GPU Depreciation and Useful Life: Burry's Case vs. the 5-6 Year Books

Current server and networking equipment useful-life assumptions (as of late 2026 filings) cluster around 5–6 years for the major hyperscalers, with GPUs typically subsumed under servers/network equipment rather than separately disclosed. Microsoft uses a 2–6 year range; Alphabet and Oracle use 6 years; Meta uses 5–5.5 years (most at 5.5); Amazon uses 5–6 years overall (with a subset at 5). These are accounting estimates for straight-line depreciation of property and equipment (servers, networking gear, and related data-center hardware), not physical obsolescence or resale values. Changes are prospective and often quantified for their one-time earnings impact in the year of adoption.[1][2]

Most changes since ~2020 extended lives (reducing near-term depreciation and boosting reported earnings/profits), reflecting longer actual utilization due to software optimizations, diversified workloads, and slower hardware turnover than initially assumed. Amazon’s 2025 partial reversal is the notable exception, attributed to faster AI/ML technology cycles. Data-center buildings and improvements generally carry much longer lives (often 15–40+ years) and were sometimes extended separately.[3][1]

Amazon (AMZN) – Multiple Adjustments, Including 2025 Shortening on AI Hardware Pace

Amazon has adjusted server lives most frequently among the group, moving from shorter periods pre-2020 to a 5–6 year range, then partially reversing one cohort in 2025 due to accelerated AI-driven obsolescence. The 2025 change applied only to a subset of servers and networking equipment already on the books or acquired that year.[2][4]

• Pre-2020: Servers ~3 years.
• Jan 1, 2020: Servers 3 → 4 years (earnings effect quantified in later disclosures as part of cumulative benefits).
• Jan 1, 2022: Servers 4 → 5 years; networking equipment 5 → 6 years. 2022 effect: depreciation reduced by ~$3.6B; net income benefit ~$2.8B ($0.28 per share).[4]
• Jan 1, 2024: Servers 5 → 6 years. 2024 effect: depreciation reduced by $3.2B; net income benefit $2.5B ($0.23 per share), primarily AWS.[2][4]
• Jan 1, 2025: Subset of servers and networking 6 → 5 years (due to faster AI/ML technology development). 2025 effect: depreciation increased by $1.4B; net income reduced by $1.0B ($0.10 per share), primarily AWS. Quarterly/9-month impacts also disclosed in 10-Qs.[3][4]
• Current (FY2025 10-K, filed ~Feb 2026): Servers and networking equipment: 5–6 years.[2]

Implication: The 2025 shortening highlights that even extended lives are not fixed; AI hardware turnover can compress assumptions for specific cohorts, accelerating expense recognition and pressuring margins in high-capex segments like AWS.

Meta (META) – Extension to 5.5 Years in 2025 Delivered Largest Relative Earnings Benefit

Meta extended most servers and network assets to 5.5 years effective 2025 after prior stepwise increases, producing a substantial one-time depreciation reduction relative to its earnings base.[5][6]

• ~2022 (Q2/Q4 steps): 4 → 5 years (cumulative effect ~$860M less depreciation in 2022).
• Prior to 2025: 4–5 years (policy range).
• Jan 1, 2025: Most servers and network assets to 5.5 years. 2025 effect (based on assets in service as of Dec 31, 2024): depreciation reduced by $2.92B; net income increased by $2.59B ($1.00 per diluted share).[5][6]
• Current (2025 10-K, filed ~Jan 2026): Servers and network assets: 5–5.5 years (most at 5.5).[1]

Implication: The $1.00/share benefit underscores how sensitive reported profitability is to these estimates amid surging AI capex; actual depreciation still rose year-over-year due to fleet growth.

Alphabet (GOOGL) – Stabilized at 6 Years Since 2023 Extension

Alphabet moved to a uniform 6-year life for servers and network equipment in 2023 and has held steady, with no further disclosed changes through 2025.[1][7]

• ~2021: Servers 3 → 4 years (effect ~$2.6B less dep in 2021).
• Jan 2023: Servers 4 → 6 years; certain network equipment 5 → 6 years. 2023 effect: depreciation reduced by $3.9B; net income increased by $3.0B ($0.24 per share).[1][7]
• Current (2025 10-K, filed ~Feb 2026): Servers and network equipment: generally 6 years. (Data centers/buildings: 7–40 years.)[1]

Implication: The 6-year assumption is now a consensus benchmark for several peers; Alphabet’s clear capex split (e.g., ~65% servers/networking in recent quarters) allows investors to model depreciation trajectories more precisely.

Microsoft (MSFT) – Broad 2–6 Year Range; Major Extension in FY2023

Microsoft uses a range rather than a point estimate and extended from 4 years in FY2023; servers/network equipment have remained in the 2–6 year band since, while buildings were separately extended prospectively.[8]

• FY2023 (effective July 2022): Servers and network equipment 4 → 6 years (due to software efficiencies and technology advances). FY2023 effect: operating income increased by $3.7B; net income by ~$3.0B ($0.40 per share).[1][8]
• Current (FY2026 10-K, filed ~July 2026): Servers and network equipment: 2–6 years. (Buildings/improvements: 5–15 years currently; extended to 25 years effective FY2027 start.)[3][8]

Implication: The range provides flexibility for varying asset cohorts (e.g., potentially shorter for some GPUs or newer hardware); the separate building extension affects lease accounting and long-term depreciation but not the core server/GPU expense profile.

Oracle (ORCL) – Extended to 6 Years in FY2025; Consistent with Peers

Oracle aligned with the 6-year consensus via stepwise extensions.[1]

• FY2023 (~June 2022 effective): 4 → 5 years (effect: ~$434M lower operating expenses).
• FY2025 (~June 2024 effective): 5 → 6 years. FY2025 effect: operating expenses reduced by $733M; net income benefit ~$573M ($0.20–0.21 per share).[1]
• Current (FY2026 10-K, filed ~June 2026): Servers and networking equipment: 6 years (computer/network/machinery equipment generally 1–6 years).[9]

Implication: Oracle’s smaller scale produces proportionally smaller dollar impacts, but the pattern mirrors the broader industry shift toward longer assumed lives to absorb AI-driven capex growth.

Summary Table of Key Changes and Current Assumptions

Useful life (servers/network equipment) by company and notable years, with disclosed earnings effects (primarily from the change year; effects are pre-tax where specified and often primarily impact cloud/AI segments). Current as of most recent 10-Ks referenced. Effects are one-time prospective impacts based on assets at the time of change.[1][1]

• Amazon: 2022 (4/5→5/6): –$3.6B dep / +$2.8B NI; 2024 (5→6): –$3.2B dep / +$2.5B NI; 2025 (subset 6→5): +$1.4B dep / –$1.0B NI. Current: 5–6 yrs (subset 5).
• Meta: 2022 (4→5 steps): ~–$0.86B dep; 2025 (to 5.5): –$2.92B dep / +$2.59B NI. Current: 5–5.5 yrs (most 5.5).
• Alphabet: 2023 (4→6 / 5→6): –$3.9B dep / +$3.0B NI. Current: 6 yrs.
• Microsoft: FY2023 (4→6): +$3.7B op. inc. / +$3.0B NI. Current: 2–6 yrs.
• Oracle: FY2023 (4→5): ~–$0.43B opex; FY2025 (5→6): –$0.73B opex / +~$0.57B NI. Current: 6 yrs.

Filing sources (examples; full 10-Ks/10-Qs on EDGAR/SEC.gov): Amazon FY2025 10-K (property/equipment note and change disclosures); Meta 2025 10-K (change in estimate note); Alphabet 2025 10-K; Microsoft FY2026 10-K; Oracle FY2026 10-K. Secondary analyses aggregate and quote these directly.[2][9]

These estimates are management judgments revisited periodically; actual hardware replacement cycles, resale values, and AI workload demands can differ. For competitors or new entrants, matching or exceeding these assumptions requires demonstrating equivalent utilization longevity through data, software, or hardware advantages—otherwise, higher relative depreciation could pressure reported margins versus peers.

