Are 6-year GPU depreciation schedules realistic?
Straight-line book value against our residual band forecasts for HGX H100, H200, B200 and B300 servers.
Residual forecasts by age vs straight-line book
Share of the price new, by age
Years since the generation first shipped
Solid lines: the forecast midpoint of each server's residual bands (conservative floor to moderate ceiling), placed at its age since the generation first shipped. Dashed: straight-line book to zero from the first-shipment date. Hollow points, the only observed values: today's listing-based A100 and V100 node values against a cited original list price; the DGX list price includes NVIDIA's own configuration, so read them as indicative.
What the data shows
These are forecasts, not observed resale at those ages. Against the residual bands in our June 2026 report, a 6-year straight-line schedule is aggressive but defensible for a server bought when its generation first shipped: book sits inside the moderate band at 8 of 10 points from years 3 to 5, and above the conservative band, the one lenders and lessors size to, at 8 of 10. A 5-year schedule sits inside the conservative band at 3 and below its floor at 7. Bought later, book runs further ahead: an HGX H100 bought used in 2026 carries 50% of cost after three years on a 6-year schedule, against 7-11% on our bands.
Updated October 17, 2026. Rack Report October 2026, data through August 2026.
Book value vs residual forecasts, years 3 to 6
| Age | 4-yr book | 5-yr book | 6-yr book | H100 forecast | H200 forecast | B200 forecast | B300 forecast |
|---|---|---|---|---|---|---|---|
| Year 3 | 25% | 40% | 50% | - | 34-59%(47) | 33-60%(47) | 34-61%(47) |
| Year 4 | 0% | 20% | 33% | 22-49%(36) | 24-51%(38) | 21-49%(35) | 23-50%(37) |
| Year 5 | 0% | 0% | 17% | 9-33%(21) | 7-37%(22) | 9-31%(20) | - |
| Year 6 | 0% | 0% | 0% | 4-11%(8) | 4-9%(7) | - | - |
Percent of the original price, for a server bought when its generation first shipped. Forecast: conservative floor to moderate ceiling of our June 2026 residual bands at that age, midpoint in brackets, read linearly between band years. Every cell falls after the report's data, so none is an observed resale price. A dash means the bands (2026 to 2030) do not reach that age for that server.
Verdict by schedule length
| Schedule | Against our bands | Evidence | Filers |
|---|---|---|---|
| 4 years | Conservative | Book sits above the moderate ceiling at 0, inside the moderate band at 0, inside the conservative band at 0, below the conservative floor at 10 of 10 forecast points (years 3 to 5); on average it runs 27 points below the band midpoint. | 1 |
| 5 years | Conservative | Book sits above the moderate ceiling at 0, inside the moderate band at 0, inside the conservative band at 3, below the conservative floor at 7 of 10 forecast points (years 3 to 5); on average it runs 15 points below the band midpoint. | 5 |
| 6 years | Aggressive but defensible | Book sits above the moderate ceiling at 0, inside the moderate band at 8, inside the conservative band at 2, below the conservative floor at 0 of 10 forecast points (years 3 to 5); on average it runs 2 points below the band midpoint. | 5 |
Points: each server at each age from 3 to 5 the bands reach, for a server bought at first shipment. At year 6 book is zero on every schedule, so those 2 points are left out. Lenders and lessors size to the conservative band; inside the moderate band is what our useful-life article calls aggressive but defensible. Filers: companies in the table below whose disclosed life is exactly that length (ranges such as 5 to 6 years are not counted). A useful life is an accounting estimate of how long an asset earns, not a resale forecast; servers under contract keep earning rent whatever they would sell for.
