AI hyperscalers slashed annual depreciation expenses by $18 billion in 2024 through a single accounting move: extending GPU useful life from 3-4 years to 6 years. Between 2023-2024, Google, Microsoft, Oracle, and Amazon all normalized on six-year depreciation schedules, reducing collective data center depreciation from $39 billion to $21 billion annually—a 46% accounting reduction. Prominent investor Michael Burry warns hyperscalers are understating billions in expenses, while Microsoft CEO Satya Nadella admitted “I didn’t want to get stuck with four or five years of depreciation on one generation.” With $300+ billion in AI capex planned for 2025, the depreciation debate determines whether hyperscalers infrastructure bets are profitable or a ticking accounting bomb.
The Coordinated Shift to Six-Year Schedules
Amazon led the accounting shift, extending AWS server depreciation from 4 years to 5 in 2022, then to 6 years in 2024. Google and Oracle quickly followed, banking on six-year useful lives for their data center GPUs. Microsoft hedges with “2-6 year equipment lifespans” in SEC filings but trends toward the longer end. By 2023, all three major hyperscalers had coordinated on the same playbook: double the depreciation period, cut annual expenses in half.
The financial impact isn’t subtle. According to Cerno Capitals financial analysis, extending depreciation from 3 to 6 years reduced collective expenses from $39 billion to $21 billion in 2024. Moreover, with $300+ billion in capex planned for 2025, the difference between 3-year and 6-year schedules now exceeds $23 billion annually. Thats not rounding error—its an $18 billion boost to reported profitability based purely on accounting assumptions, not operational improvements.
Critics argue this is accounting engineering, not economic reality. If GPUs actually depreciate faster than 6 years, hyperscalers face massive write-downs when theyre forced to acknowledge the true value of aging hardware.
Economic Life vs Technical Life: The $40 Billion Gap
The core debate centers on a distinction hyperscalers would rather ignore: GPUs may remain technically functional for 6+ years, but can they generate competitive revenue that long? Annual depreciation on 2025 GPU builds could reach $40 billion, yet rental revenues from those same assets may yield just $15-20 billion—a 2-2.7x gap between costs and returns.
Nvidias generation cycle undermines the 6-year narrative. The A100 launched in 2020, H100 in 2022, H200 in 2024, and Blackwell B200 later in 2024—a consistent 2-year cadence with performance leaps that make older generations less competitive. Blackwell B200 delivers 2.5x the performance of H200 and 11-15x faster LLM inference than H100. Furthermore, if customers demand cutting-edge performance, older GPUs risk sitting idle or commanding drastically lower rental rates.
Market pricing confirms the concern. H100 rental rates crashed from $8/hour in late 2024 to $2.85-3.50/hour by early 2025—a 64% drop in just six months. When CoreWeave charges a premium $6.16/hour while market rates collapse, it signals either exceptional service quality or a market repricing older GPU generations downward. Neither scenario supports 6-year economic viability assumptions.
CoreWeaves Evidence: Retention Under Boom Conditions
CoreWeave, a leading GPU cloud provider, offers real-world counter-evidence. According to CNBC reporting, their A100 GPUs—now 4+ years old—are “all fully booked,” and H100s from 2022 are being rebooked at 95% of original rental prices when contracts expire. This suggests GPUs can retain value beyond the traditional 3-4 year depreciation cycle.
However, context matters. CoreWeaves retention data reflects 2023-2025 market conditions—peak AI boom with insatiable demand for any available compute. When H100 market pricing drops 64% but CoreWeave maintains bookings at 95% of original rates, it reveals demand-dependent economics. Consequently, if AI demand plateaus or Blackwell adoption accelerates, older GPU utilization could crater regardless of technical capability.
Moreover, CoreWeaves profitability tells a cautionary tale. GPU rental gross margins hover at just 14-16% after labor, power, and depreciation. CoreWeave spent over $14 billion in 2025 and plans to double that in 2026. Thin margins plus massive capex equals zero room for error if rental rates stay depressed or utilization drops below the 80% financial models assume (while 60% is more realistic for mixed workloads).
Executive Doubts and Investor Warnings
Nadellas public statement contradicts Microsofts own accounting. While Microsoft files show 6-year depreciation trending, the CEO warns against “getting stuck with four or five years of depreciation on one generation” because “Nvidias pace increased in terms of their migrations.” If the executive responsible for Azures $50+ billion annual capex worries about generational lock-in, it signals internal concern that 6-year assumptions may not survive rapid GPU evolution.
Microsofts 2-6 year range is the widest among hyperscalers, suggesting the company is hedging. Oracle and Google commit firmly to 6 years, Amazon progressively extended from 4 to 6, but Microsoft leaves flexibility to adjust if GPU obsolescence accelerates. That flexibility may prove prescient.
Michael Burry—the investor who correctly predicted the 2008 housing crisis—publicly argues hyperscalers are understating depreciation by billions. His warning carries weight precisely because his track record shows willingness to call out unsustainable accounting before markets force corrections. If Burry is right and hyperscalers face write-downs when 6-year schedules prove unrealistic, shareholders will see quarters of profits evaporate overnight.
Market Implications: Utilization and Profitability Under Stress
The GPU rental market is stress-testing depreciation assumptions in real-time. When H100 rates plummet from $8 to $1/hour at budget providers while A100s trade for $0.66-0.78/hour, it demonstrates how quickly older-generation GPUs lose pricing power. Hyperscalers can absorb margin compression through integrated cloud services revenue, but pure-play GPU providers face existential risk.
Financial models assume 80% GPU utilization, but 60% is more realistic when accounting for maintenance, workload variability, and periods between contracts. That 20-point gap compounds depreciation concerns—if GPUs sit idle 40% of the time AND rental rates drop 60-80%, the economics collapse entirely.
The coming year will clarify which scenario prevails. If Blackwell adoption accelerates and H100/H200 utilization drops, hyperscalers 6-year assumptions face market rejection. If mixed workloads (premium training on latest GPUs, budget inference on older hardware) maintain utilization across generations, the extended schedules may hold. Watch GPU rental price trends and hyperscaler capex disclosures—theyll signal whether the industry believes its own depreciation schedules.
Key Takeaways
- Hyperscalers cut $18 billion in annual depreciation by extending GPU useful life from 3-4 years to 6 years (2023-2024), reducing collective expenses from $39B to $21B
- Nvidias 2-year generation cycle (A100→H100→H200→Blackwell) with 2.5-11x performance leaps undermines 6-year economic viability claims
- CoreWeave data shows A100s fully booked after 4+ years and H100s rebooked at 95%, but under peak AI boom conditions with 14-16% margins
- Microsoft CEO Nadella openly avoids “getting stuck with depreciation on one generation” despite Microsoft using 6-year schedules, signaling internal doubt
- H100 rental rates crashed 64-88% ($8 to $2.85-3.50/hour) in six months, while annual depreciation on 2025 builds could hit $40B against $15-20B revenue
- Burry warns of billions in understated expenses; market will test whether GPUs retain economic value (rental rates, utilization) for 6 years or just technical capability
