MTS

Wiki/Economics

GPU Depreciation: Booked Life vs. Economic Life

Thesis. Hyperscalers depreciate AI servers on a straight line over five to six years. The technology cycle that determines economic life — Hopper (2022) → Hopper refresh / H200 (2023-24) → Blackwell B200/GB200 (2024-25) → Blackwell Ultra / B300 (2025-26) → Rubin (2026-27) — runs at roughly half that cadence. If the economic useful life of an AI GPU is two to three years and the booked life is five to six, GAAP depreciation is roughly half of true economic depreciation. That difference flows directly into reported operating income, EPS, and the "AI payback" math that underwrites the $600 billion+ annual hyperscaler capex run-rate. This page collects the primary-source evidence on each link in that chain.

1. Hyperscaler useful-life policy: the timeline

The four U.S. hyperscalers have moved book life in the same direction (longer) for most of the post-2020 period, then begun diverging in 2024-25 as AI-specific obsolescence pressure built. Amazon is the only one to have explicitly shortened a subset of asset lives and to have cited "AI and machine learning" as the reason in plain English.

Date disclosed Company Change Annual P&L impact (per company)
Q4 FY2022 (July 2022) Microsoft Servers & network equipment: 4 → 6 years $3.7B lower depreciation in FY23 ($1.1B in Q1 alone) [The Register / Computer Weekly]
2022 Alphabet Servers: 4 → 6 years Multi-billion benefit (disclosed in 10-K)
2022 Meta Servers: 4 → 4.5 years ~$0.7B
2023 Meta Servers: 4.5 → 5 years ~$2B+
Feb 2024 (Q4 FY23) Amazon Servers: 5 → 6 years ~$900M lower depreciation in 2024 [The Register; Behind the Balance Sheet]
Jan 2023 CoreWeave Technology equipment (GPUs + networking): 5 → 6 years (originally 4) Hundreds of millions; the central plank of the IPO economics [CoreWeave S-1]
Jan 29 2025 (Q4 FY24) Meta "Certain" servers & networking: 5 → 5.5 years (non-AI only) ~$2.9B lower depreciation in 2025 [Yahoo Finance / Meta 10-K]
Feb 6 2025 (Q4 FY24) Amazon Subset of servers & networking: 6 → 5 years (effective 1/1/2025), citing "increased pace of technology development, particularly in AI and machine learning"; plus ~$920M of accelerated depreciation charges in Q4'24 for early-retired equipment ~$700M lower operating income in 2025 [AWS Q4'24 transcript; The Register; Behind the Balance Sheet]

Sources: The Register — Microsoft extends server life · Computer Weekly — $3.3bn savings · Microsoft FY23 10-K · Yahoo Finance — Meta accounting move · Meta 10-K (FY24) · The Register — Amazon Q4'24 · DeepQuarry — Amazon server life revision · Behind the Balance Sheet — Amazon AI reality check

The split. Meta explicitly excluded AI servers from its January 2025 life extension — i.e., management acknowledges that AI hardware does not fit a 5.5-year curve. Amazon went further and shortened lives for the AI-relevant subset and absorbed nearly $1 billion of accelerated depreciation for kit it is retiring early. Microsoft and Alphabet have not yet followed in disclosure, though their 10-K language ("certain" equipment) leaves them room to reverse without restating.

The asymmetry matters: every extension since 2021 has been reported as a "non-cash benefit" that added to EPS. Amazon's 2025 reversal is the first move in the other direction.

2. GPU acquisition cost — what hyperscalers actually pay

NVIDIA Hopper (H100)

  • Street / list: $25,000–$40,000 per H100 SXM5 80GB module. Peak street prices touched $40,000+ during the 2023-24 shortage. IntuitionLabs — NVIDIA AI GPU Pricing; Tom's Hardware
  • DGX H100 (8x H100 system): ~$373,000–$460,000 list. [Cyfuture; Northflank]
  • HGX H100 (8x via OEM): ~$200,000–$300,000 for the GPU baseboard before system integration.
  • H200: ~$30,000–$40,000 per module; ~$315,000 for 8-GPU HGX configurations. [IntuitionLabs]

NVIDIA Blackwell (B200 / GB200)