Bought in 2026, not at launch
Book value runs from the purchase date; the bands run from the generation's first shipment. A fleet bought after launch is already partway down the curve, so after three years it carries more book against the market than the tables above show.
| Bought in 2026 | Price | Forecast after 3 years | 5-yr book | 6-yr book | 6-yr book vs bands |
|---|---|---|---|---|---|
| HGX H100used | $195K | 7-11% | 40% | 50% | Above the moderate ceiling |
| HGX H200used | $349K | 16-50% | 40% | 50% | Inside the moderate band |
| HGX B200new | $415K | 16-46% | 40% | 50% | Above the moderate ceiling |
| HGX B300new | $425K | 28-54% | 40% | 50% | Inside the moderate band |
Percent of the purchase price, conservative floor to moderate ceiling of the 2029 band. Used servers are bought at today's node value and carried along the band ratios from 2026 to 2029, as on each server's page; new servers at the Rack Report new-price midpoint, read against the band directly. On a $1B fleet, five years books $200,000,000 of depreciation a year and six years $166,700,000.
A 6-year schedule like CoreWeave's
CoreWeave depreciates its GPU servers over 6 years (Form 10-K for 2025). It moved from 5 years to 6 years in Jan 2023.
For a server bought at first shipment, that schedule carries 50% of cost at year 3; our bands forecast 34-59% for HGX H200 servers at that age. At year 5 it carries 17%, against a forecast of 9-33% for the H100.
The one older generation we can price today held more than the Hopper bands project: HGX A100 8x 40GB nodes were valued at $59,100 in our latest data, about 30% of the DGX A100 list price at 6.1 years, against 4-11% our bands project for Hopper servers at year 6. Either the bands are conservative or Hopper will hold less than Ampere did; we use the bands. The node value is listing-based and the DGX list price bundles NVIDIA's own configuration, so read the A100 share as indicative.
Where the schedule meets risk is the conservative case: 6-year book sits above our conservative band at 8 of 10 forecast points, and further above it for any fleet bought after launch. A lender or lessor sizing to that band, as our residual report recommends, sees book value run ahead of what the servers would fetch if their contracts ended early. This compares book with our forecasts; it says nothing about whether CoreWeave's accounting is correct.
Useful lives 21 companies disclose in SEC filings
| Company | Useful life | Applies to | Filing |
|---|---|---|---|
| AkamaiAKAM | 3 to 7 years | Computer and networking equipment; servers average 6 years | Form 10-K for 2025 |
| AlphabetGOOGL | 6 years | Servers and network equipment | Form 10-K for 2025 |
| CoreWeaveCRWV | 6 years | Technology equipment (GPU servers) (GPU-specific) | Form 10-K for 2025 |
| DigitalOceanDOCN | 6 years | Servers and related equipment (including GPU Droplets) | Form 10-K for 2025 |
| MicrosoftMSFT | 6 years | Servers and network equipment | Form 10-K for fiscal 2026 |
| OracleORCL | 6 years | Servers and networking equipment | Form 10-K for fiscal 2026 |
| MetaMETA | 5.5 years | Most servers and network assets | Form 10-K for 2025 |
| Nscale | 5 to 6 years | Technology equipment (GPUs and related infrastructure) (GPU-specific) | Form S-1 |
| SpaceX (xAI)SPCX | 5 to 6 years | Servers and networking equipment (mostly AI segment GPUs) | IPO prospectus (424B4) |
| AmazonAMZN | 5 to 6 years | Servers; a subset shortened to 5 years | Form 10-K for 2025 |
| CloudflareNET | 5 years | Servers and network infrastructure (including Workers AI GPUs) | Form 10-K for 2025 |
| Hut 8HUT | 5 years | AI GPUs (its own line in the asset table) (GPU-specific) | Form 10-K for 2025 |
| IRENIREN | 5 years | HPC hardware (GPUs) (GPU-specific) | Form 10-K for fiscal 2026 |
| NebiusNBIS | 5 years | Servers and network equipment | Form 20-F for 2025 |
| WhiteFiberWYFI | 5 years | Cloud service equipment (GPU servers) (GPU-specific) | Form 10-K for 2025 |
| BaiduBIDU | 3 to 6 years | Computer equipment | Form 20-F for 2025 |
| AlibabaBABA | 3 to 5 years | Computer equipment and software | Form 20-F for fiscal 2026 |
| Applied DigitalAPLD | 3 to 5 years | Networking, electrical equipment and software | Form 10-K for fiscal 2026 |
| BitdeerBTDR | 4 years | GPU equipment (AI cloud) (GPU-specific) | Form 20-F for 2025 |
| Northern Data | 3 to 5 years | Servers and other operating equipment (Taiga Cloud GPUs) | Rumble Form S-4, Northern Data 2025 statements |
| Sharon AI | 2 to 5 years | Computer equipment (GPU servers); used equipment 1 year | Form S-1/A |
From each company's latest annual report or registration statement (18 listed companies, 3 from registration statements before listing), checked October 10, 2026. Every change since 2020: GPU depreciation schedules.