---

Recent Findings Supplement (November 2026)

No new changes to server or GPU useful-life assumptions have been disclosed in 10-Ks or 10-Qs filed after May 5, 2026, for Microsoft, Alphabet, Amazon, Meta, or Oracle.[1][2][3][4]

Recent filings (Microsoft FY2026 10-K filed July 29, 2026; Q1 FY2027 10-Q filed Oct 28, 2026; Alphabet Q3 2026 10-Q filed Oct 29, 2026; Amazon Q2 2026 10-Q filed July 31, 2026; Meta Q3 2026 10-Q filed Oct 22, 2026; Oracle Q1 FY2027 10-Q filed Sept 11, 2026) describe ongoing reviews of useful lives or restate ranges/ranges without reporting new estimate changes, earnings impacts, or revisions effective in 2026.[5][6][7]

Known Prior Changes (Disclosed in Earlier Filings, for Context)

These pre-date the post-May 5, 2026 cutoff and are not "new data" per the research criteria, but they form the baseline referenced in secondary summaries of latest reports:

• Amazon: Effective Jan. 1, 2024, servers changed from 5 to 6 years (benefit disclosed in 2025 10-K). Effective Jan. 1, 2025, a subset of servers/networking equipment shortened from 6 to 5 years due to faster AI/tech obsolescence; this increased 2025 depreciation by ~$1.4B and reduced net income by ~$1.0B (disclosed in 2025 10-K filed early 2026; policy range remains 5–6 years).[8][9]
• Meta: Effective Jan. 1, 2025, most servers/network assets extended to 5.5 years (from prior range); reduced 2025 depreciation by $2.92B and increased net income by $2.59B (~$1.00 diluted EPS) (disclosed in 2025 10-K filed Jan. 2026; policy now 5–5.5 years).[10][11]
• Microsoft: Servers/network equipment disclosed as 2–6 years in FY2026 10-K (no new change noted). Prior extension (4 to 6 years, FY2023) had a large positive earnings effect.[5]
• Alphabet: Servers/network equipment generally 6 years (per 2025 10-K and later summaries; prior extension to 6 years in 2023).[12]
• Oracle: Servers/networking equipment 6 years (per FY2026 10-K filed June 2026; prior extension from 5 to 6 years effective FY2025 had ~$733M lower operating expenses).[13]

Implications for competitors/analysts: Without fresh revisions in mid-2026 filings, the 5–6 year range (with Meta at the shorter end for most assets and Amazon’s subset at 5) remains the prevailing assumption amid AI-driven hardware turnover. Any future shortening (as Amazon did for a subset) would accelerate depreciation and pressure earnings more than extensions. Data-center/building lives (e.g., Microsoft commentary on potential 25-year extensions) are separate from servers/GPUs and not covered in the requested server/GPU focus.[14]

No regulatory updates, new publications, or other post-May 2026 developments altering these disclosures were identified in the searched filings. Secondary analyses as of Oct/Nov 2026 confirm the absence of additional changes.

CoreWeave depreciates its GPU servers (technology equipment) over a straight-line 6-year useful life per its 2025 Form 10-K, after extending from 5 years effective January 2023; this choice reduces near-term depreciation expense (e.g., ~$20M lower in 2023) and supports stronger reported margins/EBITDA while aligning with multi-year customer contracts that often feature take-or-pay commitments and prepayments.[1][2]

This policy assumes GPUs retain economic value long enough for full cost recovery through rentals, including potential second-life use (e.g., CoreWeave disclosed an A100 contract extending into 2029 on hardware introduced in 2020). It matches the schedules of major hyperscalers like Alphabet, Microsoft, and Oracle (generally 6 years for servers/network equipment) but exceeds more conservative neocloud approaches.[2]

• CoreWeave’s S-1/10-K filings and earnings commentary tie the schedule to infrastructure investments funded partly by customer contracts (typical terms around 3–5 years, with examples of attractive pricing on older SKUs).[3]
• The extension lowered 2023 expenses and was presented as reflecting extended usability amid rapid fleet growth (majority Hopper GPUs at the time of IPO filing).[4]
• Analyst and media commentary (e.g., FT) flags the 6-year horizon as aggressive in an AI context where architectures advance quickly, raising questions about resale value vs. book life, though CoreWeave points to ongoing demand and contract renewals.[4][5]

For someone entering or competing in the neocloud space, CoreWeave’s choice illustrates how a longer schedule can improve near-term financial optics and support debt financing (by showing slower expense recognition), but it heightens scrutiny on utilization, renewal pricing, and residual values—especially if customer concentration or tech shifts accelerate obsolescence.[6]

Nebius (formerly tied to Yandex) uses a 5-year straight-line schedule for servers and network equipment (per its 2025 Form 20-F), after extending from 4 years effective January 2026 to reflect “usage patterns and current utilization commitments”; this change reduced 2026 depreciation by roughly $167.6 million.[1][2] IREN (Iris Energy) applies 5 years to HPC hardware/GPUs per its fiscal 2026 10-K.[1]

These shorter schedules accelerate expense recognition relative to CoreWeave or hyperscalers, producing more conservative margins but potentially better insulating against future write-downs in a fast-evolving GPU market. Nebius’s extension was explicitly linked to observed utilization and commitments, suggesting alignment with contracted demand.[2]

• Lambda Labs is reported at 5 years in multiple analyses of neocloud policies.[7][7]
• Crusoe (private, with recent large funding rounds) lacks public filing detail on exact schedules; its CEO has publicly emphasized that GPUs can remain profitable for 6+ years (or 7–8 years per some Lambda commentary) across diverse workloads, supporting redeployment of older hardware rather than rapid obsolescence.[8][8]
• Similar neoclouds or GPU-focused filers (e.g., WhiteFiber at 5 years, DigitalOcean at 6 years for servers including GPU instances) cluster around 5–6 years.[1]

Nebius’s adjustment shows how utilization data from contracts can justify (or require) changes to estimates; competitors using shorter lives may face higher reported costs but gain credibility with auditors/investors wary of aggressive assumptions.[9]

Hyperscalers (Alphabet, Microsoft, Oracle at 6 years; Meta at 5.5 years; Amazon generally 5–6 years with a subset shortened to 5 years in 2025 citing accelerated AI tech development) largely converged on 5–6 year server/network lives after multiple extensions since 2020 (16 of 17 changes lengthened lives).[1][2]

Neocloud GPU-specific or AI-focused schedules trend slightly shorter on average (median ~5 years where explicitly named) than broad hyperscaler categories, though CoreWeave matches the longer end. Amazon’s 2025 subset shortening explicitly referenced faster AI/ML hardware pace and increased 2025 depreciation by $1.4 billion.[10]

• Most changes since 2020 extended lives, lowering depreciation and boosting earnings (e.g., Meta’s 2025 move to 5.5 years cut ~$2.92 billion in depreciation).[1]
• Where filings distinguish GPUs or AI hardware, lives are often at the shorter end of the range.[1]

The convergence suggests industry-wide confidence in multi-year GPU utility (including inference/lower-priority workloads), but neoclouds’ more AI-centric fleets may warrant (or invite) more conservative estimates than diversified hyperscalers.[11]

Depreciation schedules interact directly with contract structures: neoclouds commonly use 3–5 year take-or-pay agreements with significant customer prepayments (e.g., IREN reports 45–55% of GPU capex covered on recent 3-year deals; Nebius notes 50–60% in some Q2 2026 deals) that de-risk financing and support payback periods as short as ~2 years on compute investment.[12][13]

CoreWeave references typical ~5-year contracts that repay asset-level debt and leave residual capacity for renewals or secondary use; an A100 example running into 2029 illustrates potential beyond initial terms.[3] Nebius’s 2026 life extension cited utilization commitments, directly tying accounting to contracted demand.[2]

• Prepayments and creditworthy offtakers (e.g., Microsoft deals enabling ~6% financing rates) allow lower borrowing costs, with lenders pricing risk based more on the customer than the GPU itself.[14]
• Shorter contracts or weaker counterparties correlate with higher financing rates (e.g., double-digit yields on unsecured or lower-rated GPU debt).[6]

Longer depreciation relative to contract length assumes successful renewals or redeployment; mismatches (e.g., 5-year debt vs. 3-year contracts) require confidence in ongoing demand and pricing power.[6]

No specific SEC comment letters or auditor commentaries on these policies appear in public filings or reports reviewed; changes are treated as accounting estimates under standard GAAP/IFRS rules, with impacts disclosed (e.g., Nebius’s $167M benefit).[2]

Analyst and investor critiques focus on potential overstatement of profits if economic lives are shorter than assumed (e.g., Michael Burry’s 2025 commentary on industry-wide understated depreciation of ~$176 billion across 2026–2028, flagging hyperscalers and implying risks for neoclouds; FT and Seeking Alpha pieces questioning 6-year horizons amid rapid Nvidia generational shifts).[7][4][5]

• Book life is explicitly distinguished from resale/liquidation value in analyses; lenders are advised to use market residual data rather than accounting schedules.[1]
• Defenses emphasize workload diversity (training → inference → other AI tasks) and observed utilization keeping older GPUs (e.g., H100s) profitable and priced stably.[8]

Critiques highlight that aggressive schedules can mask true economics in a high-capex, high-depreciation business; entrants should model sensitivities around 4–6 year ranges and stress-test against resale data or impairment scenarios.[15]

Overall, neocloud GPU depreciation (clustered at 5–6 years) largely tracks hyperscalers but with more variation tied to AI focus and contract visibility; the choice materially affects reported profitability, debt capacity, and perceived risk, with longer schedules providing accounting support for growth but inviting skepticism on sustainability.[11]

Companies like CoreWeave benefit from alignment with hyperscaler norms and strong backlog visibility, while shorter schedules at Nebius/IREN/Lambda offer conservatism. No major regulatory pushback has surfaced, but ongoing monitoring of utilization, renewals, and impairments will determine if estimates hold. New entrants should prioritize contracted demand and secondary-market data over pure accounting lives when sizing capital or competing on unit economics.

