Aethir, a decentralized GPU cloud that pools graphics cards from operators around the world and rents the computing power for AI, argued that the resale value of aging chips depends on someone willing to run cheaper workloads on them.
That someone, the company said in a blog post, could increasingly be a decentralized network. The pitch: A GPU too old for frontier model training can still answer AI queries, and networks that aggregate idle silicon create the marketplace that gives a used card a price floor.
The stakes are real money. Cutting the assumed useful life of a GPU from six years to two or three could dent big cloud providers' cumulative earnings. Amazon booked roughly $920 million in accelerated depreciation after trimming server lives to five years; Meta cut expense by about $2.9 billion by extending them to 5.5.
Aethir frames itself as the venue that keeps the cascade honest, since a depreciation assumption is "self-certified rather than market-tested" without an open place to sell year-four capacity. It says it runs 430,000-plus GPU containers across 94 countries, offering H100-class chips at $1.25 an hour.
Used A100 cards still trade at $12,000 to $18,000, the company noted, though it acknowledged residual value floors near 10% to 20% of original cost.
(HOW AI WAS USED IN THE PRODUCTION OF THIS PIECE: I wrote the first draft on Claude Desktop using a custom blog-rewrite skill plus background research on Aethir. I then used our Distro Publisher MCP connector to file the piece directly to DistroVerse, our web app, where I used our built-in story editor to edit the piece.)