life for AI infrastructure Nov 2025
Our View around depreciation life for AI infrastructure
The debate around depreciation life for AI infrastructure is ultimately a debate about the true economic durability of the current AI hardware cycle. Extending useful life from four to six years boosts earnings today, but it assumes the pace of AI innovation will slow enough for existing chips and systems to remain competitive for much longer.
We find this assumption difficult to accept without qualification. The performance curve in AI compute is still steep, model architectures are shifting rapidly, and power efficiency is becoming a defining constraint. In such an environment, the economic obsolescence of AI hardware may occur well before physical obsolescence.
This doesn’t necessarily mean hyperscalers are “wrong” — they have strong incentives to smooth earnings and, unlike most companies, they can re-deploy older hardware to less demanding workloads. But it does mean that investors should be cautious in treating the new depreciation schedules as a reliable indicator of real asset life.
Investment Implications
- Margins may be overstated. If hardware turns over faster than the new schedules imply, future depreciation charges will need to rise again, pressuring operating margins.
- Capex intensity remains structurally high. Even if companies claim longer useful lives, competitive reality may force continued rapid replacement cycles.
- Valuations may not reflect true replacement costs. Markets may be underpricing the long-term capex drag embedded in AI infrastructure.
- Short sellers like Burry raise a legitimate point: when technological obsolescence is the binding constraint, accounting lives become aspirational rather than predictive.
Conclusion
Our position is that the current six-year assumptions are optimistic. The discrepancy between accounting life and economic life is likely to widen before it narrows, and this gap will become increasingly important in assessing AI-exposed equities.