Older servers the market still prices
| Server | Age | Node value | List price | Share |
|---|---|---|---|---|
| HGX A100 8x 40GB | 6.1 years | $59,100 | $199,000 | 30% |
| DGX/HGX V100 8x 32GB | 9.1 years | $12,200 | $149,000 | 8% |
Node value: the latest Rack Report's Complete systems value (observed listings adjusted to a reference configuration, not a single sale). Age from first shipment. List price: NVIDIA list price for DGX A100 (8x A100 40GB), official, May 14, 2020; DGX-1 with V100 price as reported by Network World, May 10, 2017.
Methodology
How the comparison is built
The market side is a forecast: the residual bands in our June 2026 report, GPU Residual Value: Underwriting a $3 Trillion Opportunity, projected from completed resale transactions through June 2026. Bands are by calendar year as a share of the original price; we place each year at mid-year and measure age from when the generation first shipped (H100 late 2022, H200 mid 2024, B200 early 2025, B300 late 2025), then read ages 3 to 6 linearly between band years without extrapolating. Every point falls after the report's data, so none is an observed resale price at that age.
Book value is straight-line to zero with no salvage value, from the purchase date. The main tables assume a purchase at first shipment; later purchases are in the 2026 table. Each point is placed above the moderate ceiling, inside the moderate band, inside the conservative band or below the conservative floor, and a schedule takes the position most of its year-3-to-5 points fall in: aggressive, aggressive but defensible, in line with the conservative band, or conservative, as our useful-life article reads the bands. The test is book against forecast resale; it says nothing about whether a company's accounting is correct.
Related
- GPU server values: what every complete server is worth this month
- HGX H100 server value: residual curve, earnings and payback
- HGX H200 server value: residual curve, earnings and payback
- HGX B200 server value: residual curve, earnings and payback
- HGX B300 server value: residual curve, earnings and payback
- GPU depreciation schedules by company: every disclosed life and change since 2020
- GPU residual value by year: the bands, and which one each funder books
- In-situ vs ex-situ GPU value: installed and earning against removed and sold
Discuss a transaction
Booking a residual on a GPU lease, or sizing a loan against servers? Send us the equipment and the term. We will come back with values at the dates that matter.
Prefer email? hello@amcompute.com
Questions
- How does a 6-year schedule like CoreWeave's compare with resale forecasts?
- For a server bought at first shipment, 6-year straight-line book is aggressive but defensible against our residual band forecasts, above the conservative band at 8 of 10 points from years 3 to 5, and further ahead for fleets bought later. It says nothing about whether CoreWeave's accounting is correct.
- GPU depreciation schedule: 5 years vs 6 years?
- On a $1B fleet, five years books $200,000,000 of depreciation a year and six years $166,700,000, leaving $400,000,000 against $500,000,000 of book after three years. Against our band forecasts, five years is conservative and six years aggressive but defensible.
- How fast do GPUs depreciate?
- Our band midpoints project H200, B200 and B300 servers at about 47% of their original price three years after first shipment and H100, H200 and B200 at about 21% after five. These are forecasts from our June 2026 report, not observed sales.