Do GPUs hold their value?
GPUs keep earning for far longer than they keep their hourly rate. A card from 2020 still has contracted revenue in 2029, and NVIDIA says those cards run at full utilisation today, but what an hour on one costs has moved violently, falling roughly 80% from the 2023 scarcity peak before recovering. Long life, volatile price.
What an H100 hour has cost
| Period | Quoted rate | Basis |
|---|
| 2023 | $8.00/hr | Scarcity-era on-demand |
| 2024–25 | $1.50/hr | On-demand, midpoint of a $1–2 range |
| Oct 2025 | $1.70/hr | 1-year committed contract |
| Mar 2026 | $2.35/hr | 1-year committed contract |
| Aug 2026 | $2.19/hr | Entry on-demand |
What the biggest owners tell their auditors
| Company | Useful life | Note |
|---|
| Microsoft | 6 years | Raised from 4 years in Q4 FY2022 |
| Alphabet | 6 years | Normalised to 6 years by 2023 |
| CoreWeave | 6 years | Now has A100 revenue contracted past it |
| Amazon | 5 years | Cut from 6 for a subset of servers, citing AI's pace |
The case that they last
- A GPU that launched in 2020 has revenue contracted into 2029, a nine-year earning life.: “CoreWeave signed a contract for Nvidia A100 GPUs that runs into 2029, nine years after the Ampere SKU debuted.” (Mike Intrator, CoreWeave CEO, on the Q2 earnings call)
- NVIDIA says six-year-old silicon is not sitting idle. It is fully utilised.: “Thanks to CUDA, the A100 GPUs we shipped six years ago are still running at full utilization today, powered by a vastly improved software stack.” (Colette Kress, NVIDIA CFO)
- Hardware coming off contract re-rents close to its original price rather than at a discount.: “A batch of H100s coming off an expired contract was immediately rebooked at 95% of the original price.” (Mike Intrator, CoreWeave CEO, to CNBC)
- The software stack, not the silicon, sets the retirement date: CUDA compatibility keeps older cards in the fleet.: “CUDA's compatibility and our massive installed base extend the life in our systems well beyond their original estimated useful life.” (Colette Kress, NVIDIA CFO)
The case against
- A prominent short-seller argues the six-year schedules are accounting, not engineering, and that the industry is overstating earnings as a result.: “Understating depreciation by extending useful life of assets artificially boosts earnings.” (Michael Burry, who put the effect at $176bn across hyperscalers between 2026 and 2028)
- One hyperscaler moved the other way: Amazon SHORTENED the life of some servers from six years to five, citing the pace of AI hardware.: “An increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” (Amazon, in a February filing)
- Rental rates for a current-generation card can fall hard and fast. H100 hourly pricing collapsed from its 2023 scarcity peak as supply caught up.: “Rental costs dropped from $8 per hour in 2023 to $1–$2 in 2024 and 2025 due to oversupply.”
- NVIDIA ships a new architecture roughly every year, which is faster than any of these depreciation schedules.
Sources
- CoreWeave signs A100 contracts into 2029, nine years after deployment · Tom's Hardware, August 2026
- The question everyone in AI is asking: How long before a GPU depreciates? · CNBC, 14 November 2025
- NVIDIA Q3 FY2026 earnings call transcript · NVIDIA / Rev, November 2025
- Resetting GPU depreciation: why AI factories bend, but don't break, useful life assumptions · SiliconANGLE, 22 November 2025
- AI GPU rental market trends, August 2026 · Thunder Compute, August 2026
- GPU rental price index · GetDeploying, Accessed August 2026
- H100 rental price over time (2023–2025) · Silicon Data, 2025
Browse machines · How ownership works · Earning Calculator