---

Recent Findings Supplement (November 2026)

CoreWeave maintains a 6-year straight-line useful life for technology equipment (including GPU servers), extended from 5 years effective January 1, 2023.[1][2] This non-cash change reduced 2023 expenses by about $20 million. In its Q2 2026 earnings (reported August 2026), the company highlighted an A100 contract (chips introduced in 2020) extending into 2029 at attractive pricing, framing this as evidence of multi-cycle commercial value beyond initial 5-year contracts.[3][3] Typical customer contracts average around 5 years (or 3 years in some debt contexts), with some facilities showing 5-year debt maturities against shorter contracts; management presents this as lender confidence in ongoing demand.[4][4]

• Recent quarterly depreciation and amortization reached $1.393 billion (54% of revenue), up sharply year-over-year.[5]
• Debt agreements reference a 6-year GPU Depreciated Amount with no explicit salvage/residual value terms noted in analyses of filings.[6]
• Analysts (e.g., Goldman Sachs in October 2026) model scenarios using 4.5–6-year GPU lives, estimating significant additional debt needs even under longer assumptions to fulfill backlog.[7]

This policy aligns depreciation closer to some hyperscalers (Microsoft, Alphabet, Oracle at 6 years) while providing a buffer versus shorter neocloud assumptions, but it hinges on secondary contracts and renewals materializing at viable pricing to cover debt service after initial terms.[8]

Nebius extended its useful life estimate for servers and network equipment from 4 to 5 years effective January 2026.[8][2] This accounting change is projected to reduce 2026 depreciation expense by approximately $167.6 million.[2] Filings as of mid-2026 confirm the 5-year policy (shorter than CoreWeave’s 6 years).[8]

• One October 2026 analysis contrasted Nebius trading at ~9x replacement cost of its H100-equivalent fleet versus CoreWeave near 1.2x, tying this partly to differing depreciation and capex perceptions amid rapid generational price declines (replacement costs halving roughly every 49 months per merchant data).[9]
• No auditor or SEC comment letter details emerged on this change.

The extension brings Nebius in line with IREN and some Amazon subsets while still shorter than CoreWeave or Meta’s 5.5 years, potentially improving near-term reported margins but leaving less cushion if actual economic lives or resale values prove shorter.[8]

IREN applies a 5-year useful life to HPC hardware/GPUs per its fiscal 2026 10-K.[8][8] No policy change was reported after May 2026. The company emphasizes customer prepayments (45–55% of GPU capex on recent 3-year contracts) and financing structures that cover most or all of GPU costs upfront, with payback periods around 2 years on deals generating >$20 million revenue per MW.[10][11]

• Focus in 2026 earnings materials and calls remains on contract ramp (e.g., Microsoft deal) and decommissioning mining operations by end-2026 rather than depreciation revisions.[12]
• Depreciation rose with higher capacity and GPU deployments, contributing to net losses alongside impairments.[13]

Shorter schedules versus CoreWeave accelerate expense recognition but pair with structures that front-load customer funding, reducing reliance on long post-depreciation tails.

Hyperscalers show mixed but generally longer or similar schedules: Microsoft/Alphabet/Oracle at 6 years, Meta at 5.5 years, with Amazon varying (some subsets at 5 years after prior extensions or shortenings).[8][8][2] Across 21 filers reviewed in October 2026 data, 16 of 17 useful-life changes since 2020 lengthened estimates.[2] Neocloud contract lengths (often 3–5 years) frequently align with or fall inside depreciation periods, making renewals, secondary use (e.g., inference after training), or residual value critical. Analyst critiques highlight that book lives do not equal resale or economic value—collateral can face haircuts from generational price drops (e.g., H100-era financing versus newer replacement costs), with some noting GPU rents recently exceeding 6-year schedule assumptions amid post-April 2026 demand.[14][9]

• No specific SEC comment letters or auditor reports on these policies appeared in recent searches.
• Crusoe (private) CEO comments in October 2026 emphasized that old GPUs may retain value longer than assumed via workload shifting, but provided no numerical schedule.[15][16] Lambda had no new public policy details.

For entrants or competitors, these policies underscore the need to model cash economics separately from accounting (e.g., via contract-backed financing and real-time rental/resale data) rather than relying solely on 5–6 year straight-line assumptions. Mismatches between depreciation, debt maturities, and actual utilization or obsolescence remain a focal point of scrutiny, with recent renewal evidence and life extensions providing some counter to bear cases but not eliminating collateral or funding gap concerns in a fast-evolving hardware cycle. Data is drawn from 2026 filings, earnings, and analyses published after May 2026; private firms like Crusoe and Lambda offer less granular public detail.

H100 rental prices followed a classic scarcity-to-surplus-to-rebound pattern, with spot/contract indices declining sharply from 2023 peaks before rising again in 2026 due to sustained inference and training demand.[1][2]

SemiAnalysis's H100 spot/composite index started at ~$6.57–6.62/hr in 2H 2023 (peak scarcity), fell steadily through 2025 lows around $2.79–2.91/hr (e.g., Oct 2025 spot ~$2.84/hr, 1y contracts bottoming near $1.45–2.00/hr), then stabilized or rebounded modestly to ~$2.80–2.92/hr spot and $2.00–2.70/hr for 1y contracts by early-mid 2026 amid sold-out on-demand capacity and token demand growth. Ornn, Silicon Data, and Compute Census medians in late 2026 clustered around $2.96–$3.53/hr (on-demand often $3.49 median, with floors ~$1.30–1.80 on marketplaces like Vast.ai). Hyperscaler list prices remained far higher (e.g., AWS ~$6.88/hr).[3][4]

• Marketplace/neo-cloud offers (Vast.ai, RunPod, Lambda) consistently undercut hyperscalers by 50%+ but show wide dispersion (H100 ranges often 5–18x across providers).[5]
• 1y contract indices (SemiAnalysis) rose ~40% from Oct 2025 lows (~$1.70) to ~$2.35 by Mar 2026.[6]
• Rebound tied to Blackwell supply constraints, power limits on new capacity, and inference/agent workloads favoring proven Hopper silicon.[4]

For new entrants or operators competing on H100 capacity: Secure long-term power and colo deals early; price volatility rewards those who can hold inventory through dips or pivot to spot/reserved mixes. Marketplace operators face thin margins unless utilization exceeds ~60–70% after power/hosting fees.

A100 and older generations (V100) exhibit steeper absolute price decay but retain viability for inference workloads, with A100 medians now ~$1.5–1.75/hr (floors ~$0.40–0.53 on marketplaces) versus V100 at ~$0.12–0.40/hr.[3][7]

CCIR analysis shows rental rates declining ~21% per year of chip age across generations (cross-sectional fit on 2026 data), with V100 (2017-era) recovering MSRP in ~10 months gross but now renting at fractions of newer cards; A100 (2020) crossed net recovery around 39 months and continues earning at ~0.95× MSRP-equivalent by age ~6 years.[8] SemiAnalysis and other indices show A100 softening more than H100 in 2025 corrections (broader secondary-market liquidity) but stabilizing or slightly rising in 2026 on inference demand. V100 prices have remained low and relatively stable for years, suitable only for lighter or legacy workloads.[9]

• Current snapshots (late 2026): A100 on-demand medians ~$1.75 (range $0.53–5.07); V100 ~$0.14–1.60 (floors often <$0.20 on Vast/Verda).[3][10]
• Per-bandwidth pricing narrows the gap: older cards trade at competitive $/TB/s-hr for memory-bound inference despite lower peak FLOPS.[8]
• A100 contracts extend into 2029 at CoreWeave and elsewhere, confirming sustained demand.[11]

Operators hosting older GPUs can still clear positive margins over variable costs if utilization is high and power is cheap (<$0.10–0.15/kWh effective), as power represents only 8–15% of total hosting costs for these lower-TDP cards.[12]

Approximate loaded power (GPU TDP + system/PUE overhead ~1.3–1.5×): H100 ~1.0–1.5 kW effective (~$0.10–0.30/hr at $0.08–0.20/kWh industrial rates); A100 ~0.5–0.7 kW (~$0.05–0.14/hr); V100 ~0.4 kW (~$0.03–0.08/hr). Even at marketplace floors, A100/V100 rentals exceed marginal power costs at >20–40% utilization; H100 requires higher rates or utilization to cover its higher draw.[13][14]

• Hosting economics favor A100/L40S-class cards for high-utilization inference (e.g., batch/background serving) where per-token costs can beat pricier newer GPUs.[15]
• Self-hosted enterprise utilization often lags (many clusters <50%, some as low as 5–25%), eroding margins despite low rental equivalents.[16]

For margin-focused hosts: Prioritize low-power regions and inference-optimized workloads; older chips extend payback periods but require disciplined utilization tracking. Newer chips command premiums that can offset higher power if scarcity persists.

