‘Craigslist for GPUs’ launches amid crushing Nvidia AI GPU shortages — gpulist.ai allows you to rent AI GPU clusters by the hour

The self-proclaimed “Craigslist for GPU clusters” is here; gpulist.ai bills itself as the only place on the internet where anyone can rent a GPU cluster by the card and by the hour. The service launched just yesterday and already small to medium servers are offering up hundreds of GPUs to the masses, for mere dollars an hour per card. Now anyone could theoretically harness the power of several AI GPUs without needing to build a datacenter from scratch.

The site was announced by AI startup investor and former GitHub CEO Nat Friedman, who with Daniel Gross runs the Andromeda Cluster supercomputer. It works pretty much the same as Craigslist: people with GPU clusters make listings that detail system specifications and cost per GPU per hour, and users can contact sellers to hash out an arrangement. It seems there’s nothing super complicated about gpulist.ai itself, as it simply offers a marketplace backed by a big industry player.

You would probably think it’s super unlikely that anyone who owned servers with GPU clusters (which would probably cost at least hundreds of thousands or millions at minimum to build from scratch) would list them on a Craigslist-like website, but you’d be very wrong. Already there are 13 different listings that offer A100 and H100 GPUs for rent. Two of these listings are from Andromeda, which is owned by Friedman and Gross, but the other 11 are from six distinct third-parties.

Some of these listings are pretty flexible, offering a minimum rent period of one week and as few GPUs as four, and in one case just one card. However, there are many listings that expect renters to make long-term or large-quantity commitments. Three listings for H100 clusters ask renters to rent out at least 256 cards; there is another that has a minimum of 128 and another with 50. Those same listings also require lengthy rental periods, from 12 weeks at the low-end to as high as 104 weeks, or nearly two years.

At prices ranging from roughly $2 to $3.50 per GPU per hour, you might be wondering whether these rates are actually worth it. Well, recent estimates put the price of a single H100 GPU at roughly $40,000; for that same $40,000, you could probably rent an H100 for a year to a year and a half based on current prices at gpulist.ai. Even that 104-week long listing that asks for renters to rent at least 256 H100 GPUs only costs about as much as 256 H100 cards. Considering the time and money that goes into building or upgrading a datacenter (it’s more than just the GPUs), it’s actually not a bad deal by any means.

While this all sounds very novel, it’s not quite 100% original, as vast.ai is another service that allows people to put computers up for rent. When asked about how similar gpulist.ai is to vast.ai, Friedman simply said “it’s for clusters.” That’s pretty self-evident when you compared what GPUs the two services offer. vast.ai largely offers RTX 4090s, and there are very few H100s, A100s, and other datacenter cards available to rent. The biggest single H100 server to rent on vast.ai has just eight GPUs; you can seemingly rent at least 256 and up to 1024 H100s on gpulist.ai.

It’s not clear if gpulist.ai will help usher in an age of easy-to-access AI computing power, or if it’ll just flop because of a lack of supply and/or demand. It seems unlikely any datacenter is going to set out to make GPU clusters specifically for gpulist.ai or a similar service, since the payoff wouldn’t seem to be worth it. However, if a datacenter already has GPU clusters ready to go but nothing running on them for whatever reason, quickly putting them up for rent could be a way to at least make some money rather than none at all. These ultra-expensive GPUs have to be paid-off somehow.

This post was originally published on this site