What happens when AI hardware dies?
Of short lifespans and recycling challenges.
Where do GPUs go to die? As the world races to build massive AI data centres with GPUs scaling into the millions, what happens to the old ones?
Was chatting this morning with an executive from a leading data centre operator and the topic of GPU obsolescence came up. Is anyone even thinking of this?
Live fast, die young
We are relatively new to cramming GPUs into data centres, but the number of GPUs sold has skyrocketed since Nvidia launched the A100 in 2020.
CPUs still take the lead, but the gap is narrowing. In 2025, Nvidia is forecasted to ship up to 7 million GPUs, compared to 12-13 million server CPUs by Intel.
The key difference could be in their lifespan. Servers are generally deployed for 3-5 years or more. However, there is no precedent for GPUs - we simply don't know.
The evidence is stacked against the longevity of GPUs though:
- Extreme workloads shorten its lifespan.
- Short obsolescence cycles.
- High failure rates*.
AI training does kill GPUs faster. A report from Meta last year pegged annualised failure rates of H100 GPUs at around 9%. That's over 1 in 4 dead after three years!
One GPU server maker I spoke with related how they ended up parking spare GPUs with key enterprise/SaaS customers to minimise disruptions as they fail.
There are also anecdotes of SaaS operators stuck with large numbers of underutilised A100 GPUs as customers favour newer H200 GPUs.
Recycling GPUs
Beyond the sparring between some CEOs and their accountants over depreciation schedules, why does early GPU retirement matter?
Turns out it's much harder to recycle GPUs than CPUs.
- Complex multi-component assembly.
- More manual labour required to disassemble.
- An order of magnitude more rare earth elements.
How complex are they?
Not to be confused with consumer GPUs, an AI GPU consists of:
- GPU die.
- HBM memory.
- Memory controllers.
- Power delivery components.
- Chip interconnects (e.g. NVLink).
- Massive heatsinks or cold plates.
In short, recycling of AI GPUs is considerably more difficult due to their complicated construction and various hazardous materials present.
Despite insatiable demand for GPUs in AI data centres, I don’t think anyone is recycling them at scale at the moment.
Do you think AI GPUs are a sustainability nightmare that everyone is ignoring?