Utilization evidence shows A100 fleets remaining heavily subscribed for inference (including multi-year contracts), while broader market signals indicate tight capacity overall, though self-hosted and spot reliability introduce caveats.[11]

On-demand H100/A100 capacity was frequently sold out across providers in 2026 per SemiAnalysis surveys; A100s serve as the "inference workhorse" with throughput holding steady while prices fell. CoreWeave and others extended A100 rentals through 2029. Marketplace listings (Vast.ai) often show high offer counts for A100s.[16][15]

Limitations: "Fully booked" often refers to committed/reserved capacity or specific providers; spot/interruptible offerings have lower effective utilization and reliability risks. Enterprise-owned fleets report low average utilization (VentureBeat survey: 83% ≤50%). Geography, interconnect (SXM vs PCIe), and workload fit (memory bandwidth) drive real billed hours far below listed availability.[5]

Entrants should differentiate on reliability SLAs, workload-specific optimization (e.g., vLLM tuning for A100 inference), or bundled software rather than raw price; pure commodity hosting faces utilization and power risks.

Overall market structure favors specialized operators who match chip age to workload (H100/H200 for training/frontier, A100/V100 or consumer for cost-sensitive inference) while hedging via diversified providers and term lengths. Price decay is real but slowed or reversed by AI demand growth; positive margins persist well beyond initial expectations for mid-generation cards. Data from indices like SemiAnalysis, Ornn, Silicon Data, and marketplaces (Vast/Lambda) enable precise benchmarking, but actual economics hinge on utilization discipline and power contracts. Additional primary transaction data or provider-specific utilization reports would further refine margin thresholds.

---

Recent Findings Supplement (November 2026)

Recent GPU rental price data (post-May 2026, primarily October–November 2026 observations) shows continued high demand driving price increases even for older chips, with executed-trade indices like Ornn revealing premiums for newer architectures and sustained viability for A100/V100 at lower rates.[1][2]

Ornn Compute Price Index (Executed Trades Benchmark)

Ornn’s daily settlements (volume-weighted from neocloud transactions, distinct from listings) as of late October to early November 2026 provide the most direct recent executed-price data:[1][2]

• H100 SXM: $3.00–$3.53/GPU-hour (e.g., $3.00 on Oct 31; $3.53 on Nov 3; up ~25–43% month-over-month in recent prints).
• H200: $5.64–$7.31/GPU-hour (premium over H100 reflecting higher memory/bandwidth).
• B200: $9.41–$9.54/GPU-hour (newer Blackwell commanding the highest rates).
• A100 SXM4: $1.04–$1.08/GPU-hour (up ~7–20% since early 2026 or in recent 1M periods; notable for a 2020-era chip).
• RTX PRO 6000: ~$1.52–$1.54/GPU-hour.

These reflect transacted prices rather than advertised rates. A related X post highlighted Ornn-indexed rises (H100 +~36%, A100 +~7% since Jan 1, 2026), underscoring that older silicon retains or gains value amid demand.[3]

Marketplace and Aggregator Listings (Vast.ai, Lambda, RunPod, etc.)

Live on-demand/community/spot rates from providers and aggregators (as of early November 2026 snapshots) show wide spreads by tier and availability:[4][5][6]

• H100: Marketplace lows ~$1.5–$1.8/GPU-hour (e.g., Vast.ai community or Lium); on-demand/secure often $2.7–$4.2; spot/interruptible as low as ~$1.1–$2.6. Medians across providers ~$3.3–$4.2.
• A100 (40/80GB variants): Cheapest ~$0.27–$0.60/GPU-hour on Vast.ai/marketplaces; broader medians ~$1.0–$1.85; spot lower (e.g., ~$0.5–$1.1).
• V100: Marketplace floors ~$0.05–$0.09/GPU-hour (Vast.ai spot/on-demand); higher listings up to ~$0.5+ depending on config/region.
• Other notes: H200 ~$3.5–$5+; newer Blackwell chips (B200/B300) significantly higher. Aggregators like SiliconTicker, Silicon Lien, CCIR, and GPU Finder report similar ranges with daily/weekly refreshes (e.g., H100 family public-list medians ~$4.18).[7][8]

Prices vary by region, commitment (reserved discounts), and reliability (community vs. secure tiers); spot offers deeper discounts but with interruption risk.

Price Decay by Chip Age and Margin Thresholds

Data illustrates decay with age but with recent upward pressure reversing some prior declines: H100 (2022-era) commands 3x+ the executed rate of A100 (2020), while V100 (2017) rents at 5–20x lower.[1][6]

• Older chips retain positive margins over power/opex in many cases. One hosting analysis (Oct 2026) for A100 80GB at ~$1.43 customer rate and L40S at ~$0.92 notes power costs of ~$0.04–$0.18/GPU-hour; breakeven possible below 20–40% utilization depending on electricity rates and overhead, with higher utilization yielding clear profits after fees.[9]
• V100 at $0.05–$0.09 likely clears power easily on efficient hosts but offers lower absolute margins/revenue.

No comprehensive new “chart” of multi-year decay curves appeared in results, but month-over-month and year-to-date rises (even for A100) indicate demand is sustaining older hardware longer than typical depreciation models predicted.

Utilization Evidence and Limitations

Direct utilization metrics remain limited in public reports. Ornn dashboards reference provider-level figures (e.g., one example at 91% utilization alongside pricing tools).[10] Rising executed prices for A100 (up since Jan 2026) and sustained marketplace availability for V100/A100 imply high demand/utilization for older chips, with A100s described in prior contexts as “fully booked” in high-demand periods.[3]

Limitations: Most data is from listings or indices rather than real-time occupancy; spot/interruptible capacity can mask underlying demand; regional/provider differences exist; no broad new regulatory or policy updates noted in searches. Aggregators like Vast.ai report stock history (e.g., limited or high availability percentages), but these are snapshots.

For competitors or entrants: Newer chips (H100+) offer higher absolute revenue but face premium competition and potential future supply ramps; older A100/V100 provide lower entry barriers and positive (if thinner) margins if power/hosting costs are controlled and utilization stays elevated amid ongoing AI demand. Monitor Ornn settlements and Vast.ai-style marketplaces for real-time signals, as prices have shown volatility and upward trends into late 2026. Data evolves daily; cross-reference multiple aggregators for robustness.

Secondary-market prices for used A100, H100, and V100 GPUs reflect rapid generational turnover, with A100s trading at deep discounts (~30-50% of original MSRP) as hyperscalers decommission clusters for newer hardware, while H100s retain stronger residual value (~40-60%) amid ongoing demand for inference and mid-tier workloads.[1][1]

• A100 (Ampere, 2020): 40GB SXM4/PCIe used prices cluster around $3,500–$6,500 (eBay lows ~$2,350–$6,500; indicative Q3 2026 ranges $4,000–$6,000). 80GB variants trade at $5,000–$9,500 (some listings up to ~$13,000). A full 9,600-GPU hyperscaler cluster listed on eBay equated to ~$6,219 blended per GPU (including infrastructure).[2][3][4]
• H100 (Hopper, 2022): SXM 80GB used ~$14,250–$27,000 (Q2 2026 data showed $15k–$19k realized; later ranges $22k–$27k indicative). PCIe variants lower at ~$10k–$16,500. Residual value ~42–58% after 24–30 months vs. original ~$28k–$33k list.[1][5][6]
• V100 (Volta, 2017): 32GB used ~$700–$910 (eBay sold median ~$860; indicative liquidation $1k–$2k). Far below original ~$10k–$11.5k MSRP, reflecting end-of-life status for most commercial AI.[7][8]

