GPU Economics

“One of the most exciting incentive models for me was, well, if I contribute compute to you, to your model, and if I get an ownership based on my contribution from the model, that means I get inference revenues, because inference is how you make money, and training is how you spend money… if there is some ownership rights that I can get from my contribution, I think we hit the nail on the head, because that solves everything.” — Greg Osuri, 00:38:15

“這是就是一開始 然後的需求 最好的預測是在數據上的需求 是在Gartner做的 他們說5%的GDP會在數據上的數據上 5% 2030年 那是我們有的需求” (Translation: “This is just the beginning of the demand. The best forecast of demand for data [compute] was done by Gartner — they say 5% of GDP will be [spent] on data [compute] — 5%, by 2030. That’s the demand we have.”)” — Greg Osuri, 00:00:00

“Over 50% of every single penny we spend on online services goes to one of the three cloud providers. All of us pay a cloud tax without knowing who we’re paying the tax to… [AWS] is alone a trillion dollar business for Amazon, and profit margins are over 60%, and that’s not okay — you can’t have an [ab]normally profitable layer and you call it a public utility… decentralized systems like Akash [are] 80% cheaper than what you would otherwise pay on Amazon.” — Greg Osuri, 00:05:12

“Our across-the-board utilization right now is 70%, and that was, what, 10% when we began… per GPU right now is about $20 per day, compared to $10 roughly, that was about 12 months ago.” — Greg Osuri, 00:18:06

“H200s, which are best GPUs to run DeepSeek, are 100% utilized — that tells you that H200s are high demand. H100s were pretty high utilization… so we know that H100s, A100s, H20s, 4090s have high demand.” — Greg Osuri, 00:15:57

“So far we were able to account up to 900,000 GPUs that could potentially come onto Akash that were all left from the Ethereum merge… In the next year — I think in this whole year, in 2023 — all the supply will be online, and we’ll have Akash AI as a beautiful product to deliver this value to AI devs.” — Greg Osuri, 00:48:35

“We’re anticipating from our estimations that an A100 will cost you somewhere around 50 cents, compared to a traditional cloud offering which is about $4 an hour… If you ask any AI dev and offer them an A100 for 50 cents, they will kidnap you.” — Greg Osuri, 00:20:26

“GPUs are the fuel for AI. The need for GPUs is doubling every 3.5 months as these LLM models get more advanced… GPT-3.5 is 175 billion parameters and GPT-4 is going to be a trillion parameters — that means you need that much of compute.” — Greg Osuri, 00:21:11

“The 10,000 or so enterprise data centers, their utilization rate is somewhere around 15%. So there’s enormous underutilization in these 10,000 professional one-megawatt-capacity data centers.” — Greg Osuri, 00:07:19

“We’re growing at 10x now every year… we want to get to 10,000 GPUs by end of the year… about 100,000 GPUs by end of next year, about a million GPUs within the next three years.” — Greg Osuri, 00:44:19

“A big portion of GPUs are actually with the crypto miners, especially the ones that invested quite a lot to mine ETH — post-merge, really, you have nowhere to go. So we have in our inventory about a million GPUs that we have access to in our network alone, in our community, that we potentially want to bring to market.” — Greg Osuri, 00:25:07

“For someone to really take over the world, they need GPUs. Whoever has the power — GPUs are the oil, right? Like you can have the most advanced tank, but without diesel, there’s no point. You need the fuel for AI. The ones that control that fuel will be in the best position to control the world.” — Greg Osuri, 00:26:32

“What OpenAI has that other people don’t have is GPUs — that’s where the power comes in. It’s not the model, it’s really the GPUs and the data set. That’s where the power is… OpenAI was able to secure great deals with Nvidia, with cloud providers, to be able to get ahead in the game, but I don’t think they have the advantage anymore.” — Greg Osuri, 00:22:16

“One Nvidia GB200 NVL72 rack draws about 120 kW… To reach a gigawatt of compute, about 8,300 of these racks — traditionally that would mean a dozen hyperscale campuses. But with Akash, those racks can be anywhere… We’re not bottlenecked by real estate or substation permits. We route around the constraints… Gigawatt power without a gigawatt footprint.” — Greg Osuri, 00:11:20

“In order to get decent GPUs, you need to have long contracts with cloud companies. They’re going to pick and choose the highest bidders, and it becomes prohibitively expensive. It reminds you of the Middle Ages where you had feudalism… These few lords own the grain, and we all have to stand in line to get the grain because if you don’t, you don’t get the AI.” — Greg Osuri, 00:40:45

“Our utilization right now is somewhere north of 70%. For comparison, AWS has about 40% utilization… most decentralized networks have super high supply and less utilization — I think Filecoin has 0.5% utilization.” — Greg Osuri, 00:30:40

“You can purchase GPU, CPU based compute for about 80% cheaper than what you would otherwise pay on the cloud… in many documented cases, publicly, users can deploy on Akash within two minutes.” — Greg Osuri, 00:00:52

“Since enabling GPUs, the network has scaled roughly from about 150 to about a thousand plus GPUs while sustaining roughly 54 to 55% average utilization across a very heterogeneous fleet — that’s competitive with large clouds and far above typical on-prem. It means few idle watts and more useful gradient steps per kilowatt hour.” — Greg Osuri, 00:12:32

“There are kids in colleges [who] couldn’t get access to GPUs on Amazon because Amazon would only select the customers based on who pays the highest… one student from Cornell couldn’t get GPUs on Amazon, came to Akash, could get his GPUs, built a company that eventually sold to Nvidia… it’s clear that decentralization is working for people that centralization has failed.” — Greg Osuri, 00:07:18