Jeremy Hazan - Hyperbolic - TOKEN2049 Dubai 2025

So today as uh mentioned I'm going to be talking about hyperbolic. Um we are building the decentralized infrastructure uh AI cloud and would love to chat a bit more about uh what does this actually mean. So, first off, before getting started into an explanation a little bit more in depth, um I wanted to introduce um the rest of the team and actually the founders. So, um we have two co-founders. On the left, you can see um Dr.

uh Jasper Zang, who is our co-founder and CEO. So, he's sort of like a math wizard. Um he won the uh math uh Chinese gold olympiad at the sorry the the gold medal at the Chinese math Olympiad and then um went to Becking University uh which is the best university in China where he was consistently um elected I mean ranking number one students of the faculty and he was actually uh solving very very interesting problems that he went on to solve when he moved to the US uh to pursue his PhD at Berkeley in less than two years making him the fastest ever to pursue uh a PhD in the history of Berkeley. He then went to work at Citadel Securities as a quant researcher and eventually at Avalanche uh as a senior blockchain researcher. Um and then we have Dr.

Euchenzin who's our co-founder and CTO. He's way more of uh the AI background profile. So he did his uh PhD at University of Washington in AI systems and he used to be working for a company called Octoai that was acquired by Nvidia six months ago roughly and he was working on something called Apache TVM which is which used to be back then before Nvidia killed it's actually um the number one AI compiler project uh on earth. I'm going to be talking about this a little bit later on. Um and in total we're a team of 20 people super uh super lean.

So we're also very fortunate to have amazing adviserss. Um we have uh specifically Dr. Dr. Renol Shin who's a co-founder and chief architect of data bricks. Um and then we have three amazing professors.

Uh first off professor Ruca Adapa who is maybe top five worldwide expert in computer security. She's helped us a lot into integrating um confidential computing aspects uh such as tees as well as a bunch of other things. And then we have Prof. Malemi who's an amazing professor at Columbia business school and who has helped us a lot around the tokconomics and is sort of like a legend in what he does and then Profima who's a legend in AI. So let's dig in into hyperbolic and what we do and how we actually differentiate from everyone else.

So we have four core offerings. The first one being the GPU marketplace. Um and the idea here is that we're connecting idle GPU resources from all around the world from different sources. It can be on prem. Uh it can be uh clouds.

It can be coming from data centers uh and soon from individual machines. Um again I think most of you are familiar with this. It's something that has uh been doing and it's something that a bunch of other companies have been working on. That being said, on top of this, we're also focusing on building AI services on top of the GPU orchestration layer. So, it's really building one step at a time a decentralized peer-to-peer cloud.

So, what we started working on is naturally inference because for inference you don't necessarily need tons of machines um like training job for instance. Um, and that's the reason why we're basically hosting uh amazing open source um models like the top 25 uh open source models like Deep Seek and Llama and a bunch of other damn of of other uh ones. Um three is proof of sampling. So I briefly mentioned um the fact that we're going to be opening up to individual nodes um basically you know in a permissionless way third party nodes to actually share their compute and their GPUs on hyperbolic and for that we needed to have a verification mechanism that could live it up to the expectations of um you know something that would be functional. So for that uh there are some solutions that are existing like OPML, ZKML, consensus but none of them are really practical.

it's usually a major computational overhead and that's the reason why we really created our own and we'll be talking uh in a couple slides about that and then the latest one is our Adrian framework uh which again I'll be talking more later on. So today we're really excited to say that we're the most popular decentralized infrastructure cloud um uh we might need to update those numbers because every single day things are changing but uh we're actually almost at 200k AI developers on Hyperbolic. We have amazing web two clients. Um, universities like Stanford, Berkeley, NYU, Cornell, UCSD and a bunch of other ones, but also very solid big web 2 companies like Kora, Hugging Phase, Versel, and a bunch of others that are coming in really, really soon. Uh, we're also powering top crypto AI projects.

