Analyze the publicly known economics of GPU hardware depreciation in the neocloud context — specifically how H100/H200/B200…

CoreWeave’s shift to a six-year straight-line GPU depreciation schedule (from four years in 2023) exemplifies how neoclouds stretch accounting useful lives to match revenue from long-term AI training and inference contracts, even as NVIDIA’s 18–24-month architecture cadence (Hopper H100/H200 → Blackwell B200 → Rubin in H2 2026) compresses real economic life. This creates a growing book-to-market gap on balance sheets, where assets like H100 clusters (new ~$25k–$40k per GPU) retain rental viability through workload cascading but face accelerating value erosion once newer, far more efficient silicon floods the market.[1][1]

• CoreWeave, Nebius, and hyperscalers (AWS, Google, Microsoft) largely standardized on 5–6 year schedules by 2023–2025; Nebius uses a more conservative 4 years, Lambda around 5 years.[2][3]
• Amazon shortened select server lives from 6 to 5 years in 2025 (citing AI pace), booking hundreds of millions in accelerated depreciation; Meta extended to 5.5 years.[4]
• Straight-line method is common; annual expense example for an 8× H100 server (~$250k hardware) at 6 years ≈ $41.7k/year.[5]

For new entrants or competitors, matching or beating these schedules requires either hyperscaler-scale diversification (non-AI workloads extending life) or aggressive utilization/ pricing power; shorter schedules improve credibility with lenders but pressure near-term margins and raise capital needs.

H100/H200 hardware loses 15–30% of market value in year one and another 15–25% in year two as B200 volume ramps, with secondary prices for lightly used units falling to 70–85% of new and moderate-use (2–3 years) units to 45–70%, driven by Blackwell’s 2.5×+ AI performance and dramatically better perf/watt (especially inference). This mechanism strands frontier training capacity on older silicon while older GPUs cascade to lower-margin inference or batch jobs, but only if power and demand support it; rental rates for H100s collapsed from $8–16/hr peaks to ~$2–3/hr, amplifying the mismatch with accounting assumptions.[6][7]

• Used/refurbished H100s traded as high as $50k at 2024 scarcity peaks before dropping sharply; 12-month used server nodes (~$240k new) resold around $170k in some reports.[8][9]
• CoreWeave reports counter-evidence: 2022 H100 batches rebooked at 95% of original pricing upon contract expiry, with 2020 A100s still fully utilized.[1]
• Rubin platform (shipping H2 2026) continues the annual cadence, pressuring B200 values similarly.[10]

Competitors hedging this must prioritize flexible capacity or specialized workloads; pure-play neoclouds face higher stranding risk than diversified hyperscalers unless they secure multi-year offtake that outlasts one or two generations.

Analyst commentary (notably Michael Burry) flags multi-billion-dollar write-down risk for neoclouds carrying GPU-collateralized debt, with industry estimates of ~$176B in understated depreciation if 2–3 year economic lives prove accurate versus 5–6 year accounting; the thin secondary market and power constraints exacerbate balance-sheet exposure when contracts expire or utilization dips. Rapid turnover strands assets because newer chips deliver order-of-magnitude efficiency gains, making older hardware uneconomic to run at prevailing power prices even if still rentable for niche uses.[8][11]

• Burry and others argue 6-year schedules are unrealistic given NVIDIA’s cadence; some model 2–3 year economic lives.[12]
• GPU-backed debt (neoclouds >$20B estimated) relies on residual value assumptions that lenders scrutinize; default risk rises if resale or re-leasing fails to cover amortization.[4]
• Counter-data: sustained demand and “value cascade” (frontier training → inference → batch) have kept utilization high so far.

For capital-intensive entrants, conservative modeling (shorter lives, higher residual haircuts) or off-balance-sheet structures are essential to avoid covenant breaches or equity calls when the next generation lands.

Neoclouds hedge via 3–5 year committed rental contracts that lock revenue ahead of depreciation, emerging secondary/resale channels for partial recovery, and financing structures including GPU-collateralized loans with embedded residual-value assumptions; CoreWeave additionally benefits from a $6.3B NVIDIA backstop to purchase unsold capacity through 2032. These mechanisms transfer some risk to customers or suppliers but leave residual exposure if demand or efficiency economics shift faster than modeled.[13]

• Long-tenor contracts (tracked by SemiAnalysis across 3 months to 5 years) provide visibility and support higher utilization assumptions.[14]
• Secondary markets allow resale at 50–85% depending on age/condition, though liquidity remains limited and discounts steepen post-new-gen launches.[15]
• Financing often pairs customer contracts with GPU collateral; NVIDIA’s offtake commitment acts as a de facto put option for CoreWeave.

New players should structure similar backstops or diversified offtake early; relying solely on spot/on-demand or short contracts amplifies stranding risk in a market where one generation’s economics can be upended within 18–24 months.

Overall, the publicly visible economics show a tension between accounting optimism (5–6 year lives enabling massive CapEx) and operational reality (rapid performance leaps and falling rentals), with hedging succeeding so far through contracts and backstops but vulnerable to any sustained demand or efficiency shock.

