The platform known as gpulist.ai, often dubbed as the “Craigslist for GPU clusters,” has recently emerged as a unique online marketplace where individuals can rent GPU clusters on a per-card, per-hour basis. Launched just a day ago, this service has already garnered significant traction, with numerous small to medium servers offering hundreds of GPUs for rent at remarkably affordable rates. This innovative approach eliminates the need for individuals to construct an entire datacenter from the ground up to access the computational power of multiple AI GPUs.
The brainchild behind gpulist.ai is Nat Friedman, a prominent AI startup investor and former CEO of GitHub, in collaboration with Daniel Gross, the co-operator of the Andromeda Cluster supercomputer. Functioning akin to Craigslist, the platform enables GPU cluster owners to create listings specifying system configurations and hourly rates per GPU, allowing potential renters to engage with sellers to finalize rental agreements. Despite its simplicity, gpulist.ai stands out as a reliable marketplace supported by a major industry player.
Contrary to initial assumptions, there are already 13 diverse listings on the platform offering A100 and H100 GPUs for rent. While a couple of these listings originate from Andromeda, the rest are contributed by six distinct third-party entities. The rental terms vary, with some listings offering flexibility in terms of rental duration and quantity of GPUs, while others necessitate long-term commitments and substantial quantities of cards.
The rental prices, ranging from approximately \(2 to \)3.50 per GPU per hour, may raise questions about their cost-effectiveness. However, considering the current market value of an individual H100 GPU at around $40,000, renting an H100 for a year or more through gpulist.ai appears to be a financially viable option. Even listings requiring renters to lease a minimum of 256 H100 GPUs for extended periods are relatively cost-effective compared to purchasing the same quantity outright.
Although the concept of gpulist.ai appears innovative, it shares similarities with other services such as vast.ai, which also facilitate computer rentals. When questioned about the distinction between gpulist.ai and vast.ai, Friedman succinctly remarked that “it’s for clusters,” indicating the focus on GPU clusters in the former. While vast.ai predominantly offers RTX 4090s, gpulist.ai boasts a wide selection of H100s, A100s, and other datacenter cards for rental, with the capacity to lease significantly larger quantities of GPUs.
The future success of gpulist.ai remains uncertain, as it remains to be seen whether it will democratize access to AI computing power or face challenges due to supply and demand dynamics. While the likelihood of datacenters specifically catering to gpulist.ai seems slim, existing GPU clusters could potentially find utility by being swiftly rented out through such platforms to generate revenue. Ultimately, these expensive GPUs need to be utilized efficiently to offset their substantial costs.