So, we've been powering Eliza, Virtual, Solang, Arc, Zero Bro, um, Near, uh, layer, and a bunch of other ones. So to go back on what I was explaining before which is basically the uh GPU compute marketplace. Um we're again available at scale. This doesn't necessarily change from other projects. So um it's it's just you know something that we all want to to uh to to showcase and and it's sort of like a step of success.

Um what what is really interesting is that we're we can really really cost um sorry cut the the cost. So uh we can go up to like 75% cost reductions compared to like traditional cloud. Um and the last thing is is again as I mentioned is like this permissionless access and the fact that we created a new OS called Hyperdos which is short for hyperbolic distributed operating system which is actually a a really new invention. Um and because of this new invention and the fact that it's extremely uh robust and scalable based on federated Kubernetes, we are allowing anyone to share their compute in a very easy way and to actually make sure that we can have quality computes. If it's quality compute, it's going to go in either semi-rusted or trusted and if it's not, it directly goes into untrusted and no one and the compute won't be listed in our marketplace which is a very important aspect actually.

So around open source intelligence um this is more around you know uh inference as I was saying we're offering the latest models we're trying to be very cognizant about this the second thing is uh we're offering absolute privacy so you can run your own models and keep your data yours and eventually affordable again uh I don't think this is necessarily an argument to be completely frank because uh the cost of you know inference is going down to zero either way it's becoming a commodity but at the end of the day if we're able to offer, you know, the best models optimized at a very very low cost, we're still going to be winning over the competitors. So, it's it's still at least today an important point. So, this is just a screenshot of um one of our accomplishments is uh we are now the only web 3 project that's providing serverless inference on hugging face. Uh so I don't know for those of you who know about hugging face but it's it's one of the biggest um uh company in in AI um and and it's very very interesting to be listed there because it's actually a really great distribution channel for for us and for the space in general. This is another screenshot of what we've been doing on Kora.

So we've been working very very closely to Kora especially with the the founder Adam D'Angelo who used to be the CTO of Meta before and is a very prominent figure in in uh in the Bay Area. Uh last but not least in terms of screenshots, this is basically a huge thing that happened a couple months ago. Um we are in a position where we're starting to become pretty um locked in with developers and pretty popular with developers. And this is uh Andre Karpathy. So for those of you who don't know, Andre Karpathi used to be one of the founding uh team members of Open AI and he's used to be the head of AI at Tesla.

He's today top five most prominent AI figures worldwide. And he's now educating people on AI and he made a video a couple months ago on on YouTube and he was saying that Hyperbolic is his favorite place on earth actually to use um base models which is a type of uh of way of serving uh inference models. So going back on the web three side as I was explaining a little bit earlier um we are at the end of we needed to actually finish this black box of AI and especially when you're doing it in a decentralized way and when you're accepting in a permissionless way third party notes. So that's the reason why we created our proof of sampling verification mechanism. Um something that's basically you know uh we can verify the outputs but in two ways.

We wanted to have a very high security like what ZK promises but also very high levels of scalability what OP can do. And so it's it's actually very funny to be able to uh work on both and and this is what we've been pushing. Um so yeah as as it explains here u the way we're doing this is is pretty easy. Basically we have um proof of sampling is is going to be reducing a lot the overhead the computational overhead compared to consensus. So as you know there's going to be a 10x overhead you know if it's reproduced by 10 nodes and obviously reducing also compared to ZK because it's 100x to a 1000x overhead.

Um and then in terms of scalability we're we've been thinking about something similar uh basically to what OPML is is doing. uh and again we've had the genius uh coming in from our CEO in math um to to really focus on the right parameters for this pure Nash equilibrium to actually happen and and really accomplish those goals. So in the agent framework I didn't talk about this uh earlier but the idea is that we are building an infrastructure and because we're building an infrastructure we are in a position where we can have a self- evvolving intelligence. So what does this mean? We started creating an agents.

