Do GPUs hold their value? Long life, volatile price.
A GPU from 2020 has revenue contracted into 2029, and NVIDIA says it still runs at full utilization. Its hourly rate is another story: an H100 hour fell roughly 80% from its 2023 peak before recovering.
Compiled August 2026 from 6 public sources, including the ones that argue against us.
What an H100 hour has cost.
H100 hourly rental rates from 2023 to August 2026, each on its source's basis.
| Period | Quoted rate | Basis |
|---|
| 2023 | $8.00/hr | Scarcity era, reported as over $8 |
| 2024–25 | $1.50/hr | Midpoint of a reported $1–2 range |
| Aug 2026 | $2.19/hr | Lowest tracked on-demand rate |
- ~80%: Off the 2023 peak by 2024–25
- $1.09/hr: Lowest tracked A100 rate, 6 years after launch
Three measures. Different meanings.
A purchase price, an accounting schedule and a rental quote answer different questions.
Acquisition
Your agreed acquisition price is the starting cost. It does not establish a future resale price or a guaranteed rental return.
Useful life
The cited owners report five- or six-year accounting lives. Those schedules allocate cost on their books; they do not promise how long your machine will last or what it can be sold for.
Rental hour
The A100 observation is a published hourly rental floor, six calendar years after launch. It is not a resale transaction, typical earned rate or promise of available demand.
No resale value is inferred here. The cited rental quotes and accounting schedules do not establish your machine’s future sale price.
The books say six years.
Depreciation schedules are audited and move earnings, so they are the closest thing to a public estimate of useful life. One owner went the other way.
| Company | Useful life | Note |
|---|
| Microsoft | 6 years | Raised from 4 years in Q4 FY2022 |
| Alphabet | 6 years | Normalized 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 for, and against.
We sell these machines, so weigh the case for with that in mind. The strongest case against gets equal space.
They last.
- Old cards still get long contracts. “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)
- Six-year-old silicon is still busy. “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 off contract is rebooked at once. “A batch of H100s coming off an expired contract was immediately rebooked at 95% of the original price.” (Mike Intrator, CoreWeave CEO, to CNBC)
- Software, not silicon, sets the retirement date. “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)
They don’t.
- The long schedules may be accounting, not engineering. “Understating depreciation by extending useful life of assets artificially boosts earnings.” (Michael Burry, who put the effect at $176 billion across hyperscalers between 2026 and 2028)
- One hyperscaler shortened its schedule. “An increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” (Amazon, in a February filing)
- The rental market flipped to oversupply. “Don’t buy H100s. The market has flipped from shortage ($8/hr) to oversupplied ($2/hr) because of reserved compute resales, open model finetuning, and decline in new foundation model co’s.” (Eugene Cheah, CEO of featherless.ai)
- NVIDIA ships a new architecture roughly every year, faster than any of these schedules.
What our model assumes.
20%: fall in hourly rates, every year · 50%: Floor, as a share of the starting rate
Our payback model assumes rates fall every year until they reach the floor, and every payback figure on this site is computed after that decline. These are modeling assumptions, not forecasts, and not a promise about any machine.
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, November 14, 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, November 22, 2025
- From $8 to $1 per hour: GPU rental prices are crashing · Cybernews, February 24, 2025
- AI GPU Rental Market Trends (August 2026): Complete Industry Analysis · Thunder Compute, August 13, 2026 · archived August 28, 2026
Rates and schedules move, so check the sources. Nothing here is investment advice.
Weighing it up?
Ask the people who build the machines the hard questions.
Next: Run the numbers yourself · Book a call · How ownership works · Browse machines