These prices signal a vibrant but tiered resale market: A100s flood secondary channels from hyperscaler refreshes (e.g., 9,600-unit eBay lot from a hyperscaler environment), H100s command premiums due to versatility, and V100s approach liquidation.[9][4] For entrants or competitors, buying used A100/H100 offers capital-efficient entry for inference or fine-tuning but requires factoring in higher power draw, potential reliability risks on older silicon, and competition from cloud rentals (A100 ~$0.34–$1.75/GPU-hr; H100 higher).[10]

Nvidia maintains an annual architecture cadence, with Hopper (H100/H200, ~2022) giving way to Blackwell (B200/GB200/GB300, ramping 2024–2025/early 2026) and then Vera Rubin (Rubin GPUs + Vera CPUs, first production shipments H2 2026, NVL72 systems generally available by fall/November 2026).[11][12][13]

• Hopper introduced FP8 Transformer Engine and strong NVLink scaling.
• Blackwell added native FP4, higher bandwidth (HBM3e up to 8 TB/s per GPU), and rack-scale NVL72 designs for massive efficiency gains.
• Rubin (shipping/ramping as of Nov 2026) features HBM4 (288 GB/GPU, ~19.2–22 TB/s bandwidth after spec refinements), NVLink 6, and co-designed Vera CPU; some bandwidth specs were quietly adjusted downward (~13% HBM4, ~17% NVLink per GPU) to reflect volume-production realities, but overall platform targets remain aggressive.[11][14][15]
• Next steps include Rubin Ultra (~2027) and Feynman (~2028).

This ~12-month cadence compresses product lifecycles and accelerates secondary-market supply as operators upgrade for perf/watt advantages. New entrants face a moving target: designing around one generation risks quick obsolescence unless focused on software optimization or niche workloads.

Performance-per-watt and cost-per-token gains compound dramatically across generations, driven by precision formats (FP8 → FP4), memory bandwidth, interconnects (NVLink scaling), and full-stack co-design (rack-scale NVL systems, software like TRT-LLM).[16][17]

• A100 to H100: ~1.8x FLOPS/W (FP16/BF16 tensor; e.g., ~0.78 to ~1.41 TFLOPS/W system-level in some comparisons); FP8 support enables further effective gains. Cost-per-token for inference drops substantially (H100 ~$0.27–$0.28/MTok on 70B FP8 vs. higher on A100 equivalents).[18][19]
• Hopper to Blackwell (B200/GB300): Up to 25x perf/watt on certain MoE models (SemiAnalysis InferenceX); 10–50x throughput per MW; 15–40%+ lower cost-per-token (e.g., B200 ~2.5x throughput at ~1.5x price → ~40% cost/token cut; some benchmarks show 35x lower token cost vs. Hopper). FP4 enables major inference efficiency.[16][17][20]
• Blackwell to Rubin: Up to 10x higher inference throughput per MW and ~10x lower token cost vs. Blackwell on select workloads (e.g., agentic models); benchmarks show 9–21x+ tokens per dollar vs. GB300 at matched speeds, with some previews citing higher multiples (up to 30x in Nvidia commentary).[21][22][23]

These gains (often 1.6–2.5x+ effective per generation, amplified by software) imply economic lives of 2–4 years for flagship use before power/performance economics favor replacement, though inference extends viability. Power-constrained data centers amplify this: newer gens deliver more tokens per watt (and thus per MW of grid allocation), but higher per-GPU TDP (A100 400W → H100 700W → B200 1,000W+) means fewer GPUs per rack/MW, offset by vastly higher output.[24]

Implication: Competitors or buyers must model TCO including power (PUE effects), utilization, and residual value. Used older GPUs win on capex for low-intensity inference; new gens dominate high-volume or frontier workloads. Long-term, power availability (not just silicon) becomes the binding constraint.

Hyperscalers and analysts report accelerating retirement of A100 clusters (e.g., large decommissioned supercomputers entering secondary markets) as Blackwell ramps, with H100s shifting to inference or secondary uses rather than immediate full retirement; older chips like V100 linger in research/legacy roles.[25][26][27]

• A100 fleets are being displaced for training but repurposed for inference/fine-tuning; CoreWeave extended A100 contracts into 2029, showing retained earning power.[25]
• H100 supply remains tight in secondary markets due to strong productivity (rental rates above depreciation assumptions per a16z analysis); some owners redeploy internally rather than sell.[26]
• Analysts note GPU useful life may exceed traditional 3–6 year depreciation due to inference demand; resale volumes rising (e.g., 980k units projected annualized Q2 2026).[1][25]
• V100s approach end-of-life liquidation for commercial AI but retain niche value.

For market entrants: Plan for multi-generation fleets—buy used A100/H100 for cost-sensitive inference while reserving capex for Blackwell/Rubin on power-dense, high-margin workloads. Monitor hyperscaler refresh cycles for secondary supply gluts. Software optimization (quantization, frameworks) extends older hardware economics more than raw specs suggest. Power and cooling infrastructure decisions made today will dictate upgrade flexibility for years.

---

Recent Findings Supplement (November 2026)

Secondary-market prices for used A100, H100, and V100 GPUs have stabilized or declined modestly in late 2026, with A100 80GB cards trading around $12,500 and H100 SXM around $14,250 realized, reflecting sustained demand for inference workloads amid Blackwell/Rubin ramps.[1]

• As of November 4, 2026, used A100 80GB cards list from $12,500 on eBay (stable since September), with new/refurbished higher at ~$22,000; weekly tracking shows consistency through October.[1]
• October 29, 2026 dealer data shows realized H100 SXM 80GB sales at $14,250 (dealer asks ~$16,576), H100 PCIe $10k–$14k, and H200 SXM modeled ~$22,500 (thin market); A100 values appear in broader $4,800–$18,900 ranges across sources, with some liquidation values lower (~$5,400 fair market for SXM).[2]
• V100 32GB used prices reached ~$910 on eBay as of November 4, 2026 (median asks ~$1,099), with fluctuations but availability in the secondary channel.[3]
• Broader October–November 2026 indices (e.g., American Compute Rack Report through August data) show used H100 retention at 75–85% after 18–24 months and A100 steeper declines to 30–50% of original list, with resale indices down to 75 (H1 2025 baseline = 100) in August 2026.[4]

This pricing dynamic means secondary buyers can acquire capable inference hardware at 40–60% of new equivalents, pressuring new-gen premiums while extending fleet utility.

Nvidia maintained its annual cadence with Blackwell (Hopper successor) ramping in 2025 and Vera Rubin entering production shipments and NVL72 general availability by late 2026, with minor spec revisions but strong hyperscaler uptake.[5]

• Rubin GPU shipping specs (as of October 31, 2026) include 288 GB HBM4 (unchanged), revised down to 19.2 TB/s memory bandwidth and 3.0 TB/s NVLink scale-up (from preliminary 22 TB/s and 3.6 TB/s), with scale-out raised slightly; NVL72 systems now generally available at hyperscalers.[5]
• Jensen Huang highlighted platform revenue per gigawatt rising to ~$40 billion for Vera Rubin (vs. $25 billion for Grace Blackwell and $18 billion for Hopper), tied to productivity X-factors and AI infrastructure spend projections of $3–4 trillion by 2030.[6]
• First NVL72 systems became generally available in fall 2026 on the original H2 2026 schedule, with production hardware arriving and purchase orders from major customers.[7]

Competitors entering the secondary or cloud market must time purchases around these annual waves, as Rubin availability will likely pressure H100/H200 values further in 2027.

Blackwell and Rubin deliver substantial performance-per-watt and cost-per-token gains over Hopper, with Rubin showing up to 30x throughput per megawatt and 45x lower token costs versus GB300 NVL72 on agentic workloads, driven by higher memory bandwidth, NVFP4 support, and rack-scale optimizations.[8]

• SemiAnalysis benchmarks (updated through October 2026) show Vera Rubin NVL72 achieving ~2.5x FP4 inference and ~2.3x training compute per rack versus GB300 NVL72, with memory bandwidth ~2.4x higher (1,400 TB/s vs. 576 TB/s) and NVLink ~1.7x; power draw rises to 156–227 kW per rack.[9]
• Agentic inference results indicate up to 5.7x throughput per MW on specific models versus GB300, with peaks supporting 272M tokens/s per MW; cost-per-token advantages reach 45x lower on some workloads due to efficiency at scale.[10]
• Blackwell (e.g., B200/GB200) examples show ~10x+ throughput per MW and 15–35x lower cost per million tokens versus Hopper on MoE/inference models, with software optimizations (TensorRT-LLM) adding further 4–5x gains post-launch.[11]

These gains accelerate replacement economics in power-limited environments, where older chips lose relative value despite absolute capability.