Um, we're at the very early steps of agents, but things are moving really fast and what we've been doing so far is very fun. It's like we've been creating with the CDP kits from Coinbase. We've been working with Coinbase on that. Um, we have injected some crypto funds in the wallet of uh our agent and basically our agent is able to call the API to check which compute is available in hyperbolic and can actually rent out the GPU and literally live on hyperbolic. it can ingest a specific data set um and become a more verticalized version of itself like by doing fine-tuning on the GPU.

It can become an operator on layer. It can become a validator on Ethereum. This is something we've we've done already which is really really fun. So yeah, so this is a really promising um aspects and offering that that we're we're giving. So economic layer.

So I I didn't mention this but we're Hyperbolic is based in San Francisco. are really at the center at the heart of everything that's happening in AI today and most of the people in San Francisco are in the Bay Area in AI are not necessarily inclined towards web three but they have similar views they have similar mindsets they all talk about open-source AI they all talk about you know giving access to uh people all around the world developers all around the world without the barriers of the huge margins taken by the clouds and So, one thing that we're very proud of at Hyperbolic is that because of this aspect of web 3, I mean crypto plus AI because we're going to be launching our token soon, is that we're in a position where we can actually help uh creators all around the world monetize especially open source. The big the best example for there are two examples that I want to give. The first one is um Meta. So, so far before Deep Seek Moments, Meta was the only company in the world that was actually capable of saying, "Okay, let's burn $15 billion to train and build Llama.

We don't care because the return on investment will get back on our stock market." Uh, which is a very very rational uh decision. The problem they have 1 billion users on Llama today. The problem is that Deepseek arrived and changed the entire space because now people started realizing, okay, with just a few million dollars, I'm not saying it's easy, but a lot of founders all around the world can raise 2 to3 million can actually build a model that's a competitor to Chad GPT and a model that's competitor to Claude. And this is a tremendous thing that changed in the industry.

And because of that, we're very very proud to say that on Hyperbolic, it's something that you can do. Um it cost $5.5 million on uh for DeepSeek for training apparently. Um if we made the calculation if they did it on a hyperbolic it would have cost them $2.1 million.

Um so that's the first thing and then the second thing is that we can help monetize open source which we started doing with Black Forest Lab the company that's behind Flux which is really really fun. So as sort of like a recap and summary of everything that I explained, this is our ecosystem architecture which I think is very important to take a step back um to understand things a little bit better. The idea is that um as you can see in the bottom uh you know we have the decentralized orchestration layer that's coming from those different sources. It's coming from data centers, mining farms, individual machines, on-prem machines, etc. It's basically on top of that we have the AI services ecosystem.

So we everything that's in purple is built by Hyperbolic by the way. Everything that's in white is companies that are building it and we're we're partnering with them. So uh we started with the inference engine um with genen, LLM, thanks to speech, videogen etc. and then uh agent framework but then there are other things that we're either going to be collaborating on especially you know uh the training decentralized training engine a lot of amazing web3i companies are working on that um and then we're going to be adding fine-tuning very very soon. The second aspect of this is the verification layer.

So in order to start grasping all of the different you know uh like the web 3 world and and start having all of those machines coming in you know in a permissionless way um into our our marketplace then basically we needed to have this verification layer that's going to that we're going to be implementing at some point and then potentially launching our own blockchain which can be an SP rollup on top of proof of sampling. Uh the last step is privacy layer which is basically you know uh trying to target even more web2i companies thanks to te. So uh this is actually my my last slide. Um it's more around like the ecosystem map that we've had so far. We've had amazing research partners and clients.

Um LMS chatbot arena is one of them. I don't know if you know about this website but I highly recommend you to go there. Uh it's the best number one's evaluation service platform. uh Elon Mus, Sam Altman, Zuckerberg, they're all on this website consistently trying to make sure that their model is number one on the leaderboard. Um and then data layer training services like Ian layer and a bunch of other ones as well as the models that we've been trying to help monetize with revenue sharing as I explained before.

Um so yeah, thank you very much for listening. Um if you want to follow us, uh you can basically scan uh any of those uh those QR codes and I'll be here if you have any questions. Thank you very much.