Previous lesson
What is GPU utilization
Continue the Model Costs track.
← Back to Compute College
Compute College
How Fast Does an H100 Depreciate?
GPU depreciation spreads an accelerator's purchase cost over its useful life, which drives the real cost of every GPU-hour.
Compute & Pricing LessonsLearning path
One concept connected to AI compute market decisions.
5-8 minutesRead time
A practical introduction designed to be completed in one sitting.
Depreciation / Hardware / CostTags
Useful for operators, analysts, investors, and procurement buyers.
Plain-English definition
Plain-English definition
Depreciation is how the cost of a GPU is spread over its useful life. Asking how fast an accelerator depreciates is asking how quickly an expensive chip loses economic value — through wear, newer generations, and falling rental rates — which sets how much of its purchase price must be recovered in each year it operates.
Why it matters
Why it matters
The assumed useful life of a GPU drives the cost of every GPU-hour an owner sells or uses. A longer assumed life spreads the purchase cost over more hours and lowers the hourly cost; a shorter life concentrates it. The AI-capex debate turns on whether operators are depreciating GPUs over too many years and understating the true cost of compute as newer chips arrive.
- The assumed useful life sets how much purchase cost each GPU-hour must recover.
- Newer generations and falling rental rates can shorten an accelerator’s economic life faster than physical wear.
- A long depreciation schedule can understate the real cost of compute and flatter reported margins.
Simple example
Simple example
Suppose an accelerator costs an illustrative $30,000. Spread straight-line over 3 years that is $10,000 a year; over 5 years, $6,000 a year. At an illustrative ~70% utilization (about 6,100 productive hours a year), the hardware-only cost is roughly $1.64 an hour at 3 years versus about $0.98 at 5 years — before power, networking, or facilities. The schedule alone moves the hourly cost by around 40%.
- A longer assumed life lowers hardware cost per hour but bets the chip stays competitive.
- Utilization matters as much as the schedule: idle years still depreciate.
- Purchase price, useful life, and utilization here are all illustrative assumptions.
Example figures are illustrative calculations, not current quoted market prices.
Market signal
How to read the market signal
Disagreement over GPU useful life is a signal about the true cost of compute and about operator economics. If newer generations arrive faster or rental rates fall, effective useful life shortens and per-hour cost rises. Watch how operators state depreciation schedules and how quickly prior-generation rental rates decline.
- A faster generational cadence shortens economic life regardless of physical durability.
- Declining prior-generation rental rates reveal real-world depreciation in the market.
- Depreciation assumptions are a lever on reported compute margins.
Market read: the depreciation schedule is a hidden driver of compute cost and operator margin. Evidence discipline: state the purchase price, useful life, and utilization behind any per-hour cost, and treat schedules as assumptions, not facts.
Common mistake
Common mistake
Treating a GPU’s hourly cost as fixed. Most of it is recovered purchase cost spread over an assumed life and utilization — change either and the cost moves a lot. Assuming a long life also ignores that a newer chip can make an older one uncompetitive well before it physically wears out.
Practical takeaway
What you can do with this
When you see a GPU-hour cost or an operator margin, ask what depreciation schedule and utilization it assumes, and test it against a shorter useful life.
- Buyers and operators: model hardware cost per hour under both a conservative and an aggressive useful life.
- Analysts: stress-test operator margins against shorter depreciation as newer generations ship.
- Separate physical wear from economic obsolescence; the latter usually bites first.
- Treat purchase price, life, and utilization as stated assumptions, not observed facts.
- Keep modeled per-hour costs separate from observed rental quotes.
Decision check: before trusting a low GPU-hour cost, check whether its assumed useful life survives the next generation of accelerators.
Helpful memory trick
Helpful memory trick
A GPU usually goes obsolete before it wears out — economic life, not physical life, sets the clock.
Compute College
Turn the lesson into a number
Use the GPU-Hour Cost Calculator, AI Training Cost Calculator, or Model Serving Cost Calculator.
Compute College track
Model Costs
Continue this Compute College lesson path
Next lesson
What is model flops utilization mfu
Continue the Model Costs track.