Power-constrained data center economics imply shorter effective economic lives for less-efficient generations (potentially 3–5 years for frontier training, longer for inference), as each watt favors chips maximizing tokens per MW; however, real-world retention exceeds traditional 3-year assumptions due to secondary demand and long-tail workloads.[12]

• Revenue-per-GW metrics and benchmarks tie directly to power envelopes: Rubin’s efficiency supports higher output without proportional power increases, while hyperscalers face 30–45% shortfalls in planned 2027 capacity due to grid/electrician constraints.[13]
• Older GPUs retain value when they remain the lowest-cost option for specific jobs; e.g., CoreWeave A100 contracts extend to 2029 (ninth year of the chip), with prior-gen pricing holding or rising amid shortages.[12]
• Depreciation trackers show H100 at 60–75% retained value by year 3 and floors around 25–35% beyond year 5, better than book assumptions but pressured by next-gen efficiency.[14]

Firms competing here should model TCO around actual MW utilization rather than nominal depreciation, prioritizing inference repurposing.

Hyperscalers and operators are actively retiring or rotating older fleets (especially H100 for Blackwell space and air-cooled incompatible racks), with rapid secondary-market absorption, repurposing for inference, and new platforms like Compute Exchange facilitating sales; V100-era retirements confirm multi-year lives.[15]

• 2026 sees the largest concentrated AI GPU retirement wave, with H100 clusters displaced by higher-density Blackwell systems and air-cooled facilities unable to host liquid-cooled successors; neoclouds (CoreWeave, Lambda) actively buy retired hyperscaler inventory.[16]
• Secondary volume has grown sharply (projected ~980k units annualized in Q2 2026 data), with H100 days-to-sell shortening and institutional intermediaries rising; Azure’s prior V100 retirement (pre-2026) and ongoing T4 activity illustrate 7–9 year useful lives in practice.[17]
• July 2026 launch of Compute Exchange secondary marketplace for H100/A100 reflects enterprise/cloud demand for non-frontier hardware; H100 supply remains tight in physical secondary channels despite new-gen ramps.[18]

New entrants or refurbishers can capture value by targeting inference buyers and timing sales ahead of major GA events, but must navigate export controls and compliance on large lots.

Michael Burry’s November 2025 claim (detailed in his X post and Substack “Unicorns and Cockroaches: Blessed Fraud”) asserts that hyperscalers (primarily Microsoft, Alphabet/Google, Amazon, Meta, and Oracle) are systematically extending the estimated useful lives of servers and networking equipment—including Nvidia GPU-based AI hardware—from prior shorter schedules (often 3–5 years) to 5–6 years (or 5.5 years at Meta).[1][2]

This flattens annual depreciation expense on rapidly cycling hardware (Nvidia’s product cadence of roughly 1–2 years, with each generation offering major efficiency gains), overstating near-term earnings. He quantified the cumulative effect as ~$176 billion in understated depreciation (and thus overstated profits) across 2026–2028 for the group, with specific 2028 earnings overstatements of ~26.9% at Oracle and ~20.8% at Meta.[1][3]

Burry did not publish the full bottom-up model, but it derives from scaling the companies’ disclosed AI-related capex (hundreds of billions cumulatively, with 2025 hyperscaler capex already in the $200–300B+ range and rising sharply) against the difference between a 2–3 year economic life (tied to product cycles and obsolescence for frontier workloads) and the booked 5–6 year straight-line schedules. Independent reconstructions (e.g., assuming ~$600B+ relevant GPU/server capex and a 3-year vs. 6-year life) produce annual understatements in the $50–60B range, compounding to roughly Burry’s three-year total.[4][4]

Prior extensions produced large, disclosed one-time benefits that illustrate the mechanism:
- Microsoft (FY2023): 4→6 years added $3.7B to operating income / $3.0B to net income.[1]
- Alphabet (2023): Shift toward 6 years cut depreciation by ~$3.9B and added ~$3.0B to net income.[5]
- Meta (2025): Extension to 5.5 years for most servers/network assets reduced 2025 depreciation by ~$2.92B (boosting net income by ~$2.59B or $1.00/share).[6]
- Oracle (FY2025): 5→6 years lowered operating expenses by ~$733M.[1]
- Amazon had earlier extensions (including 5→6 years) with multi-billion benefits, then partially reversed course in 2025 (subset of servers/networking from 6→5 years due to “increased pace of technology development, particularly in AI”), adding ~$1.4B to 2025 depreciation and cutting net income by ~$1.0B.[1][7]

Other analysts have modeled the earnings impact of shorter lives (3–4 years). Reconstructions suggest material overstatements, often in the high single digits to low double digits or more as a percentage of operating income for exposed names like Meta and Oracle (whose smaller overall earnings bases amplify the percentage effect). For Meta, a further one-year shortening could add >$5B to annual depreciation in some models. Barclays analysts modeled higher-than-consensus depreciation at Meta and concluded shares were 5–25% more expensive than Street assumptions implied. Morgan Stanley has flagged rising depreciation risk not fully in consensus forecasts, with potential for significant P&L pressure as AI capex scales. Broader estimates tie the accounting choice to roughly a third of reported AI-related operating income in some analyses.[8][6]

Counterarguments from hyperscalers, sell-side analysts, Nvidia, and accounting perspectives emphasize tiered utilization, observed longevity, and disclosure transparency rather than a uniform 2–3 year cycle.[9][7]

Key points include:
- Tiered/“cascaded” lifecycles and inference reuse: Newer GPUs handle frontier training workloads (where obsolescence is fastest); older ones (e.g., A100s from 2020 or H100s) shift to inference, fine-tuning, or lower-intensity internal/edge uses where performance-per-watt demands are lower. Utilization often exceeds 85%, rental rates for H100s have held or risen in some periods, and older chips remain rentable profitably years later because operating costs (power + colocation) are low relative to revenue. Examples: A100s still earning multiples of running costs after 6 years; P100 instances running 9 years later at Google Cloud.[10][11]
- Useful life of A100s and empirical data: Nvidia and customers cite observed utilization and longevity supporting 4–6 year schedules. Resale markets, ITAD channels, and continued demand (including long-term contracts extending into 2029 for some A100 capacity) indicate residual value and earning power beyond 2–3 years. Jensen Huang has pointed to platform durability and software improvements extending value; some operators (e.g., Lambda, CoreWeave) describe effective lives of 6–8 years.[12][11]
- Accounting and disclosure: Changes are estimates under GAAP (straight-line is a convention, not a mandate; any systematic method reflecting expected benefit consumption is allowed). Impacts are quantified and disclosed in 10-Ks/earnings releases (as shown above). This is not “fraud” (a legal term requiring intent to deceive); it is management judgment. Longer schedules can understate near-term earnings relative to a shorter life but may align better with multi-year cash flows from diversified workloads. Amazon’s partial reversal shows companies adjust when evidence warrants.[13][7]
- Sell-side and expert views: Bernstein and others argue 5–6 years is reasonable given margins on legacy hardware. Some analyses note that rapid generational gains do not automatically obsolete prior fleets if demand (especially inference) is elastic and software stacks improve older silicon. Construction-in-progress assets and lease commitments add future depreciation pressure regardless of current schedules.[9][14]

The evidence leans toward the counterarguments on magnitude while acknowledging directional validity in Burry’s critique. Amazon’s 2025 reversal for AI-specific assets lends credence to concerns about accelerated obsolescence for certain hardware cohorts. Rapid Nvidia cadence and front-loaded training value create real risk of future impairments or accelerated depreciation if utilization or pricing for legacy fleets weakens. However, widespread real-world data—high sustained utilization, resilient rental/resale economics, tiered workload migration, and continued profitable operation of multi-year-old chips—supports longer effective economic lives than a strict 2–3 year product-cycle view implies. No major impairments or broad reversals have materialized as of late 2026, and disclosed benefits have been transparent. The $176B figure appears directionally plausible as an upper-bound sensitivity on aggressive assumptions but likely overstates the gap once tiered use and observed longevity are factored in.[1][12]

For competitors or investors, the debate highlights the importance of monitoring actual depreciation run-rates versus capex (depreciation has lagged capex growth at some names), utilization metrics, rental pricing trends, and any further life changes. Companies with more diversified or inference-heavy workloads may sustain longer schedules better than pure training-focused players.