  • B200 SXM module: $30,000–$40,000 list; OEM quotes for 192GB SXM as high as $45,000–$50,000. NVIDIA's chip-level gross margin on B200 is estimated at ~82% against a ~$6,400 bill of materials. Epoch AI — B200 cost breakdown; Northflank
  • GB200 superchip (1× Grace + 2× B200): $60,000–$70,000.
  • GB200 NVL72 rack (72 Blackwell GPUs + 36 Grace CPUs, fully integrated, liquid-cooled): ~$3 million per rack, 13.5 TB of pooled GPU memory. [Tech-Insider; Modal]

AMD MI300X

  • List / street: $10,000–$15,000. Citi estimated Microsoft pays AMD ~$10,000/unit as the largest MI300X buyer. Other enterprise customers pay ~$15,000. The MI300X is therefore priced at roughly 1/3 to 1/4 of an H100 for nominally competitive memory bandwidth and capacity. Tom's Hardware

Hyperscaler discounts

Bulk hyperscaler pricing is not publicly disclosed but is universally believed to run 20–40% below street, consistent with NVIDIA's reported >70% gross margins on data-center products. CoreWeave's S-1 implies an effective fully-loaded H100 cost (GPU + chassis + interconnect + initial software) of roughly $40k–$45k per GPU at the system level.

3. Rental-rate trajectory: the H100 collapse

The on-demand price of an H100-hour is the single cleanest market signal for GPU economic depreciation, because it is set by competitive supply against contestable demand.

Period H100 on-demand $/GPU-hr Notes
Q2-Q3 2023 $4.70–$8.00 Pre-launch capacity sold out. Some spot listings >$10 [SemiAnalysis]
Q4 2023 $8.00–$11.00 Peak; AWS p5.48xlarge at $98/hr list ($12/GPU-hr)
Q1 2024 $5.00–$7.00 Lambda, CoreWeave reserved rates settle
Q2 2024 $4.00–$5.50 New neocloud capacity (Crusoe, Nebius, Together) online
Q4 2024 $3.00–$4.00 Blackwell ship date approaching
Q1 2025 $2.50–$3.50 Lambda 8×H100 at $2.99/GPU-hr; broad neocloud convergence
Q3 2025 $2.13–$2.85 Cycle low on the spot/short-term market [Silicon Data; IntuitionLabs]
Oct 2025 $1.70 (1-year contract low) SemiAnalysis H100 Rental Index trough
Mar 2026 $2.35 (1-year contract) ~40% recovery — driven by Blackwell supply tightness and inference demand [SemiAnalysis]

Sources: SemiAnalysis — H100 Rental Price Index · Silicon Data — H100 Rental Price Over Time · IntuitionLabs Nov 2025 H100 cloud comparison · Introl — GPU Cloud Price Collapse Dec 2025

What this means for booked depreciation. A six-year straight-line schedule implies an H100 must generate roughly $40k / (6 × 8760h × 0.85 utilization) ≈ $0.90/GPU-hr of gross margin (not revenue) every hour, every year, for six years, to recover cost. At a spot rate of $2.13/hr against a fully-loaded cost (power, cooling, facility, networking, software, sales) of $1.50–$1.80/hr, the H100 spent much of late 2025 generating sub-cost gross margin on the merchant market. Long-term contracts rescue the math — but only because customers (often VC-funded labs) are willing to lock in 2-3 year commitments at $2.20-$2.40/hr. Dylan Patel / Dwarkesh interview

4. Resale and the "depreciated GPU pool"

For accounting useful life to be defensible, an asset must retain economic value across that life. Used-GPU prices are the test.

GPU Vintage New (peak) Used median 2024 Used median 2025 Implied annual decline
V100 (32GB) 2018 ~$10,000 $1,500–$2,500 $1,000–$1,800 ~25%/yr; Azure NCv3 retiring Sept 2025 (7.5-yr life achieved via export-restricted markets and inference repurposing)
A100 (80GB) 2020-21 ~$15,000 $7,000–$10,000 $5,500–$8,000 ~20%/yr; supported by China export-control arbitrage
H100 (SXM) 2023 ~$35,000 61% of contemporaneous new ($21k) 69% of new ($22-24k) Slower-than-expected — see below
H100 server (8-GPU) 2023 $300k+ $150k–$180k $150k–$180k ~40-50% off new at the system level

Sources: Silicon Data — Illusion of Stability · Introl — Secondary GPU Markets Guide 2025 · Applied Conjectures — Hyperscaler depreciation · Hashrate Index — Used GPU pricing

The H100 anomaly. A naive cycle reading predicted H100 residual values would collapse when B200 shipped in volume in mid-2025. Two things prevented that:

  1. Blackwell supply slippage and liquid-cooling integration problems kept B200/GB200 shipments below plan through 2H 2025.
  2. Inference demand for open-weight models (Llama-class, DeepSeek-class) absorbed H100 capacity that frontier labs were vacating.

