GPU Economics

“There is a ton of capacity sitting in data centers. About 85% of the capacity sitting in 4.8 million data centers is not used… there’s this incredible capacity that’s not being used. And there is this, you know, a few companies that are capitalizing on knowledge. And we felt that it was fundamentally broken.” — Greg Osuri, 00:06:06

“Every attempt to price compute has ended up in an utter failure, EOS being I think a big example… Every time you commoditize something, turns out… there’s always going to be room for people to cheat… compute is very, very hard to commoditize. Instead of trying to commoditize, [we] create a free market that’s driven by auction… it’s like diamonds and it’s not like gold.” — Greg Osuri, 00:37:34

“It really comes down to price performance. So when you have a price performance metric, where the cost is insignificant, latency becomes insignificant as well… when you do batch optimality… the requirement really is driven by cost. So our thesis is the cost is exponentially lower, about like eight times, nine times lower, latency becomes less important for batch optimal workloads.” — Greg Osuri, 00:18:21

“Machine learning happens to be our biggest use case, even though Akash is generic compute… because of the cost advantage that we’re seeing with machine learning applications.” — Greg Osuri, 00:48:56

“Usually we’re looking at north of 85% of capacity that sits idle in these data centers… a lot of them happen to be like GPU clusters as well… they use these massive clusters to design cars… but they only use them like two hours tops a day… All in all, there are about 8.2 million data centers in the world with excess capacity that is just sitting idle.” — Greg Osuri, 00:06:02