---

Recent Findings Supplement (June 2026)

CoreWeave (and to a lesser extent other neoclouds) continues to defend 6-year straight-line GPU depreciation schedules into 2026 by citing sustained utilization and re-leasing data, but variations across peers (Nebius at 4 years, Lambda at 5 years) and hyperscaler adjustments highlight growing divergence in reported economics versus observed generational turnover.[1][1]

• Early 2026 server costs for an 8x H100 SXM configuration ranged from $250,000–$400,000 depending on OEM bundling and networking.[1]
• CoreWeave’s 6-year schedule produces ~$50,000 annual depreciation per $300k server versus ~$75,000 on a 4-year schedule (a $25k/server/year gap that scales dramatically at fleet level).[1]
• CoreWeave publicly extended its technology equipment useful life from 5 to 6 years starting in 2023; Nebius uses 4 years and Lambda 5 years.[1][2]
• Hyperscalers largely converged on 5–6 years (Meta more conservative at 5–5.5 years), with Amazon shortening a subset of servers/networking from 6 to 5 years effective January 2025, producing hundreds of millions in higher 2025 depreciation expense.[3]

This accounting choice directly affects balance-sheet presentation and perceived risk for capital-intensive neoclouds, as longer schedules lower near-term expense but amplify potential future impairments if economic life proves shorter.

Michael Burry’s November 2025 critique—that hyperscalers and peers are understating depreciation by stretching GPU lives to 5–6 years when economic reality is closer to 2–3 years—remains a focal point of analyst discussion through mid-2026, with his cumulative $176 billion overstatement estimate for the top five hyperscalers (2026–2028) frequently referenced.[4][5]

• Multiple 2026 analyses converge on ~2–3 year economic half-life for heavily utilized AI GPUs due to thermal/electrical stress and obsolescence, versus accounting lives of 4–6+ years.[6]
• Goldman Sachs (May 2026) modeled the sensitivity of annual depreciation to useful-life assumptions ranging from 3 to 7 years, underscoring material earnings impacts.[7]
• CoreWeave CEO Michael Intrator countered in November 2025 (still cited in 2026 commentary) with data-driven evidence: 2020-era A100s remain fully booked for inference, and a batch of 2022 H100s re-leased immediately at 95% of original pricing after contract expiration.[8][3]

These debates create tangible financing and valuation friction for neoclouds reliant on debt collateralized by GPU fleets.

Secondary-market pricing for H100s has stabilized at materially lower levels in 2026 after earlier sharp declines, illustrating the stranding risk from rapid Hopper → Blackwell turnover while also providing a partial hedge via resale liquidity.[9]

• Used/refurbished H100 prices stabilized around $18,000–$22,000 per GPU in early 2026 (down from peaks near or above $40k–$50k during 2024 scarcity).[10][9]
• Reports indicate H100 secondary values fell as much as 85% from peaks in some cases as Blackwell supply ramped; refurbished units retain 80–90% of contemporaneous new pricing better than used units (65–75%).[11][12]
• Rental rates for trailing-edge H100s recovered ~40% between late 2025 and early 2026 amid capacity tightness before softening again by May 2026.[13]

This volatility underscores asset-stranding potential as Blackwell (B200/GB200) and upcoming Rubin generations accelerate obsolescence, particularly for training workloads, though inference demand provides a partial buffer.

Neoclouds are hedging primarily through multi-year take-or-pay contracts, asset-backed non-recourse financing structures (including SPVs), and workload cascading to inference, rather than relying solely on secondary markets or accounting assumptions.[1]

• CoreWeave derives nearly all revenue from 2–5 year fixed-rate commitments, with a $66.8 billion backlog at end-2025 supporting $21.4 billion in debt (and $1.2 billion in 2025 interest expense).[1]
• Asset-backed loans (often 60–70% LTV, 12–36 month terms, SPV structures) isolate GPU collateral risk and close faster than traditional bank financing; non-recourse variants limit borrower exposure to the hardware itself.[14]
• “Value cascade” models (detailed in early 2026 analyses) posit GPUs shifting from frontier training (years 1–2) → production inference (years 3–4) → batch/analytics (years 5–6), supported by CoreWeave’s re-leasing data and hyperscaler service lives extending to 7–9 years in practice.[3]
• Rental models (vs. ownership) fully transfer depreciation/obsolescence risk to the provider; some neoclouds emphasize this for customers while absorbing it on their own balance sheets via debt.[15]

For new entrants or competitors, success hinges on securing long-duration offtake contracts and favorable financing terms before deploying capital, as shorter economic lives compress payback periods and increase the required rental yield to cover depreciation plus interest.[16]

Blackwell’s performance-per-watt gains and volume availability in 2026 are accelerating the economic obsolescence of H100/H200 fleets for certain workloads, amplifying the mismatch between 18–24 month NVIDIA architecture cycles and 4–6 year accounting lives.[3]

• H100 depreciation curves cited in 2026 financing guides: 20–30% in year 1 (new-gen announcements), 15–25% in year 2 (B200/B300 production), 20–30% in year 3 (next-gen adoption), with overall 30–50% annual value decline on owned hardware.[14]
• Post-24 months, H100 values are projected to depreciate an additional 10–20% annually as Blackwell expands; inference workloads mitigate but do not eliminate the pressure.[17]
• Analysts note that while H100 rental pricing held or recovered in spots into 2026, sustained Blackwell supply could force further repricing or accelerated refresh cycles for neoclouds.[18]

Competitors must model aggressive refresh assumptions (potentially 3–4 years for training-heavy fleets) and build secondary-market or trade-in channels early, as reliance on 6-year schedules risks sudden impairments or refinancing challenges when collateral values reset.

These developments, drawn primarily from 2025–2026 disclosures and analyses, show no fundamental resolution to the depreciation mismatch; instead, they highlight ongoing experimentation with contract structures and financing to manage the risk.