The End-of-Life Model
See every retirement before the first rack arrives.
Our End-of-Life Model turns a capacity plan or an asset register into a retirement schedule, a decommissioning line for the total cost of ownership, a residual-value sensitivity and embodied carbon avoided. This is what it produces for an illustrative 150 MW AI gigafactory.
- First retirements
- 2030
- Accelerator racks retiring in the 5-year TCO
- 733
- Decommissioning in the TCO
- €7.3m
- Net resale in the TCO, sensitivity
- €136m to €340m
Illustrative: 150 MW installed between 2028 and 2032, refreshed every two years, with a fifth of accelerators redeployed for two more years. Figures are not a forecast for any project.
Continuous refresh means retirements every year
With annual adoption of new GPU generations, hardware starts leaving the site within two years of going live, so some of it retires inside a five-year total cost of ownership.
- 2028 capacity
- 2029 capacity
- 2030 capacity
- 2032 capacity
- CPU, storage and network racks
Show the data
| Year | 2028 capacity | 2029 capacity | 2030 capacity | 2032 capacity | CPU, storage and network racks | Total |
|---|---|---|---|---|---|---|
| 2028 (TCO) | 0 | 0 | 0 | 0 | 0 | 0 |
| 2029 (TCO) | 0 | 0 | 0 | 0 | 0 | 0 |
| 2030 (TCO) | 160 | 0 | 0 | 0 | 0 | 160 |
| 2031 (TCO) | 0 | 160 | 0 | 0 | 0 | 160 |
| 2032 (TCO) | 200 | 0 | 213 | 0 | 0 | 413 |
| 2033 | 0 | 200 | 0 | 0 | 0 | 200 |
| 2034 | 200 | 0 | 267 | 267 | 40 | 773 |
| 2035 | 0 | 200 | 0 | 0 | 40 | 240 |
| 2036 | 200 | 0 | 267 | 333 | 53 | 853 |
| 2037 | 0 | 200 | 0 | 0 | 0 | 200 |
| 2038 | 200 | 0 | 267 | 333 | 67 | 867 |
| 2039 | 0 | 200 | 0 | 0 | 0 | 200 |
| 2040 | 200 | 0 | 267 | 333 | 40 | 840 |
A small cost in the TCO, a large value in the sensitivity
Decommissioning is a fully costed line paid at every refresh. Resale value stays out of the base case and is shown as a sensitivity, net of selling costs and the provider's share.
The questions boards and lenders ask, answered in minutes
Each line changes one input. Owners run the same tests on their own figures, including the sensitivities lenders and investment committees ask for.
| Scenario | First retirements | Racks retiring in the TCO | Decommissioning in the TCO | Net resale, low / high |
|---|---|---|---|---|
| Base: 2-year refresh, 150 MW by 2032 | 2030 | 733 | €7.3m | €136m / €340m |
| 3-year refresh instead of 2 | 2031 | 320 | €3.2m | €59m / €148m |
| Phase 1 only: the first 30 MW | 2030 | 360 | €3.6m | €67m / €167m |
| Later start: TCO window 2029 to 2033 | 2030 | 933 | €9.3m | €173m / €432m |
| High case at 40% residual, not 25% | 2030 | 733 | €7.3m | €136m / €543m |
Carbon that stays in use
On low-carbon power, the hardware's embodied carbon is a large share of the footprint. Every accelerator that is reused instead of replaced keeps its manufacturing emissions in service.
Method: accelerators reused or resold × the manufacturer's published product carbon footprint per accelerator (NVIDIA HGX B200 summary, ISO 14067, baseboard share). A lower bound: it excludes the rest of the server.
- Accelerators reused in the TCO window
- 42,221
- Embodied carbon kept in use
- 12,000 tCO2e
What goes in, what comes out
The model works from a capacity plan before procurement, or from an asset register by serial number once hardware is live.
You provide
- Capacity by year and site, or an asset register by serial number
- Your refresh cycle and second-life use
- Rack and system design
- Your TCO window
- CAPEX, with ITAD rates and market cases from us
You receive
- Retirement schedule by year, generation and serial number
- Decommissioning line inside and beyond the TCO
- Residual-value sensitivity, net of costs and shares
- Embodied carbon avoided through reuse
- Figures ready for a bid, a lender or a board paper
Interactive demo
Try it yourself
Sign in with your work email and change the refresh cycle, capacity, timing and resale cases on an illustrative AI gigafactory. Clients run the same model on their own capacity plan or asset register through the secure portal; files are processed in memory and not stored.