---

Recent Findings Supplement (November 2026)

Michael Burry's core claim, first made in a November 10, 2025 X post and reiterated into 2026, is that hyperscalers (Microsoft, Alphabet/Google, Amazon, Meta, Oracle) have extended useful lives of servers/network equipment (typically to 5–6 years) while rapidly cycling Nvidia GPUs on 2–3 year product generations, understating depreciation expense by roughly $176 billion cumulatively from 2026 to 2028 and overstating earnings (e.g., Oracle by ~26.9% and Meta by ~20.8% by 2028).[1][2]

He framed this as a common modern-era accounting maneuver that artificially boosts earnings without changing cash flows. No detailed public derivation or full model from Burry has appeared in sources; analyses treat it as a plausible back-of-envelope estimate tied to the capex ramp and the gap between assumed (5–6 year) vs. economic (3–4 year) lives.[1]

One reverse-engineered check (using ~$526B annualized infrastructure spend for four hyperscalers in early 2026, assuming ~50% equipment share, straight-line depreciation, and batches stacking over 2026–2028) shows the $176B figure falls between scenarios for true 4-year (~$132B extra dep) and 3-year (~$263B extra dep) lives. It requires believing AI servers lose economic value faster than the companies assume.[1]

Recent company-specific disclosures quantify the earnings impact of life extensions (or the one reversal) in the billions, expressed relative to operating income or net income in the year of the change. These predate or align with the post-May 2026 discussion window:

• Microsoft (FY2023 change, servers/network 4→6 years): +$3.7B operating income, +$3.0B net income (~$0.40/share).[1]
• Alphabet (2023 change to 6 years): –$3.9B depreciation, +$3.0B net income (~$0.24/share).[1]
• Meta (2025 change to 5.5 years for certain servers/network): ~–$2.9B depreciation expense (i.e., earnings boost of similar magnitude, ~$1.00/diluted share).[1]
• Oracle (FY2025 change 5→6 years): –$733M operating expenses, +$573M net income.[1]
• Amazon (2024 extension to 6 years for servers, then 2025 reversal shortening a subset back to 5 years citing “increased pace of technology development, particularly in AI/ML”): The 2025 shortening added $1.4B depreciation and reduced net income by $1.0B (plus ~$920M accelerated depreciation on early retirements, impacting 2025 op income by ~$0.6B).[1][3]

Barclays analyst modeling (cited in Nov 2026 coverage) projected Meta depreciation well above consensus, implying shares 5–25% more expensive than Street assumptions under higher realistic depreciation.[4] No comprehensive new third-party estimates of overstatement as a precise % of operating income per company (beyond Burry’s earnings figures) emerged in post-May 2026 sources.

Counterarguments from hyperscalers, Nvidia, sell-side, and accounting perspectives emphasize that useful-life estimates are management judgments under GAAP, disclosed with quantified effects, and supported by ongoing utilization rather than rapid obsolescence.[3]

• Tiered lifecycles and reuse: Older GPUs (e.g., A100s, now ~6 years post-launch) remain fully utilized for secondary workloads, inference, testing, or lower-intensity tasks after frontier training shifts to newer silicon; high utilization and resale activity contradict short economic lives.[3]
• Resale/rental data and Nvidia pushback: Nvidia’s Sept 2026 investor materials highlighted infrastructure retaining value beyond accelerated schedules. H100 rental prices rose ~14% YoY into late 2026. Resale value analyses (mid-2026 report) show 6-year straight-line schedules align with or stay within moderate residual bands (conservative bands support 5 years); 4-year schedules fall below conservative floors in later years.[3]
• Accounting and disclosure: Changes are estimates reviewed annually; effects are footnoted and audited. Not “fraud”—standard practice. Efficiency gains (newer chips far more power-efficient per FLOP) extend viable economics. One strategist called Burry’s fraud framing an overreach, noting companies know their assets better than outsiders.[5]
• Other context: Princeton CITP (Oct 2025) had flagged 1–3 year realistic lives, but real-world data since then (utilization, rentals, resale) has been cited against it. Capex continues outpacing depreciation growth (e.g., Meta capex +87% vs. dep +18%; Microsoft FY2026 capex +80% to $115.9B vs. dep +56%), but this lag is expected with heavy buildout.[4]

Evidence as of late 2026 is mixed but leans toward the companies’ assumptions being defensible for now, with Amazon’s partial reversal as the clearest signal supporting Burry’s direction on pace of change. High ongoing utilization, rising/ stable secondary-market values for recent generations, and disclosed (non-hidden) accounting effects weaken the “understatement as fraud” framing and the most aggressive 3-year life scenarios. However, the capex–depreciation divergence, Amazon’s AI-cited shortening, and potential for future impairments or accelerated charges if demand/utilization disappoints keep the debate live—particularly into 2027–2028 as more hardware enters service.[1][4]

No major new regulatory or policy updates on depreciation accounting for AI hardware appeared in post-May 2026 sources. Watch for further life changes in 10-K/10-Q filings, impairment charges, or quarters where depreciation growth accelerates sharply relative to revenue.

GPU-collateralized loans and ABS in the AI infrastructure sector typically route through bankruptcy-remote SPVs that own the GPUs (often serial-numbered NVIDIA H100/H200/B200/Blackwell units), assign customer contracts (or take-or-pay MSAs), and pledge datacenter leases as collateral. Lenders (private credit like Blackstone, Apollo, Macquarie; banks; ABS buyers like PIMCO) advance against contracted cash flows more than raw hardware resale value, with delayed-draw mechanics aligning draws to hardware delivery and installation.[1][2]

This structure ring-fences risk: the SPV borrower isolates collateral from the parent operator, while parent guarantees (full or limited to “bad acts”) and equity pledges provide recourse. CoreWeave’s DDTL series (e.g., DDTL 4.0 $8.5B at SOFR + 2.25% / ~5.9% fixed, rated A3/A(low) and tied to a Meta take-or-pay MSA per Bloomberg reporting; DDTL 5.5 $2.6B at SOFR + 5.50%, rated Ba2/BB+) exemplifies the model, with each facility in its own SPV (e.g., CoreWeave Compute Acquisition Co. VIII, LLC or Financing DDTL V-V, LLC).[3][4]

Lambda follows a similar pattern with SPV or subsidiary facilities secured by GPUs plus contracted cash flows (e.g., its $926M Term Loan B at SOFR + 3%, rated Baa2, fully amortizing to 2030 and aligned with one IG offtaker; its $1.008B fixed-rate DDTL at 6.78%, rated A(low)/Baa1, backing three deployments with two unnamed IG offtakers and maturing May 2033).[5][6]

Fluidstack uses Macquarie-led GPU-secured senior debt (capacity up to $10B in some reports) paired with hyperscaler backstops (e.g., Google guarantees on hosting agreements), enabling scale beyond its ~$653M equity.[7]

xAI structures involve SPVs or sale-leasebacks (e.g., Valor Compute Infrastructure SPV leasing NVIDIA GPUs back to xAI for ~5 years; Apollo providing ~$3.5B debt on a $5.4B transaction with NVIDIA equity participation up to $2B; larger SpaceX/xAI master leases up to $40B).[8][9]

Meta/Blue Owl-style vehicles are off-balance-sheet project bonds or JVs for entire campuses (e.g., Hyperion/Beignet Investor LLC: $27.3B 6.581% senior secured notes due 2049, A+ rated, fully amortizing; Blue Owl funds ~80% equity/~$7B cash, Meta 20% with distribution; Meta leases back with 4-year initial term + renewals and a 16-year residual value guarantee).[10][11]

These are not pure hardware ABS but hybrid contract-backed financings; true GPU-only securitizations remain rare and smaller-scale.

Reported LTV/advance rates range 60-90% of GPU server cost (higher for IG contracts, e.g., 90% Microsoft vs. 75% OpenAI in CoreWeave DDTL 2.1), often sized via projected debt service coverage (DSCR) rather than independent appraisals. Amortization is typically monthly or quarterly straight-line or cash-flow-matched, with full or near-full amortization over the loan term (3-6+ years) to avoid large balloons; some include cash sweeps, reserves (e.g., 3 months debt service + opex), and re-sizing on contract renewal.[12][13]

Contract coverage is central: loans require eligible customer contracts (often IG or take-or-pay) sufficient to service debt, with covenants like minimum DSCR (1.15-1.40x post-draw), contract realization ratios (e.g., 0.85x cash conversion), and renewal tests (new qualifying contracts must support 2.40x remaining loan value or trigger prepay in DDTL 5.5).[14][1]

Pricing tiers sharply by customer credit: IG-backed (Meta, Microsoft) at SOFR + 2.25-3.00% or ~5.9-6.78% fixed (IG-rated facilities); non-IG or shorter contracts at SOFR + 4.5-5.5% or higher (earlier deals reached +8.75-13%).[15]

Loans are highly sensitive to assumed GPU useful life (commonly modeled 4-6 years for depreciation/financing, vs. Nvidia’s claim of 8+ years) and residual value because many facilities extend beyond initial contract terms, shifting reliance to re-leasing, renewal, or resale. CoreWeave DDTL 5.5 (~5-year maturity, ~3-year average contracts) exemplifies this: repayment depends on renewal/re-lease tests tied to DSCR and contract-value ratios; failure triggers prepayment.[4][16]