CoreWeave reported its 2022-vintage H100 contracts re-booked at ~95% of original pricing on expiration — a striking data point that the bull side of this debate leans on heavily. SiliconANGLE — Resetting GPU Depreciation

But — and this is the bear counter — the H100 cycle is unusually generous because of the singular convergence of (a) inference-demand explosion, (b) Blackwell supply problems, and (c) the China export-control overhang sustaining demand for sanctioned chips. None of those is a structural property of the technology cycle. Rubin (2026-27) and beyond may not get the same reprieve.

5. Power, thermal, and facility stranding

Chip-level obsolescence is half the story. The other half is that the building the chip lives in may also become uneconomic. Rack power density tells the story:

Generation Rack power density
Pre-2020 enterprise 6–10 kW per rack
H100 HGX (8-GPU air-cooled) 30–40 kW per rack
H200 / B100 transitional 40–60 kW per rack
GB200 NVL72 (liquid-cooled) 120–140 kW per rack [NVIDIA, Schneider, Tone Cooling]
GB300 NVL72 (Blackwell Ultra) ~135–150 kW per rack
Rubin (announced 2025) targeting >200 kW per rack

Sources: NVIDIA GB200 NVL72 page · Tone Cooling — GB200 NVL72 requirements · SemiAnalysis — GB200 hardware architecture · IntuitionLabs — HGX physical requirements

The retrofit problem. A data center designed in 2021-22 for 30-kW racks and air cooling cannot host GB200 without substantial rebuild: liquid loops, CDUs, busway upgrades, structural reinforcement for liquid-filled racks (≈3,000 lb each), and frequently new substations to handle 3-5x power per square foot. Operator commentary across DCD, DataCenterDynamics, and Schneider Electric briefings suggests retrofits run $8–$15M per MW versus $10–$14M per MW for greenfield purpose-built AI capacity — so retrofit is often not economic, and the older shell is left holding GPUs that themselves are aging. The capex you booked over 15-20 years of facility depreciation is mismatched against IT equipment that turns over every 2-3.

6. The Hopper-to-Blackwell handoff: what actually happened

The cleanest test of the obsolescence thesis was the H100→B200 transition. In 2023-2024, sell-side consensus expected that Blackwell volume shipments (which began ramping Q4 2024) would crater H100 demand and prices. The actual outcome was more nuanced:

  • H100 on-demand spot prices did collapse — from $8+ to $2.13 trough — consistent with the obsolescence thesis on the marginal hour.

  • H100 reserved 1-year contracts declined more slowly and then rebounded 40% from Oct 2025 to Mar 2026. [SemiAnalysis] This rebound reflects three factors none of which contradict the obsolescence concern:

    1. Blackwell supply was constrained through 2025 (liquid cooling, NVLink switch yield).
    2. Inference workloads absorbed the H100 fleet.
    3. CoreWeave / Crusoe / Lambda's customer base of VC-funded model labs needed any GPU capacity, not specifically Blackwell.

  • Contracted-but-undelivered H100 capacity. Several reports (including SemiAnalysis and Bloomberg) note that some hyperscale and neocloud H100 orders placed in 2024 are arriving in 2025-26 into a market where the per-FLOP cost of new Blackwell is ~40-50% lower. Those H100s start their 6-year book life already economically stale.