If lives are shorter than assumed (e.g., rapid obsolescence from newer Blackwell/ Rubin chips or power/cooling limits on legacy racks), collateral coverage erodes quickly. One analysis finds economic life of 4.3-7.2 years depending on TDP, with a 700W accelerator retaining only ~25% of value after 4 years under income-anchored modeling; structures extending past ~4 years can lose coverage within 18 months.[17]

Resale data shows variability (older H100s at 60-70% after ~4 years in some reports, but sudden drops possible); no liquid futures market yet fully hedges this, though emerging GPU rental futures are noted.[18]

Shorter lives trigger defaults, forced sales at depressed values, or equity calls; lenders shift underwriting toward contract cash flows and DSCR tests rather than standalone LTV on hardware.[19]

Disconfirming evidence for overstated bear case (longer lives correct): Hyperscalers have extended depreciation schedules to 5-6 years (Meta to 5.5, others to 6), aligning with observed utilization; Nvidia cites fungibility, durability, and high residual value, arguing GPUs are “lowest risk” collateral with 8+ year potential.[20][21]

Strong demand (prepay-heavy take-or-pay contracts, hyperscaler offtake) supports renewals at attractive rates; SPV structures and parent guarantees provide buffers. Some facilities explicitly size advances on depreciable cost + projected coverage, not aggressive residuals.[12]

Cases where bear case understates risk: Demand slowdown (e.g., AI capex pause or cheaper alternatives like custom TPUs) could crater renewal pricing or utilization, amplifying concentration risk (many facilities tied to 1-2 large counterparties like Meta/Microsoft/Google). Price crashes in secondary GPU markets (already variable) or power constraints would worsen residuals. Counterparty concentration (e.g., single-MSA deals) or correlated failures across neoclouds could cascade.[19]

Longer-dated structures (e.g., Meta/Blue Owl 2049 bonds with 16-year RVG) embed more residual exposure; off-balance-sheet vehicles still expose sponsors via guarantees/leases.[10]

For competitors or new entrants: Success hinges on securing IG or high-quality take-or-pay contracts to access lower-cost capital (pricing spreads of 200-300+ bps between IG and non-IG); building diversified offtake and remarketing capabilities reduces renewal risk. SPV isolation and covenant-heavy structures favor operators with strong execution track records. Pure hardware plays without contracts face higher rates or equity requirements (20-40% skin in the game common). Monitoring Nvidia depreciation shifts, secondary market data, and emerging futures will be key to navigating residual uncertainty.[15]

Sources drawn from SEC filings, rating agency releases, company announcements, and analyst reports (2025-2026). Specific numerical claims (sizes, rates, covenants) cite the listed web results.

---

Recent Findings Supplement (November 2026)

CoreWeave's DDTL 5.5 ($2.6B, closed ~August 10, 2026) exemplifies the maturing SPV structure for shorter-duration contracts.[1][2] The facility finances GPU servers and infrastructure in a dedicated subsidiary (CoreWeave Financing DDTL V-V, LLC) against customer contracts via delayed draws through December 2026. It matures September 1, 2031 (~5 years), prices at SOFR + 5.50%, carries Moody's Ba2/Fitch BB+ ratings, and includes a full unconditional parent guarantee from CoreWeave Inc. Customer contracts average ~3 years (diverse AI, financial services, and tech counterparties), shorter than the debt, so the structure explicitly permits renewal or re-leasing subject to eligibility tests and coverage covenants.[1][3]

• Key terms from filings and analyses: DSCR minimum of 1.35x (starting after draw period or Dec. 31, 2026); cash trapped in collateral accounts; security on substantially all SPV assets (GPUs, contract rights/payments, data center leases, equity pledge); borrowing base and sizing tied to contracted cash flows and depreciated GPU value rather than pure hardware cost.[4][5]
• Advance rates/LTV around 70% (with reports of ~80% of purchase price at closing, capped at 85% of depreciated value under a 6-year straight-line depreciation assumption used in covenants).[6][7]
• Similar mechanics appear in DDTL 5.0 ($3.1B, closed May 18, 2026; SOFR + 4.50%; two non-IG customers; same ratings and DSCR).[1]

This shifts risk from pure residual hardware value toward contract performance and re-leasing capability, with parent support as a backstop.

Lambda's August–October 2026 facilities show the template scaling to IG offtakers with full amortization. The $926M Term Loan B (Aug 27, 2026; Moody's Baa2; SOFR + 3.00%) funds an NVIDIA take-or-pay contract in a Quebec SPV, fully amortizing to ~2030.[8] The $1.008B fixed-rate DDTL (closed Oct 1, 2026; DBRS A(low)/Moody's Baa1; 6.78% fixed, semi-annual) supports three deployments across two unnamed IG offtakers (GB300/Vera Rubin GPUs in multiple sites), with delayed draws tied to commissioning milestones, full amortization to May 2033, and security on GPUs/infrastructure plus contracted cash flows.[9][10]

• These are among the first IG-rated GPU facilities marketed to insurers/fixed-income investors, with the rating tracking the offtake credit rather than the borrower.
• Amortization aligns repayment with contracted cash flows and assumed useful life, reducing balloon risk at maturity.

xAI/SpaceX structures emphasize large-scale sale-leaseback/SPV leasing. A October 2026 $40B master finance lease commitment (largest on record) with an Apollo-managed entity (anchors including Goldman Sachs, BlackRock, KKR) covers NVIDIA Vera Rubin GPUs for xAI terrestrial data centers. Leases run ~5 years with quarterly payments recovering full cost; schedules commence on 2027 deliveries.[11][12] Earlier xAI SPVs (e.g., Valor-led with Apollo/Diameter debt and NVIDIA equity participation) follow similar GPU-owning SPV + lease-to-operator patterns.

Meta/Blue Owl-style vehicles remain more campus/lease-focused with residual guarantees. Examples include Meta's Hyperion JV (~$27B bonds, residual value guarantees) and Fluidstack-related leases backstopped by Google. These contrast with pure GPU SPV loans by emphasizing real estate and long-term tenant credit.[13]

GPU useful life and residual value sensitivity center on 5–6 year assumptions with re-leasing risk. Covenants (e.g., in CoreWeave DDTL 2.0) explicitly use 6-year straight-line depreciation for "GPU Depreciated Amount."[14] Contracts often run 3–6 years, so facilities like DDTL 5.5 explicitly price in renewal/re-lease at "favorable GPU lease rates" (Fitch rating case).[15] LTVs start at 60–80% (higher for IG contracts) and are managed via amortization (typically 25–30% per year) to offset depreciation.[16]

• If lives are shorter than assumed (bear case): Faster obsolescence (e.g., new generations halving replacement costs every ~4 years) erodes collateral value faster than amortization, pushing LTVs higher and risking covenant breaches (DSCR, contract realization ratios ~0.85x). Lenders could force paydowns, repossess, or face lower recovery on foreclosure/sale in a weak secondary market. Demand slowdown, counterparty concentration, or price crash would amplify this.[17]
• Disconfirming evidence (longer lives or overstated bear case): Sustained hyperscaler demand supports renewals at attractive rates; actual resale data and Nvidia ecosystem effects maintain values; insurance products for residual risk are emerging; strong contract coverage (e.g., take-or-pay from IG names) and parent guarantees provide buffers. Fitch and others explicitly model favorable re-leasing for facilities like DDTL 5.5.[15]

Emerging ABS and middle-market activity diversifies exposure but introduces pooling risks. October 2026 deals (Stonebriar ~$869M ABS with ~15% GPU loans; Wingspire $407M with ~20%) mark early entry of GPU loans into equipment ABS pools.[18] Smaller facilities (e.g., USD.AI-led $106M for NexGen Cloud, 4-year term, take-or-pay contracted) show standardized lending at 70–80% LTV with 3–5 year amortization.[19]

Implications for competitors or new entrants: Success hinges on securing high-quality, long-duration or renewable offtake (IG hyperscalers command 200+ bps tighter spreads) and demonstrating re-leasing capability. Pure hardware plays face higher rates and lower advance rates; structures with strong cash-flow covenants and SPV isolation reduce but do not eliminate re-contracting or obsolescence risk. Watch secondary market development, actual recovery data from any workouts, and regulatory scrutiny on valuation models (e.g., Korean banks citing gaps). No major post-May 2026 policy shifts appear in public sources, but deal volume and ratings activity confirm rapid market maturation around contracted cash flows over raw silicon value.