CNBC — How long before a GPU depreciates? · SiliconANGLE — Resetting GPU Depreciation

7. Analyst and skeptic positions

Jim Chanos (Kynikos / now Chanos & Co)

Chanos is short Oracle, CoreWeave, and the data-center landlord complex (Equinix has appeared in his commentary). His central claim: using CoreWeave management's own stated 2-3 year technical useful life for GPUs, against the 6-year book life, the company generates ~0% ROIC. He has also raised the "circular financing" critique — that hyperscaler AI demand is being sold to VC-funded labs that don't generate positive unit economics, creating recycled-capital loop. Monetary Matters podcast with Jack Farley · Yahoo Finance — depreciation time bomb · TipRanks — GPU-backed loans

Michael Burry (Scion)

Disclosed short positions in NVIDIA and Palantir in late 2025; subsequently published written arguments that hyperscalers will understate depreciation by ~$176B cumulatively over 2026-2028. His specific estimates: Oracle's earnings overstated by ~27% by 2028 if useful life is corrected to 3 years; Meta's by ~21%. CNBC · Sherwood News · Motley Fool

Sell-side

  • Barclays has trimmed 2025-26 EPS estimates for AI-exposed names by up to 10% to reflect more aggressive depreciation assumptions. Barclays note via Stanley Laman summary
  • Morgan Stanley projects $3T of cumulative data-center spend through 2029 with a $1.5T financing gap; raised 2027/2028 Google TPU shipment estimates to 5M/7M units, implicitly assuming faster Hopper/Blackwell displacement at Google.
  • Goldman Sachs — its "$7.6T AI infrastructure ledger" framework explicitly names chip useful life as "the single most influential variable" determining cumulative AI capex. Goldman Sachs — Tracking Trillions
  • JPMorgan Asset Management Eye on the Market 2026 ("Smothering Heights") raises the depreciation question for hyperscaler free cash flow conversion.

Princeton CITP (academic)

Direct technical assessment by network and systems researchers: "1 to 3 year useful life reflects engineering reality" for frontier AI GPUs, "driven by NVIDIA's annual release cadence." CITP — Lifespan of AI Chips: the $300B Question

Ed Zitron (Where's Your Ed At?)

The most acerbic of the skeptics, but data-grounded. Key points: (a) a 100MW AI data center costs ~$4.4B fully loaded and generates ~$1.06B in annual revenue, which makes the 6-year depreciation curve barely workable even before considering technical obsolescence; (b) inference margins on subscription products like ChatGPT and GitHub Copilot are negative when fully costed; (c) the "subscription as subsidy" model conceals true inference cost. AI's Economics Don't Make Sense · The AI Compute Demand Story Is A Lie

SemiAnalysis (Dylan Patel)

Less polemical, more data-rich. Patel's published view: H100s have a longer-than-expected useful life for inference but a short one for training, and the rental-rate recovery in late 2025 is supply-driven (Blackwell shortfall) not demand-driven. He notes contract H100 prints as high as $2.40/hr for 2-3 year terms in late 2025, which is consistent with positive unit economics on long-duration contracts but not on the merchant spot market. SemiAnalysis — Great GPU Shortage · GPU Cloud ClusterMAX

8. Putting it together: the EPS arithmetic

A simplified back-of-envelope, using Burry's framework, on the four hyperscalers:

  • Combined 2025 AI-related capex: ~$300B (Microsoft $80B + Meta $65B + Google $75B + Amazon $80B, AI-allocated subset).
  • If true economic life is 3 years rather than booked 5.5–6 years, annual depreciation expense rises by ~$50–60B in aggregate (~80–100% increase on the affected asset base).
  • That flows ~1:1 to operating income, implying a ~10-25% hit to non-Amazon hyperscaler operating profit if marked to economic reality.
  • The "2-year AI payback" headline metric (cited by Microsoft, Meta, Google in various forms during 2024) implicitly assumed continued 5-6 year book life; under 3 years it stretches to roughly 5+ years before the GPU has paid for its cost of capital — and that is before the facility refresh problem in §5.

This is the core of why GPU depreciation has become the central accounting controversy of the AI capex cycle. None of the disputed numbers are themselves hidden — they sit in plain English in 10-Ks. What is disputed is which side of the policy choice (6 years vs. 3 years) reflects economic reality. Amazon's January 2025 reversal, Meta's carve-out of AI servers from its life extension, and CoreWeave's stock action after Q3 2025 all suggest the market is pricing in some non-trivial probability that the 6-year camp is wrong.

Key sources (consolidated)

10-K / earnings primary documents

Trade press

Industry analysis

Skeptic commentary